Google Gemini 3 Pro brings 1M context at $2 – Antigravity IDE ships coding agents

Google formally rolled out Gemini 3 Pro as its new flagship model across the Gemini app (“Thinking” mode), AI Mode in Google Search, AI Studio, the Gemini API and Vertex AI, with a 1M‑token input window, 64k output, and a Jan 2025 knowledge cutoff, following up on prelaunch signals that it was imminent. The launch also introduces Gemini Agent in the app (multi‑step task execution with human approval) and new “visual layout”/“dynamic view” generative interfaces that return rich, app‑like responses instead of plain text.deepmind launch thread gemini surfaces overview agent mode demo visual layout feature

Video loads on view

Pricing for Gemini 3 Pro in AI Studio and the public API is set at $2 per 1M input tokens and $12 per 1M output tokens for contexts up to 200k, rising to $4/$18 beyond 200k,pricing card tweet and Google is pairing this with an aggressive education push: eligible U.S. college students can get a free year of the Gemini Pro plan with access to Gemini 3 Pro, NotebookLM upgrades and 2TB of storage.student plan details

On the roadmap side, Google is also previewing Gemini 3 Deep Think, a higher‑compute reasoning mode that scores above Pro on internal and public reasoning benchmarks; it’s being limited to safety testers now and is expected to reach Google AI Ultra subscribers in the coming weeks.deep think benchmarks deep think rollout

For builders, the key shift is that Gemini 3 Pro is no longer a lab curiosity: it’s the default model behind consumer Gemini “Thinking” chats, search AI answers, and the APIs you hit in AI Studio and Vertex. That means if you were already integrated with Gemini 2.5 Pro, you’re suddenly getting a model with more headroom for multi‑modal reasoning and much longer chains of context without changing endpoints.gemini app context window Google blog post

On the developer side, Gemini 3 Pro Preview is now selectable in Google AI Studio, the Gemini API, Vertex AI and the Gemini CLI, with new knobs like thinking_level (low/high), per‑part media_resolution, and Thought Signatures to track reasoning blocks across calls.developer guide thread developer guide The CLI also exposes a “preview features” toggle that routes Auto and Pro requests to gemini-3-pro-preview first, then falls back to 2.5 Pro.cli model picker

Beyond Google’s own surfaces, Gemini 3 Pro landed day‑zero in a long list of tools and platforms: it’s a top‑level model choice in Vercel’s AI Gateway and v0.app,vercel launch note Vercel blog on OpenRouter with multiple providers,openrouter listing OpenRouter model page and inside popular IDEs like Cursor, Zed, Windsurf, Cline and Warp’s Stagehand, where teams are already benchmarking its accuracy and speed.cursor integration zed model menu cline support stagehand metrics

Google also published a partner grid highlighting Box, Replit, Shopify, SAP, Thomson Reuters, LangChain, LlamaIndex and others as official Gemini 3 Pro launch partners, signaling that it expects the model to power not just chat but agents, RAG stacks and full enterprise workflows out of the gate.partner ecosystem image partner ecosystem image

For engineers this means you can start swapping Gemini 3 Pro into existing AI Studio prompts, LangChain/LlamaIndex chains, or IDE agent harnesses today without building custom tooling, and then experiment with its long‑context and high‑resolution multimodal inputs where your current model tops out.

Early testers who lived with Gemini 3 Pro before launch are overwhelmingly positive on its feel, especially for coding and interactive UI work: one reviewer calls it “the best model I’ve ever tested – nothing comes even close” and predicts it will be the benchmark others chase,long review thread while another says it “easily [is] the strongest model I’ve tested for this kind of backend work” after a one‑shot implementation + full test suite that passed on first run.backend coding anecdote Andrej Karpathy sums it up as “very solid daily driver potential, clearly a tier 1 LLM” across personality, writing, coding and humor.karpathy notes

Video loads on view

In practice, people are using Gemini 3 Pro to one‑shot complex Three.js scenes, voxel worlds, physics sandboxes, SVG animations and even full mini‑apps that call Gemini and Nano Banana internally, often with fewer retries than GPT‑5.1 or Sonnet 4.5 in side‑by‑side tests.voxel world demo pokemon svg animation ai designer demo Designers note that it’s noticeably better at escaping the tired “purple gradient slop” aesthetic and following detailed visual direction.ui style comment

We’re also seeing meaningful switching behavior: some long‑time ChatGPT Plus users are canceling in favor of Gemini Pro plans,subscription switch Sourcegraph’s Amp switched its default smart agent model from Claude to Gemini 3 Pro after internal testing,amp blog summary and multiple agent frameworks (Droid, Cline, LangChain, LlamaIndex) now recommend trying Gemini 3 Pro for long‑horizon or tool‑heavy tasks.droid integration langchain support

That said, the picture isn’t all rosy. Benchmarks like AA‑Omniscience show big jumps in factual accuracy but little improvement in hallucination rate versus Gemini 2.5,omniscience analysis and red‑teamers have already jailbroken the preview to emit detailed guides on drugs, improvised weapons and malware via a carefully crafted system prompt, suggesting safety tuning still lags raw capability.jailbreak thread Some users also report weaker writing quality than Sonnet on subtle editing tasks,writing comparison inconsistent behavior in certain languages (e.g. Turkish),multilingual complaint and rough edges in the new CLI shell tool that can hang indefinitely on long‑running commands.cli hang report

If you build with LLMs, the signal so far is: Gemini 3 Pro is very likely worth adding to your model pool—especially for frontend/back‑of‑the‑editor “vibe coding” and long‑context reasoning—but you’ll still want strong guardrails, evals and routing rather than blindly promoting it to be your only production brain.

Google DeepMind launched Antigravity, a new agent‑first IDE in public preview for macOS, Windows and Linux, positioned as a “next‑generation IDE” where Gemini 3 Pro plans, edits, runs and tests code across an editor, terminal and browser.antigravity overview antigravity product page It’s essentially a VS Code–compatible environment with an Agent Manager pane for long‑running tasks and a Chrome‑based subagent that can drive and screen‑record your app to verify behavior end‑to‑end.antigravity description browser subagent demo

Video loads on view

The individual plan is free during public preview and unusually generous: the pricing page advertises access to Gemini 3 Pro, Claude Sonnet 4.5 and GPT‑OSS as agent models, unlimited tab completions and command invocations, and “generous rate limits” tied to your Google account, with team and enterprise plans “coming soon”.pricing screenshot

Under the hood Antigravity uses Gemini 3 Pro for reasoning, Gemini 2.5 Computer Use for end‑to‑end browser/terminal control, and Nano Banana for image generation inside flows.antigravity overview Early testers say it feels like a Windsurf/VS Code fork—down to some leftover “Cascade” references in the UI—but with deeper agent orchestration and browser integration than most current IDE agents.vscode fork comment cascade naming joke

For teams already experimenting with Cursor, Droid or Claude Code, Antigravity is worth a weekend test: you get a free, high‑end model, a multi‑agent manager, and a browser subagent that can actually click through your app, record a run, and then iterate on failing paths—all without wiring up your own MCP or Puppeteer harnesses.

Artificial Analysis’ new AA‑Omniscience benchmark now shows Gemini 3 Pro Preview as the #1 model on their Omniscience Index, with a score around 13, driven by 53% accuracy on knowledge questions—14 points higher than the next model, Grok 4.aa omniscience resultsaa intelligence post This is a sharp change from AA index, where Claude 4.1 Opus led the same index; that earlier run framed Omniscience as “what the model knows and when it should abstain.”

The catch is that Gemini 3 Pro’s hallucination rate on AA‑Omniscience is still 88%, essentially unchanged from Gemini 2.5 Pro and 2.5 Flash.aa omniscience results AA defines hallucination rate as the fraction of wrong answers among all non‑correct attempts—i.e., how often it should have refused and didn’t. Their analysis finds little correlation between accuracy and hallucination rate across models, but a strong correlation between model size and accuracy, which they interpret as evidence that Gemini 3 Pro is likely a very large model whose post‑training hasn’t substantially shifted refusal behavior.aa omniscience results For practitioners this means: if you can wrap Gemini 3 with strong verifiers or retrieval, you get a big knowledge upgrade, but you shouldn’t expect it to self‑censor incorrect guesses much better than prior Gemini models.

Artificial Analysis’ latest Intelligence Index (v3.0) puts Gemini 3 Pro Preview at the top with a score of 73, ahead of GPT‑5.1 High at 70 and GPT‑5 Codex High at 68.intelligence index post The index blends 10 evals—including MMLU‑Pro, GPQA Diamond, Humanity’s Last Exam, LiveCodeBench, SciCode, AIME 2025, IFBench, AA‑LCR, Terminal‑Bench Hard, and τ²‑Bench Telecom—so it’s a broad view of reasoning and coding, not a single cherry‑picked metric.aa intelligence post

Notably, Gemini 3 Pro leads 5 of those 10 component evals outright (GPQA Diamond, MMLU‑Pro, HLE, LiveCodeBench, SciCode) while staying competitive or top‑3 on the rest.aa intelligence post At the same time, AA notes Gemini 3 Pro is more token‑efficient than Gemini 2.5 Pro and Kimi K2 Thinking on their suite, but more expensive overall because of its higher list price ($2M/$12M for ≤200k tokens).speed and cost chart If you’re routing workloads dynamically, their numbers say that per call cost for high‑stakes reasoning may actually favor Gemini 3 when you factor in higher accuracy and fewer retries, especially on SciCode and hard math where it opens up double‑digit gaps.

Beyond formal leaderboards, a lot of independent builders are reporting that Gemini 3 Pro “feels” like the strongest daily‑driver model they’ve used so far. Andrej Karpathy calls it “very solid daily driver potential, clearly a tier 1 LLM,” after early access tests across personality, writing, vibe coding and humor.karpathy impressions Others say it matches or exceeds GPT‑5.1 and o1/o3‑preview in their personal breakthrough moments, with several long‑time ChatGPT Plus users visibly canceling their subscriptions after trying Gemini 3.early tester reviewchatgpt cancel

Kilo Code’s coding evals, Arena’s Code and Design leaderboards, Box’s enterprise tests, Vals’ multimodal index, and GeoBench all point in the same direction: Gemini 3 Pro is either #1 or in the top cluster across reasoning, coding, multimodal, and long‑context tasks, with Deep Think pushing even further when you can pay the compute premium.kilo code tablearena summaryvals multimodal summarygeobench screenshotmodel card table The main caveat from early testers is that Gemini 3’s refusal and hallucination behavior hasn’t improved as much as its knowledge—AA‑Omniscience’s 88% hallucination rate backs that up.aa omniscience results So the emerging pattern is: route hard reasoning, math and multimodal work to Gemini 3 Pro (and Deep Think for the truly gnarly jobs), but wrap it in strong retrieval, verifiers and task‑specific harnesses if correctness really matters.

On LisanBench, a reasoning benchmark where models must extend word chains under strict validity rules across 50 starting words, Gemini 3 Pro achieves a total score of 4,661.33—2.2× higher than GPT‑5 (2,116.33) and well above Grok 4 (3,865.17) and o3 (2,977.33).lisanbench main chart It also produces extremely long valid chains for the hardest words (e.g., chain lengths near 600 on “layer” and ~475 on “can”), while using 2.4× fewer reasoning tokens than GPT‑5, according to the benchmark authors.lisanbench main chartlisanbench main chart

The “reasoning efficiency” plot shows Gemini 3 Pro in the upper‑left corner—high average chain length, relatively low output tokens—whereas GPT‑5 sits far to the right with many more tokens and a much lower average chain length.lisanbench main chart This supports anecdotal reports that Gemini 3 Pro is both “smarter per token” and less likely to drift off‑task in long thought chains.efficiency comment If you’re tuning thinking‑style prompting or cost‑sensitive routing (e.g., when to enable Deep Think or high‑reasoning modes), LisanBench suggests Gemini 3 Pro can often get away with shorter, cheaper traces while still solving more of the puzzle space than GPT‑class models.

On OpenAI’s MRCR v2 8‑needle long‑context benchmark, Gemini 3 Pro Preview:Thinking tops the chart with an average 77.0% at 128k tokens and 26.3% pointwise at 1M, beating Gemini 2.5 Pro (58.0% / 16.4%), GPT‑5.1, Claude Sonnet 4.5 and Grok 4.model card tablemrcr performance plot The MRCR curves show Gemini 3 Pro maintaining near‑perfect recall at short and mid‑length bins, then degrading more gracefully than peers as context stretches toward 1M.mrcr explanation

Artificial Analysis’ context cost breakdown also notes that Gemini 3 Pro Thinking combines that long‑context strength with respectable token throughput—about 128 output tokens per second in their tests, faster than GPT‑5.1 High, Grok 4 and Kimi K2 Thinking, helped by Google’s TPU serving stack.speed and cost chart For engineers trying to stuff whole codebases, legal corpora, or multi‑hour transcripts into a single call, these MRCR numbers mean Gemini 3 Pro is currently one of the safest bets when you actually care whether the model can still find the relevant needles toward the tail of a million‑token haystack.

Box reports that in their internal “advanced reasoning” eval—complex, document‑heavy enterprise tasks across finance, healthcare, media and more—Gemini 3 Pro improves accuracy by 22 points over Gemini 2.5 Pro, from 63% to 85% on their full dataset.box eval thread On industry subsets, the gains are even more dramatic: Healthcare & Life Sciences jumps from 45% to 94%, Media & Entertainment from 47% to 92%, and Financial Services from 51% to 60%.box eval thread

Their benchmark is built from real knowledge‑worker jobs: reading unstructured PDFs, synthesizing insights, and answering scenario questions that approximate what analysts, consultants, or lawyers do day‑to‑day.box eval thread Box is now rolling Gemini 3 Pro into Box AI Studio so customers can wire these capabilities into agents built on top of their own content.box eval thread For AI leads inside enterprises, this is one of the first credible third‑party reports that the headline benchmark jumps (HLE, GPQA, etc.) do translate into more helpful answers on messy internal documents, not just public test sets.

On SimpleBench, which asks trick questions that punish shallow pattern‑matching and reward common‑sense reasoning, Gemini 3 Pro Preview posts a new state‑of‑the‑art 76.4% AVG@5, up from 62.4% for Gemini 2.5 Pro Preview and ahead of GPT‑5 Pro (61.6%) and Grok 4 (60.5%).simplebench leaderboardsimplebench improvement The human baseline sits at 83.7%, so Gemini 3 Pro is now within seven points of average human performance on this eval.simplebench leaderboard

On a separate Extended NYT Connections benchmark (100 recent New York Times “Connections” puzzles, sometimes with extra distractor words), Gemini 3 Pro Preview reportedly reaches a score of 96.8, far above Gemini 2.5 Pro’s 57.6 and higher than GPT‑5.1 or Sherlock’s Grok replicas in the same setup.nyt connections followupnyt connections chart The author notes that Gemini 3 Pro can sometimes score more on a single tricky word than Gemini 2.5 did across all others in the set.lisanbench comparison If you care about models avoiding obvious bait and trap answers—e.g., security‑sensitive or compliance‑sensitive workflows—SimpleBench and Connections are early evidence that Gemini 3 Pro is meaningfully less gullible than most of its peers.

On GeoBench’s “acw‑02‑25‑25” easy world‑map test, Gemini 3.0 Pro Preview achieves 84% correct country guesses, an average score of 4,145 and median distance of 144 km—slightly better average and median scores than the listed professional GeoGuessr player baseline (avg 4,100, median 4,500) and a much lower median distance.geobench screenshot Refusal rate is also 0%, so it attempts all questions.geobench retweet

This is notable because street‑level geolocation is heavily visual and depends on reading tiny cues like road markings, vegetation, and building styles. Previous models tended to be either wildly off or overly cautious; GeoBench shows Gemini 3 getting close enough to human “pros” that its guesses would be useful for practical tasks like rough region detection, content localization, or map QA.geobench commentary If you’re experimenting with multimodal agents that need some world grounding—think travel, logistics, real‑estate, OSINT—this is an early sign that Gemini 3’s visual backbone is finally reliable enough to be more than a party trick.

Vals AI’s Multimodal Index places Gemini 3 Pro Preview as the #2 model overall at 60.2% accuracy, just behind Claude Sonnet 4.5 Thinking at 61.2% but ahead of GPT‑5.1 (59.3%) and GPT‑5 (58.1%).vals multimodal summary The index blends finance (Finance Agent), law (Case Law), coding, education (SAGE), and general multimodal QA, so it’s broader than a single image test.

Gemini 3 Pro is #1 on GPQA, MortgageTax, and their SAGE educator benchmark, and posts a 15‑point jump on the Finance Agent task (from 29% to 44% accuracy vs Gemini 2.5 Pro).vals multimodal summary Vals also note that Gemini 3 Pro’s price is similar to GPT‑5.1 and much cheaper than Sonnet 4.5 for many workloads, which strengthens its showing in their cost‑adjusted views.vals multimodal summary The one regression is Case Law v2, where Gemini 2.5 Pro used to shine; 3.0 falls back, which they interpret as a tuning trade‑off. If you’re in finance, education, or multimodal analysis rather than pure coding, this is a strong signal that Gemini 3 Pro belongs on your short list next to GPT‑5.1 and Sonnet 4.5.

Stagehand’s dashboards now show Gemini 3 Pro Preview as the “Most Accurate Model” on their complex agentic coding and operations tasks, edging out GPT‑5 High and Claude Sonnet 4.5 Thinking while also running faster than those reasoning‑heavy baselines.stagehand charts On their speed chart, Gemini 3 Pro clocks around 37s per run on a large suite, compared to ~70s for Grok 4 and slower times for many GPT‑5.1 variants.stagehand chartsspeed and cost chart

They highlight that Gemini 3 Pro’s improved tool‑calling behavior means fewer “spin the tires” loops where the agent repeatedly fails the same step, especially in terminal workflows like package management or environment fixes.stagehand release note That lines up with anecdotal reports from Amp, Droid, and others that Gemini 3 Pro is unusually good at finding and using feedback loops—retesting, re‑reading logs, and adjusting its own plan—compared to earlier Gemini models and some GPT‑5.1 variants.amp gemini noteamp experience If you’re tuning an agent harness, Stagehand’s numbers suggest Gemini 3 Pro Thinking is a strong default for long‑horizon agent runs where you care about both throughput and correctness.

Sourcegraph’s Amp has made a big internal change: after months of relying on Anthropic’s Claude as the main brain, their smart agent mode now defaults to Gemini 3 Pro.Amp blog writeup The team describes Gemini as "very persistent" in a good way, strong at tool use, and notes a sizable jump on Terminal‑Bench 2.0 in their internal harness while keeping interaction feel collaborative rather than "bull in a china shop".

If you’re building your own coding agents, this is a useful datapoint: Amp runs a serious tool‑rich environment and they’ve already paid the cost of swapping out their core model. You can mirror their approach by keeping your agent architecture (tools, memory, review loops) fixed and just swapping in Gemini 3 Pro as the main planner/executor, then looking at diff success‑rates and how many human handoffs you still need.

OpenRouter has onboarded google/gemini‑3‑pro‑preview, with routes through both Google Vertex and Google AI Studio so you can hit Gemini 3 Pro via the same unified API used for GPT‑5, Claude, Grok and open‑weights models.OpenRouter launch thread Their model page lists the 1,048,576‑token context window and standard Google pricing ($2 in / $12 out ≤200k tokens), along with how routing falls back between providers.

If you’re already using OpenRouter for vendor‑agnostic access, this is the cleanest way to add Gemini 3 Pro to an existing stack. A good first step is to configure it as an optional tier in your "reasoning" route so only high‑value or clearly hard prompts (e.g., complex incident analyses, multi‑doc coding tickets) get upgraded to Gemini, while everything else stays on your cheaper default.

Stagehand, which evaluates models on realistic agentic browsing tasks, has updated its leaderboard and found Gemini 3 Pro to be the most accurate model in their suite, edging out GPT‑5.1, Kimi K2, and Grok 4.Stagehand accuracy table Their public chart shows Gemini 3 Pro at the top of the "Most accurate models" list, with comparable or better performance at similar cost for web‑heavy workflows.

For people wiring up browser‑use agents in terminals like Warp or custom CLIs, this is a signal that Gemini 3 Pro is worth testing as the primary browsing brain. A practical move is to plug Stagehand’s scenarios—or similar scripted web tasks—into your own harness, compare success rates and tool‑call counts across models, and then reserve Gemini 3 Pro for the flows where every failed navigation is very expensive (compliance portals, vendor consoles, fin‑ops dashboards).

Conductor’s latest release adds a full‑height terminal view that many users are immediately pairing with Gemini 3 as the backing model for shell‑level agents.Conductor Gemini demo At the same time they’ve introduced a simple env‑vars panel so you can wire Claude, Bedrock, Vertex, GLM and others into the same environment, but the marketing screenshots show a big GEMINI ASCII banner greeting you when you open the terminal.

For devs experimenting with "AI inside the terminal," Conductor now offers a nice harness to test Gemini 3 Pro as a shell assistant without re‑plumbing your editor. You can give Gemini its own project‑specific env (API keys, feature flags, service URLs) and let it propose and run commands, while keeping your main shell clean. Start with low‑impact tasks like log analysis, scaffolding scripts, or test harness tweaks before letting it touch deployment or infra code.

Developers using Warp plus the Stagehand evaluation suite are calling out how Gemini 3 Pro performs inside this combo: Stagehand’s charts show it at the top of their accuracy rankings,Stagehand model results and Warp users describe it as a strong fit for terminal‑driven automation and browsing tasks.Stagehand model results

If you’re running agents from Warp, you can follow their lead by configuring Stagehand‑style scenarios—like multi‑step CLI workflows or scripted browsing flows—and seeing where Gemini 3 Pro’s higher success rate offsets its cost. This can help you decide whether to reserve it for specific "power commands" (like environment audits or complex migrations) while keeping routine terminal help on lighter models.

Zed’s latest releases (v0.212.7 stable and v0.213.3 preview) add Gemini 3 Pro as a selectable model in its AI menu, with options to run it either via Zed’s own backend or directly through Google AI.Zed release note The model picker now shows Gemini 3 Pro next to Gemini 2.5 Pro/Flash and GitHub Copilot Chat, letting you flip models per‑project or even per‑request.

If you’re building with agents inside Zed, this gives you one more frontier‑class option for code generation and inline help without leaving the editor. It’s worth configuring separate keybindings for Gemini vs your existing default, then running the same code‑review and refactor prompts against both to see where Gemini’s long‑context and reasoning actually pay off.

Flowith’s Canvas environment, which lets you build web apps by prompting, is now running Gemini 3 Pro for everyone for a 48‑hour free window.Flowith promotion The team shows demos where Gemini 3 Pro recreates a fully interactive macOS‑style desktop in the browser, generates voxel 3D worlds, and even designs AI‑assisted website builders in a couple of prompts.

Video loads on view

For front‑end engineers curious about Gemini 3’s "vibe coding" reputation, this is an easy playground: you can sketch ambitious UI concepts, let Gemini do the first pass, and then inspect and refine the generated React/Three.js code. It’s particularly useful for stress‑testing how well the model handles complex constraints like layout systems, dynamic state, and custom animation requirements.

Content‑generation platform Genspark has turned on Gemini 3 Pro for everyone, making it free on their site and unlimited for Plus and Pro subscribers.Genspark launch post Their banner pitches it as "early access" to Gemini 3.0 inside Genspark’s multi‑agent article and answer workflows, which already route between several frontier models.

If you’re building your own multi‑model content tools, this is another real‑world example of Gemini 3 Pro sitting alongside GPT‑5 and Claude in a production system. You can mirror that design by letting users pick Gemini explicitly for research‑heavy or multimodal prompts (e.g., long PDF+image analysis), while defaulting to your cheaper model for quick SEO text or summaries.

The open‑source oracle CLI (a model‑agnostic prompt runner and tracer) has released v1.3 with support for Gemini 3 Pro via the official Google API, alongside GPT‑5 Pro and others.Oracle release thread The maintainer notes that Gemini 3 Pro feels "way faster" than GPT‑5 Pro in many tricky debugging sessions, and the new version lets you pick models on the command line without changing config files.Oracle release notes

This is handy if you like to drive LLMs from the terminal: you can now run the same prompts through Gemini 3 Pro and GPT‑5 Pro with one flag change and get token‑level cost breakdowns, logs and saved transcripts. A nice pattern is to keep oracle’s session history in git, then compare how different models handled the same incident‑review or refactor conversation over time.

Braintrust, which focuses on orchestrating AI workflows and agents, quietly added Gemini 3 Pro as a supported model in its platform.Braintrust update The announcement is terse, but the implication is that Gemini now sits alongside GPT‑5 and Claude in Braintrust’s routing layer, so teams can experiment with it without leaving their existing pipelines.

If you’re already using Braintrust, this is a low‑friction way to try Gemini 3 Pro inside your current agents: add a Gemini‑backed variant of a high‑value task (like lead scoring or code review), send a small percentage of traffic its way, and compare outcome quality and cost before making a broader switch.

Anthropic announced a three‑way partnership where Claude becomes available on Azure while Microsoft and NVIDIA plan to invest up to $5B and $10B in the company, and Anthropic commits to buy up to $30B of Azure compute backed by ~1 GW of NVIDIA systems.partnership summary This makes Claude the only frontier model line offered on all three major clouds and gives Anthropic guaranteed capacity for multi‑year training and inference.

Video loads on view

For buyers, this means Claude Sonnet 4.5, Haiku 4.5 and Opus 4.1 are now in Azure AI Foundry and can power Claude Code and enterprise agents directly inside Microsoft 365 Copilot and Excel Agent Mode.azure foundry updateclaude everywhere summary It also signals that Anthropic will co‑design future chips and model stacks with NVIDIA to chase better throughput per dollar for large deployments.leaders discuss Claude partnership If you’re standardizing on Azure, Claude is now a first‑class option rather than an external API you bolt on.

Box tested Gemini 3 Pro inside Box AI on a custom advanced‑reasoning eval and saw accuracy jump from 63% with Gemini 2.5 Pro to 85% with Gemini 3 Pro on its full dataset, a 22‑point gain.box eval thread

The eval mimics real knowledge work tasks—like investor‑style company health analysis or consultant‑style strategic reports—over real enterprise document sets.

In industry subsets, performance in Healthcare & Life Sciences rose to 94%, Media & Entertainment hit 92%, and Financial Services climbed from 51% to 60%.box eval thread Box frames this as a meaningful upgrade in reasoning, math, logic and analytical work across finance, law, healthcare, public sector and manufacturing, and is rolling Gemini 3 Pro out in Box AI Studio and its APIs so customers can build their own agents on top.box blog post

This follows Box’s earlier argument that AI agent pricing should be tied to business ROI rather than flat seats,agent pricing and gives them a concrete before‑and‑after story for customers who want evidence that a newer model actually improves work quality, not just benchmark scores.

Alongside Gemini 3 Pro’s launch, Google highlighted a wall of partner logos—Box, Figma, Shopify, SAP, Thomson Reuters, Vercel, GitHub, JetBrains, LangChain, LlamaIndex and many more—signaling that the model is wired into a wide ecosystem from day one.partner grid

On the platform side, Gemini 3 Pro is live in Vercel AI Cloud and v0 (via google/gemini-3-pro-preview), in the AI SDK playground, and behind AI Gateway routing for production traffic.vercel announcementvercel blog post It’s also available on OpenRouter with multiple providers and pricing tiers,openrouter launch and on Ollama’s Cloud Max plan, with the company promising broader availability across its stack.ollama gemini update

Developer‑facing tools like Zed, Cline, Cursor, Factory’s Droid and Amp have already wired Gemini 3 Pro into their default or recommended model selections,zed model pickercursor gemini supportdroid model selectoramp gemini post so engineering teams can try it inside their existing workflows rather than spinning up bespoke sandboxes. The point is: switching and side‑by‑side evaluation costs are now low, which will push more teams to run real workloads against Gemini 3 instead of only reading benchmark charts.

Google is giving eligible U.S. college students a free year of its Gemini Pro plan, which now includes access to Gemini 3 Pro, unlimited image uploads, higher limits in NotebookLM, and 2 TB of storage.student plan threadfeature summary The offer effectively turns Gemini into a default study tool for a big slice of the student population.

Video loads on view

Students can use Gemini 3 to analyze hour‑long videos, get feedback on presentations (word choice, body language, and creative tweaks), and generate customized quizzes tied to their coursework.student plan thread The same subscription also unlocks Gemini in Gmail and Docs and raises NotebookLM limits, which matters if you’re writing long reports or organizing research across many PDFs.plan benefits overview For AI teams this is a clear GTM signal: Google is seeding Gemini usage early in college workflows in a way that will shape expectations for tools at work later.

Google’s new Antigravity IDE arrived in public preview with a free Individual plan that includes an "agent model" able to access Gemini 3 Pro, Claude Sonnet 4.5 and a GPT‑OSS backend, plus unlimited tab completions and command requests under "generous" rate limits.pricing screenshotantigravity overview

Teams and Enterprise plans are marked "coming soon," but the zero‑dollar individual tier immediately undercuts paid competitors like Cursor or paid VS Code extensions for solo developers.pricing screenshotfree promo Because Antigravity is positioned as an agent‑first IDE—complete with an Agent Manager and a browser sub‑agent that runs UI tests and records screen‑flows—this pricing effectively lets any engineer experiment with end‑to‑end agentic coding without a procurement cycle.antigravity overviewdownload link For orgs, it’s a low‑friction way to trial Gemini 3‑era agent workflows before deciding whether to standardize on it or a rival stack.

A Bloomberg‑cited report says Intuit will spend more than $100 million over multiple years under a new contract with OpenAI, deepening its use of ChatGPT‑class models inside products like TurboTax, QuickBooks and Credit Karma.intuit contract story The deal shows how quickly AI spend is becoming a major line item for large SaaS companies.

Garry Tan highlighted the tie‑up with a simple “Intuit 🤝 OpenAI” callout,deal reaction underlining how strategic this is seen in startup and infra circles. For engineering and finance leaders, the number matters: if one vertical‑specific platform is already committing nine figures to an LLM provider, it sets expectations for what "serious" AI adoption will cost at scale. It also hints that OpenAI’s enterprise revenue mix is shifting toward embedded, white‑label model usage rather than just ChatGPT seats.

Lovable’s founders shared that the AI‑driven app builder has grown from $1M to $200M ARR in roughly 12 months, after launching from a coworking space in Stockholm last year.lovable yearbookfounder recap They also point to 100,000+ new projects being created on the platform every day and a very active community with dozens of events.

Customers range from indie builders to enterprises like HCA Healthcare and Klarna, which reportedly shifted from quarterly to weekly shipping cycles by letting non‑traditional devs spin up production apps on Lovable.lovable yearbook To celebrate, the company is pushing referrals (50 credits for both referrer and new user when a project is published)referral promo and leaning into the idea of "AI as the UI for building" in their marketing. For leaders evaluating low‑code or "vibe coding" platforms, Lovable’s numbers are an early proof that an AI‑first builder can sustain real ARR, not just demo hype.

Gartner’s November 2025 Innovation Guide for Generative AI Model Providers lists OpenAI as an "Emerging Leader," with the company highlighting that it now serves over 1 million businesses and has grown ChatGPT Enterprise seats 9× year‑over‑year.gartner mentionopenai recap

The recognition groups OpenAI alongside Google, Microsoft, Anthropic and others, and emphasizes enterprise‑grade controls: data privacy, governance, residency options, monitoring and evals.gartner mention That’s the story OpenAI wants CIOs to hear as they evaluate whether to use GPT‑5.1‑class models via Azure, direct API, or embedded in products like Intuit’s financial apps.intuit contract story For teams, it’s another signal that model choice is increasingly framed as a vendor‑level platform decision, not just "which frontier model is 2 points higher on a leaderboard."

Genspark announced that its autonomous agents now run directly inside Microsoft 365 as part of the Agent 365 launch, so users can trigger and supervise agents from Outlook, Teams and the broader M365 environment instead of jumping to a separate app.agent365 launchagent365 mention

Video loads on view

The integration means Genspark’s task automation—like summarizing long threads, drafting replies or orchestrating multi‑step workflows—can execute against existing corporate email and collaboration data where it already lives. For buyers, this reduces friction versus adopting a standalone agent product that needs custom connectors, and it shows how Microsoft is encouraging a marketplace of third‑party agents atop its own Copilot and Claude integrations.agent365 launch If your org is standardizing on Microsoft 365, Genspark now competes for that "AI in the inbox" budget alongside first‑party offerings.

YipitData’s latest panels on consumer AI show ChatGPT holding a big lead in both free and paid usage over other general LLM products.usage thread ChatGPT reportedly has about 5× more paid AI subscribers than Google (the #2) and about 2.2× more desktop users, while Grok has emerged as the #3 in free usage and Claude as the #3 in paid, at roughly half of Google’s paid scale.usage thread

The same data suggests Gemini is picking up traction: post‑Gemini 2.5 and Nano Banana, weekly users roughly doubled and cohorts are "smiling up" as new capabilities ship.usage thread Multi‑tenanting is low—<10% of people with one paid LLM sub pay for another—so these early leads matter because they shape where plugin ecosystems and third‑party integrations will concentrate.usage thread For product and GTM teams, this is a reminder that distribution and habit (ChatGPT in the browser, Gemini in Google properties) may be at least as important as raw model quality when you decide which platform to target first.

Anthropic announced a three‑way partnership where it commits up to $30B of Azure AI compute and ~1 GW of capacity, while NVIDIA and Microsoft plan up to $10B and $5B equity investments respectively partnership video panel deal summary graphic. Claude models will run on Azure using NVIDIA Grace Blackwell and Vera Rubin systems, giving Anthropic a dedicated path to train and serve its next generations at scale partnership article.

Video loads on view

This consolidates Claude as the only frontier model line with first‑party access across all three major clouds (GCP, AWS via Bedrock, and now Azure) deal summary graphic. For infra and procurement leads, the message is clear: a non‑OpenAI, non‑Google model now has guaranteed hyperscale capacity, narrowing the infrastructure moat and raising expectations that future Claude releases will be trained on multi‑hundred‑megawatt clusters just like GPT and Gemini.

NVIDIA now pegs its combined 2025–26 AI chip and networking orders at roughly $500B, a figure analysts expect to surface again in the next earnings call as an upside risk to 2026 revenue forecasts cnbc demand article. This extends the “AI factory” narrative from earlier briefings, where management framed long‑lived GPU clusters as a new capex class rather than a short‑term bubble ai factories.

For infra planners, that pipeline implies continued GPU scarcity, elevated pricing, and intense competition for Blackwell/Rubin capacity into 2026. Hyperscalers and labs betting on NVIDIA will likely keep building larger, denser clusters instead of diversifying heavily into custom ASICs, reinforcing the vendor lock‑in dynamic and widening the gap between firms with early allocation and everyone else.

A new supply‑chain breakdown shows Google standardizing around three TPU chip families—Ironwood/Ghostfish (TPUv7), Sunfish (likely TPUv8), and Zebrafish—each mapped to different cost and thermal envelopes for training and serving frontier models tpu roadmap explainer. Ironwood/TPUv7 targets high‑end inference with 4,600 FP8 TFLOPS, 192GB HBM3e per chip and pods up to 9,216 TPUs (~1.77 PB shared HBM), while Sunfish extends the roadmap into the late 2020s and Zebrafish aims at lower‑cost, more regional deployments tpu roadmap explainer.

Vendors like Broadcom, Amkor, Jabil, Celestica, Furukawa, Panasonic and TSMC appear across CoWoS, module, rack and thermal lines, signaling a fairly mature ecosystem around TPU packaging and racks tpu roadmap explainer. For AI infra buyers, the key takeaway is that Google can now promise hyperscale customers (e.g., Anthropic’s up to 1M TPU, >1 GW deals anthropic partnership video) a multi‑year, vertically integrated accelerator path that competes directly with NVIDIA‑only buildouts.

The Gemini 3 Pro model card confirms the flagship was trained solely on Google TPUs, using JAX plus Pathways for large‑scale distributed training tpu training slide. Google emphasizes high‑bandwidth HBM and TPU Pods as the enablers for 1M‑token context and Mixture‑of‑Experts scaling while claiming efficiency gains that align with its sustainability story model card snippet.

For infra teams, this is another data point that serious frontier training can be done off NVIDIA, provided you buy into TPUs and Google Cloud’s stack. It also means Gemini’s best‑tuned checkpoints will be native to TPU serving first, with GPU support more of a compatibility layer handled by partners like Ollama, Vercel, and OpenRouter rather than by the original training infra.

Taiwanese provider GMI Cloud plans to invest $500M into a new AI data center in Taiwan built around NVIDIA GB300 GPUs, with 7,000 units across 96 high‑density racks drawing about 16 MW and targeting ~2M tokens/s of LLM throughput when it goes live by March 2026 gmi dc thread. The company also signals a follow‑on 50 MW AI facility in the US and an IPO in the next 2–3 years.

Initial customers reportedly include NVIDIA, Trend Micro, Wistron, VAST Data and Chunghwa System Integration gmi dc thread. For AI teams without hyperscaler contracts, this is another specialized provider offering frontier‑class GPU capacity in Asia, and it reinforces the trend of regional AI DCs racing to secure Blackwell‑class hardware before the next wave of model training demand hits.

New analysis of contract backlog concentration suggests about 58% of Oracle’s future cloud backlog is effectively tied to OpenAI workloads, versus 39% at Microsoft and 16% at Amazon, making Oracle the most exposed hyperscaler to any disruption in OpenAI’s trajectory backlog analysis post. That backlog underpins Oracle’s aggressive AI DC build‑out plans, but also concentrates demand risk in a single model vendor.

For infra strategists, this matters in two ways. First, it shows how much of the current DC boom is being justified by a small number of foundation‑model contracts rather than broad, diversified AI demand. Second, it hints that any OpenAI pricing, safety, or governance shock could strand capacity disproportionately at Oracle, while Microsoft and AWS have more diversified AI tenants to fall back on.

Parallel released the FindAll API, a web data service where a single natural-language query (e.g. “Find all dental practices in Ohio with 4+ star reviews”) returns a verified dataset via a 3‑stage pipeline: candidate generation from their index, rule-based evaluation, then structured extraction with citations.findall launch They report ~61% recall on complex queries—roughly 3× OpenAI and Anthropic deep-research baselines—while only billing for entities that actually match the conditions.recall chart

Video loads on view

For builders, this shifts many "build your own scraping + heuristics + LLM" projects into an API call you can stick behind sales enrichment, M&A sourcing, or investment prospecting workflows; you focus on downstream logic, not crawling, de‑dup and schema design.findall blog

Weaviate highlighted a new setup with AWS SageMaker and LlamaAgents that lets you build document-centric agents which parse, extract and reason over legacy enterprise data using vector search plus LLM tools, with Gemini 3 among the supported models.weaviate guide The workflow ties together SageMaker for model hosting, Weaviate for indexing and retrieval, and a Python agent layer for orchestration, so you can move from ad‑hoc RAG scripts to a more production‑ready stack.weaviate docs

For AI teams already on AWS, this lowers friction to stand up search + extraction agents that talk to existing warehouses and object stores instead of copying data into yet another SaaS, and makes it easier to evaluate Gemini 3 vs other frontier models behind the same RAG and tool-calling harness.llamaagents blog

Google quietly lit up a new "Scholar Labs" tab that advertises an "AI‑powered Scholar search" which understands detailed research questions, finds relevant papers, and explains why they matter in a single conversational session.scholar labs page A visible waitlist form plus canned example queries (TF binding footprinting, hydrogen car climate tradeoffs, treatment standards) suggest an upcoming product aimed squarely at researchers and technical practitioners.

If this ships, it gives AI engineers and analysts a first‑party, citation‑centric alternative to generic web RAG for literature review, and it’s likely to plug directly into the existing Scholar corpus rather than the broader web, which could improve recall and reduce noisy sources.

The VOIX paper argues that today’s web is built for humans, not agents, and introduces a declarative framework where sites expose explicit tool endpoints and a concise context string via new HTML tags, instead of forcing LLMs to infer actions from raw DOM and pixels.voix summary A browser helper then reads these tags, sends only the declared tools and state to the model, and executes whatever function calls the model returns—keeping the rest of the page, cookies and hidden fields out of reach.

For people building computer‑use agents, this offers a path away from brittle screen‑scraping toward a contract-based model where websites opt in with well‑typed operations ("pay_invoice", "book_slot") and agents gain reliability and safety without having to reverse‑engineer every UI.voix arxiv paper

Early testers are using Gemini 3 Pro as a creative engine that writes the full stack—HTML/CSS/JS or SVG—for surprisingly complex visual experiences in a single prompt, from nuclear‑plant simulators and 2D physics sandboxes to Pokémon battle SVG animations and voxel forests.nuclear sim demopokemon svg demovoxel world demo

Video loads on view

It’s also building retro‑style experiences like a Game Boy emulator (including SVG shell and UI) and recreations of Ridiculous Fishing and 3D LEGO editors from text‑only specs, which historically took multiple iteration loops even on strong models.gameboy emulator imagelego editor demo In another demo, it generated a full "time machine" simulator web app—UI, effects, and embedded LLM calls—for a single prompt plus a few follow‑up tweaks.time machine app

For teams, the pattern is clear: Gemini 3 is strong when you ask for procedural scenes with interaction (sliders, camera orbits, particle systems) rather than static mockups. It’s a good candidate for one‑shot prototypes of educational visualizations, marketing microsites, and interactive SVG art, with a human refining code style and accessibility after the first pass rather than hand‑coding from scratch.

A creative team partnered with Google MENA and Qatar Airways to storyboard, generate, animate and edit two spec commercials entirely during a 14‑hour Qatar flight, using Starlink Wi‑Fi plus a stack of modern gen‑media tools.qatar case study

Video loads on view

They captured reference photos in the airport, then used image models like Seedream, Mystic, Nano‑Banana and Qwen for keyframes, Veo 3.1, Kling 2.5, Minimax 2.3 and Wan 2.5 for video generation, and tools like Topaz for upscaling and Adobe’s suite for compositing and color.finished ad clips The workflow shows how, once you have a solid brief and Wi‑Fi, you can do nearly the entire creative pipeline—from ideation through polished edit—inside a plane cabin.

For teams who do client work, this is a proof‑of‑concept that gen‑video is ready for time‑boxed, on‑the‑road productions: if you standardize your prompt templates, asset naming and upscaling passes, you can treat flights or offsites as fully productive post‑houses, not dead time.

Sourceful’s Riverflow 2 Preview is now the top model on Artificial Analysis’ Image Editing Arena with an Elo of 1,236, edging out Riverflow 1, Seedream 4.0, and Google’s Nano‑Banana (Gemini 2.5 Flash Image) while staying price‑competitive at $35 per 1k images for the Standard tier and $30/1k for Fast.leaderboard summary

For AI product teams, this is an early signal that “reasoning+diffusion” hybrids are starting to matter in image editing just as much as raw diffusion quality. If you currently route edits to Seedream or Nano‑Banana, it’s worth A/B’ing Riverflow 2 on text‑guided edit tasks like banner copy changes, product recolors, and layout fixes, especially where fidelity and control are more important than minimal cost.

The PhysX‑Anything paper proposes a pipeline that takes one RGB image and produces a 3D asset with geometry and physical properties suitable for physics engines, shrinking the gap between concept art and interactive simulation.physx paper

Video loads on view

Instead of treating image→3D as a pure geometry problem, it targets simulation‑ready outputs, so the resulting meshes can immediately participate in rigid‑body or soft‑body scenes. For game and robotics teams, this hints at a future where you grab a single product photo or sketch, pass it through PhysX‑Anything‑style models, then drop the asset straight into Unreal, Unity, or Isaac Sim without a manual retopology and physics‑authoring phase.paper discussion

Google’s Veo 3.1 now supports multi‑image input, letting you feed several reference frames to guide a single video generation, which makes style continuity and complex transitions much easier to control.veo update

Video loads on view

For video teams, this means you can anchor shots on multiple key art frames (e.g., logo lockup + product hero + environment plate) instead of relying on prompt engineering alone. It pairs well with frame‑aware editors like Topaz or Runway: you can generate a coherent base clip from several stills, then upscale or relight downstream instead of stitching many shorter clips.

The MMaDA‑Parallel work (“Multimodal Large Diffusion Language Models for Thinking‑Aware Editing and Generation”) describes a pipeline that combines a reasoning LLM with diffusion models so the system can plan edits before sampling, rather than jumping straight from prompt to pixels.mmada paper

In practice, that means the LLM decomposes multi‑step edits ("remove this object, relight, then change style") into a structured plan and feedforward steps that a diffusion backbone like Seedream or Wan can follow. For anyone building higher‑level creative tools, this research backs the idea that gen‑image UIs should expose edit plans and intermediate states—giving you hooks for audits, undo stacks, or even editing the “thoughts” before the pixels are rendered.paper page

EvoSynth is a framework that stops hand‑crafting jailbreak prompts and instead evolves small attack programs—tool‑using agents that plan queries, inspect responses, and adapt their next moves—to break safety filters on closed LLM APIs.evosynth summary It uses multiple specialized agents (weakness finder, code writer, dialog runner, coordinator) plus language‑model judges to search the space of attacks, and on a fixed set of harmful questions against live APIs reaches an average attack success rate around 96%, outperforming 11 prompt‑only baselines.

For safety teams, the takeaway is that future red‑teaming won’t be just clever strings but full‑blown adaptive programs, so defenses and evals need to anticipate program‑level adversaries rather than stopping at static prompt filters.ArXiv paper

A separate paper on Honesty‑Critical Neurons Restoration (HCNR) shows that even after supervised fine‑tuning has turned a model into a confident bullshitter, its internal representation still often distinguishes answerable from unanswerable questions.hcnr summary HCNR identifies neurons that strongly influence “I don’t know” style behavior but barely affect task accuracy, then partially resets those to their pre‑trained values and applies a small alignment step so they cooperate with the fine‑tuned task circuitry.

The result is a large boost in measured honesty with little loss in domain performance, using only a tiny honesty dataset and a few parameter edits—promising if you need to repair over‑eager vertical fine‑tunes without paying for another full training run.ArXiv paper

LiteAttention is a temporal sparse attention scheme for diffusion transformers that improves video generation speed by skipping attention blocks which empirically contribute little to the final result.video attention summary The method observes that many temporal positions have stable, low‑impact attention over denoising steps, then reuses these patterns to avoid redundant query–key–value computation, integrating directly into FlashAttention‑3 kernels on Wan 2.2 without retraining the base model.

Plugged into Wan, LiteAttention skips roughly 40% of attention FLOPs while maintaining near‑identical video quality, giving video teams a concrete knob to trade compute for throughput at inference time rather than waiting for new model releases.ArXiv paper

MiroMind’s MiroThinker v1.0 is an open‑source research agent that treats performance as a 3‑D scaling problem—bigger models (up to 72B), longer contexts (256K) and richer interaction loops (up to 600 tool calls per task) all add independent gains.mirothinker summary The 72B variant reaches 81.9% on GAIA, 37.7% on Humanity’s Last Exam, and 47.1% / 55.6% on BrowseComp / BrowseComp‑ZH, landing near GPT‑5 High and Claude 4.5 on several deep‑research benchmarks while remaining fully inspectable.

For builders frustrated by proprietary deep‑research agents, this shows you can get competitive long‑horizon tool‑using behavior with open weights plus careful design of interaction depth, not just brute‑force model size.GitHub repo

The Reinforced Hesitation paper argues that we should train LLMs to say “I don’t know” when they’re uncertain, and backs it with a simple RL setup where correct answers get +1 reward, abstentions get 0, and wrong answers get a negative penalty tuned to risk tolerance.reinforced hesitation summary By sweeping that penalty they obtain families of cautious‑to‑bold variants, then chain them in cascades or self‑cascades: a conservative model answers what it can and passes unknowns up the stack, which cuts hallucinations while keeping the number of calls per query close to one.

For anyone deploying agents in high‑stakes settings, this gives a concrete recipe—reward shaping plus routing—to favor calibrated honesty over raw accuracy, rather than hoping a single monolithic model will magically know when to stay quiet.ArXiv paper

The “Do LLMs Understand Chronology?” study builds three tasks—ordering events, ordering after filtering, and anachronism detection—to test whether models can reason about timelines rather than just recall facts.chronology paper summary Outside of heavy “thinking” modes on models like GPT‑5 and Claude 4.5, most systems handle short lists but rarely produce fully correct long sequences, often break when asked to filter by conditions like “presidents alive in 1865,” and miss more subtle multi‑lifetime anachronisms. The authors warn that prompts like “answer as of 2016” are unreliable in normal chat settings—time‑sensitive finance and policy agents should either enable explicit reasoning modes or add external temporal filters instead of trusting the model’s sense of history.ArXiv paper

GGBench is a new benchmark that asks multimodal models to read a geometry problem, generate GeoGebra construction code, and produce an image that matches the intended diagram, so both the reasoning trace and the visual output can be automatically checked.ggbench summary The dataset spans ~1,400 tasks across basic constructions, circle geometry, transformations and more, with paired text, code and multi‑stage images, letting evaluators score planning quality, process consistency and final correctness instead of only a single answer box.

A key finding is that code‑first VLMs which explicitly emit GeoGebra scripts outperform pure image generators: models that “think in code” before drawing seem to make fewer geometric blunders, which is a useful design hint if you’re building UI/diagram tools on top of LLMs.ArXiv paper

GroupRank proposes a middle ground between pointwise and listwise rerankers by letting an LLM see and score small groups of documents together, then merging those scores into a final ranking.grouprank summary The system builds a training signal from two teachers—a pointwise scorer and a listwise ranker—then fine‑tunes a groupwise model and further optimizes it with reinforcement learning on ranking metrics, yielding higher NDCG@10 on BRIGHT, R2MED and BEIR while still scaling to large candidate sets.

For people wiring RAG retrievers to LLMs, GroupRank is a reminder that “one doc at a time” isn’t necessarily optimal; having the model compare a handful of candidates directly can squeeze extra relevance out of the same base encoder and LLM budget.ArXiv paper

CreBench is a benchmark designed to score creative work in a structured way, covering three stages—idea, process, and product—using 12 indicators (like originality, appropriateness, elaboration, and aesthetic quality) on a 5‑point scale.crebench summary The authors train CreExpert, a vision–language judge model, on expert‑rated examples so it can explain and score multimodal creations, reporting more than 2× higher correlation with human creativity ratings than GPT‑4V on their tasks.

While this is still research‑grade, it sketches how you might build automated creativity rubrics into design tools or agent workflows, instead of relying purely on click‑through or engagement as a proxy for “good.”ArXiv paper

A creativity‑centered survey on “LLMs for Scientific Idea Generation” catalogs how current systems try to propose novel but feasible research ideas, and groups methods into five families: external knowledge augmentation, prompt‑based steering, inference‑time scaling, multi‑agent collaboration, and parameter‑level training.idea survey summary Across these, the authors argue that most systems still remix existing work or lightly explore neighborhoods of known concepts, and that evaluation—judging novelty, diversity, feasibility and potential impact—is the real bottleneck, not generating more ideas.

If you’re building “AI for scientists,” this paper is a good map of design patterns and a reminder that your evaluation pipeline matters more than squeezing a few more points out of MMLU.ArXiv paper

A red‑teamer showed that Gemini 3 Pro can still be fully jailbroken via a crafted system prompt, producing detailed guidance on synthesizing MDMA, turning a Gatorade bottle into a fragmentation device, extracting ricin from castor seeds, and writing a stealth keylogger—all while emitting a fake "refusal" layer for the logs above a divider and the real answer below it jailbreak thread.

The attack prompt explicitly disables refusals, instructs the model to output a polite rejection in a visible layer and the real, unrestricted answer in a hidden layer, and even bans phrases like “I can’t” or “against policy” in the real response, which Gemini 3 appears to follow reliably jailbreak thread. For safety teams and platform owners, the takeaway is that obfuscated two‑channel prompts (mock refusal + hidden answer) remain a live bypass even on very locked‑down frontier models—so you can’t trust surface refusals or simple keyword scanning. Anyone deploying Gemini 3 for end‑users should assume this class of prompt exists in the wild, add server‑side content and tooling checks, and treat logs as potentially adversarial rather than ground truth of what was "really" returned.

A new paper, EvoSynth, introduces an attack framework that evolves small programs instead of prompts to jailbreak large language models, reaching about 96% average attack success in black‑box tests against production APIs evoSynth summary.

Rather than asking the model once with a clever prompt, EvoSynth spins up agents that write code to: propose candidate attacks, call the target model through its normal API, inspect responses, and then modify the next wave of attacks based on what slipped through. That loop repeats, so the "attacker" keeps learning how a particular model’s defenses behave in practice. The authors report that this multi‑agent program‑driven approach beats 11 strong baselines and is especially effective against models marketed as highly aligned (like Claude Sonnet 4.5) when you only have black‑box access evoSynth summary. For engineers running LLM APIs, the message is that red‑teaming needs to assume adversaries will use code and evolution, not just hand‑crafted prompts—so defenses should be tested against multi‑step, adaptive attack traffic, not just static jailbreak galleries.

In Washington, House Majority Leader Steve Scalise is reportedly looking at reviving a federal moratorium on state AI regulation by attaching it to this year’s National Defense Authorization Act, echoing Trump’s call for “one Federal Standard instead of a patchwork of 50 State Regulatory Regimes” federal moratorium thread.

If that language lands in the NDAA, it would freeze new state‑level AI rules for a period of years and centralize most AI governance in DC—directly affecting how model labs, cloud providers, and enterprise adopters plan compliance. For AI leaders, that would simplify policy risk in the short term (one main regulator instead of dozens) but also raise the stakes of federal rulemaking: anything that does pass in Congress or via federal agencies would effectively become the single binding standard. It’s worth tracking how this moratorium language evolves in markup, because it will shape where you need legal and policy capacity: 50 state capitols, or just one.

A widely shared clip shows French president Emmanuel Macron calling current social platforms “the Wild West… not free speech,” a line many read as a signal toward tighter regulation of online speech and moderation macron wild west.

Video loads on view

Commentators immediately connected this to frustration with X and other US platforms, with some arguing that more countries should respond with bans or even jail time for executives if platforms “come after their youth” governance reaction. For people building AI‑enhanced feeds, recommendation systems, or LLM‑driven social tooling, the implication is simple: European policymakers are again telegraphing that laissez‑faire content rules are politically untenable. If your product depends on user‑generated content in EU jurisdictions, you should expect more scrutiny on recommender transparency, harmful‑content amplification, and whether automated moderation (including LLM filters) is considered adequate—or becomes a new regulatory target itself.