GPT‑5.2 Thinking keeps near‑perfect recall to 256k – METR unifies strength scores

Following up on Unitree concert, where G1 units did Webster flips on stage, Unitree’s spec card confirms the G1 humanoid is about 132 cm tall, ~35 kg, runs ~2 hours on a quick‑release 9000 mAh battery, and starts at roughly $13,500 G1 spec image. That price puts a capable humanoid in the range of a mid‑range workstation or industrial arm, not a bespoke one‑off.

A long technical thread explains why G1 is intentionally short: smaller actuators and links cut bill of materials and shipping costs; a lower center of mass improves balance and tip‑over margin; reduced limb length lowers required joint torques, inertia, and impact energy in falls design rationale. Shorter limbs also mean less bending stress and smaller end‑effector position errors per joint error, which boosts durability and precision. For labs and startups, that combination—compact folding form factor (~690 mm when folded), onboard depth camera and 3D LiDAR, and a research‑friendly price—makes G1 a realistic platform for everyday experimentation rather than a museum‑piece robot. If you’re planning embodied AI work, this is probably the first mass‑market humanoid you can budget like a serious dev tool instead of a moonshot.

Tesla’s Optimus V2.5 humanoid showed up at xAI’s holiday party, walking into a crowded room and waving, with onlookers calling out its increasingly human‑like hands Optimus appearance. Compared to earlier factory‑floor clips, this is a more social setting, with the robot moving near people in normal clothes and lighting.

Video loads on view

For robotics teams, this kind of public, less‑staged appearance matters: it suggests Optimus’s locomotion, balance, and teleop/autonomy stack is now trusted enough for live events, not only tightly controlled demos. The emphasis on hands indicates Tesla is prioritizing general‑purpose manipulation—grasping, gesturing, and eventually tool use—over pure locomotion tricks. If you’re building competing platforms or applications on top of humanoids, this is another data point that the race is moving from can it walk to can it use its hands safely around people.

Unitree’s H1 humanoid robots paired with Go1 quadruped "robot dogs" are now shown patrolling real corridors as a coordinated team in China, not just in lab demos Unitree patrol clip. The setup is framed as a way to scale patrol hours without scaling headcount, with robots handling routine sweeps while human officers focus on judgment and de‑escalation.

Video loads on view

For AI and robotics engineers, this is a concrete example of multi‑robot deployment in a safety‑critical, human environment. It implies a working stack for perception, localization, and simple task coordination robust enough to trust on unscripted patrol routes, plus an operations model where robots and officers are explicitly divided by role. If you’re building embodied agents, this is a signal that customers are ready to pilot mixed human–robot security teams, not only novelty demos.

A Kyber Labs demo shows a robotic arm doing delicate, high‑throughput assembly of small electronic components onto circuit boards, the kind of precise, repetitive work usually associated with blue‑collar line jobs assembly clip. The arm rapidly places tiny parts with accurate alignment and without damaging them, underscoring how far contact‑rich manipulation has come.

Video loads on view

For people designing factory automation or robotics tooling, this is a reminder that the bottleneck is shifting from raw capability to integration and economics. If a system can already handle this level of dexterity, the open questions become: how quickly can you retarget it to new SKUs, and how do you fuse perception + planning + quality checks into a single, reliable agent loop? It’s also the kind of workload where learned policies and vision‑language models could start to replace pure hand‑tuned motion code.

A recent bipedal robot video shows a humanoid walking over a textured surface at speed, with subtle stumbles and recoveries rather than choreographed “dance” moves, and commentary highlighting progress in contact‑rich control at 100–500 Hz as the real milestone locomotion clip. The robot maintains balance through continuous small corrections instead of relying on perfectly repeatable trajectories.

Video loads on view

For control and RL engineers, this is the behavior that actually matters if you want robots in warehouses, factories, or homes. High‑frequency, low‑latency control loops that can handle friction changes, foot scuffs, and unmodeled dynamics are much closer to deployable reality than a single, rehearsed backflip. This kind of demo signals that more labs are focusing on robustness and rapid recovery rather than one‑off stunts, which in turn opens room for higher‑level planners and world models to sit on top of a reliable locomotion substrate.

A clip circulating in China shows a young girl crying as she says goodbye to a broken AI learning robot, with her father sharing the moment online emotional robot story. The device appears to be a dedicated educational companion, not a general tablet or phone, and the child’s reaction suggests it had become a day‑to‑day presence.

For product leads and alignment folks, this is a small but telling signal: in some markets, “AI robots” are already embedded enough in family routines that kids form emotional attachments. That raises the bar on reliability and repairability, but also on behavioral design—how these systems handle shutdown, failure, or replacement without causing distress. If you’re working on home robotics or child‑facing agents, this is a reminder that your off‑states and failure modes are as much part of the UX as your best demo.

A new MRCRv2 long‑context chart shows GPT‑5.2 Thinking sustaining almost 100% mean match ratio across 4‑needle tasks from 8k up to 256k tokens, while GPT‑5.1 Thinking decays from ~90% at 8k to ~45% at 256k mrcr chart.

For long‑horizon agents and retrieval systems this effectively makes 5.2 Thinking a "no‑drop" model across today’s practical context lengths, and explains why some builders expect it to surpass Opus 4.5 on METR once it’s fully evaluated there mrcr chart. The main constraint becomes cost and thinking latency rather than basic recall, so infra leads should start routing the most context‑heavy workloads to 5.2 while keeping cheaper models for short prompts.

After asking METR to publish runtimes and costs explicitly, independent analysts dug into the raw YAML and concluded that working_time and usd are per single full evaluation run (one of the 8 attempts), not totals or averages across all runs metr runtime request yaml discovery.

Using those fields, they infer that GPT‑5.1‑Codex‑Max takes about 2.6× longer working_time than Claude Opus 4.5 to complete the benchmark, while scoring worse but being roughly 7.6× cheaper per run, which helps explain why Anthropic feels faster for coding agents while OpenAI wins on raw dollar efficiency runtime comment cost comment. The takeaway for teams reading METR is that you can now treat time and cost as comparable per‑attempt quantities between models, but you still need to recompute effective price/performance for your own success‑rate targets rather than relying on the headline p50 horizon alone yaml interpretation.

ContextArena published a full MRCR breakdown for Gemini 3 Flash, showing how Base/Low/Medium/High reasoning modes trade cost for long-context recall at 128k–1M tokens, following up on Gemini3Flash MRCR which first highlighted Medium’s strong 128k AUC contextarena update.

At 8‑needle / 128k, Base hits 36.5% AUC for about $7 output cost, Low jumps to 54.5%, Medium to 69.2%, and High to 71.6% while costing ~$158, with similar but compressed gaps at 1M tokens contextarena update. Medium looks like the sweet spot: it nearly ties High on 4‑needle AUC (87.1% vs 85.5% at 128k) while using ~45% fewer output tokens, which is the practical setting builders should probably standardize on unless every extra point of recall is worth a big spend contextarena update.

New analysis finds the log of METR time-horizon scores correlates extremely strongly with other frontier benchmarks, with r≈0.96 vs Epoch Capabilities Index, r≈0.95 vs ARC‑AGI, and r≈0.91 vs GPQA‑Diamond on overlapping models metr correlation table.

Lower but still positive correlations appear even for softer tasks (e.g. writing), suggesting METR’s long-task horizon is mostly a proxy for general capability rather than a totally distinct axis metr correlation table. For engineers and analysts, this means you can often treat METR time-horizon as a single, comparable scalar of "how strong is this model overall" while still checking domain‑specific suites when they matter (e.g. code, math, tool use) arc-agi comment.

New WeirdML results put Gemini 3 Flash Preview at ~61.6% average accuracy across 17 synthetic‑but‑tricky tasks, essentially tied with GPT‑5.1‑high (~60.8%) and not far behind GPT‑5.2‑medium (~63.4%) weirdml chart.

The kicker is cost and code length: Flash clocks in at only $0.222 per full WeirdML run with the shortest average solution (149 lines of code), making it the clear price‑performance outlier among the frontier models listed weirdml chart. For engineers doing algorithmic or "toy math"‑style work at scale, that suggests Flash is already good enough to replace older GPT‑5.1‑tier models in many workloads while cutting both API bills and execution time.

OpenRouter’s live performance dashboards now show Claude Opus 4.5 responding faster in practice than Claude Sonnet 4.5, despite Opus being the larger, more capable model openrouter speed post.

That’s based on aggregated latency and throughput stats published on separate tracking pages for each model, which developers can inspect before choosing routing defaults (sonnet metrics, opus metrics ). For agents and long‑running coding sessions, this flips the usual "Sonnet for speed, Opus for depth" intuition and suggests many stacks can safely standardize on Opus without paying a latency penalty.

Builders are standardizing on a "Continuity Ledger" file for GPT‑5.2 Codex so long-running coding agents survive context compaction without losing the plot. The pattern keeps a single markdown ledger (Goal, Constraints, Key decisions, State, Done/Now/Next, Open questions, Working set) that the agent reads and updates on every turn instead of relying on raw chat history, and it’s auto-injected into all sessions via ~/.codex/AGENTS.md so you configure it once and forget it continuity prompt ledger details agents file docs. This has already enabled coherent 3‑hour Codex runs on large workspaces, and gives teams a template they can tweak per repo (for example tightening how uncertainty is marked or how plans sync with tool outputs).

The OpenCode ecosystem is quickly picking up the emerging AgentSkills spec: one developer had OpenCode implement the spec itself via prompting, then wired skills into their workflows so the agent can follow reusable, declarative capabilities instead of ad‑hoc prompts agentskills run. On top of that, the new "Open orchestra" plugin lets you define specialized agent profiles and orchestrate them into workflows, offering a lighter‑weight, lower‑token alternative to heavy "oh‑my‑opencode"-style setups for autonomous coding runs open orchestra demo plugin commentary.

A new example from the AI SDK community underlines how much agent quality depends on using provider-executed tools that the underlying models were trained on, like OpenAI’s webSearch or Anthropic’s code execution sandbox ai sdk example. In their setup, these tools are imported as typed functions (with structured inputs like userLocation) and wired into a ToolLoopAgent, so GPT‑5’s websearch or Opus 4.5’s code runner behave predictably even over long interactions, while client-executed tools (local shell, computer control) sit alongside them with the same type-safe interface.

Warp now exposes its terminal agent as a GitHub Action so teams can run repo-aware automations in CI, like scanning new issues and auto-labeling underwritten bug reports with "needs info" before a human even looks github action demo. The action is configured by referencing warp-agent-action in your workflow, passing a prompt plus API key, and letting the agent operate against the checked-out codebase; Warp also published a walkthrough showing how to set this up end-to-end for issue triage pipelines setup guide.

The Athas Code Editor shipped a syntax‑highlighted diff viewer and a new Pull Requests view, tightening the feedback loop between its coding agent and Git workflows diff viewer video pr view feature. The maintainer publicly thanked a core Rust contributor for the diff work and says they’ll "stop adding new features for a while" to concentrate on fixing agent bugs and editor issues, signaling a shift from rapid feature growth to hardening the agent-assisted coding experience athas praise roadmap note.

Video loads on view

Multiple power users are now calling Codex’s CLI UX its weakest link, arguing it should adopt Claude Code’s explicit modes for "plan", "ask to edit", and "edit without permission" instead of relying on carefully worded prompts ui critique codex plan rant. The complaint is that with Codex you’re often "begging" it not to touch files while it thinks, whereas Claude’s UI makes planning and approval a first-class workflow primitive, which matters when you’re letting agents loose on large, shared codebases.

One engineer reports finally getting "decent test coverage" by flipping their TDD workflow: write the implementation first, then ask an LLM to inspect the code and draft a battery of tests that they review for intent and fill in gaps test coverage story. This pattern turns the model into a test author rather than a spec follower, which seems to work well for legacy code where specs are fuzzy but concrete behavior is embodied in the current implementation.

A practitioner describes planning concrete technical design documents by convening a small "council" of models—GPT‑5 Pro (extended thinking), GPT‑5.2 Extra High via Codex, Opus 4.5 in Claude Code, GPT‑5.2 Codex Extra, and Gemini 3 Pro—and giving them the same review prompt plus full context, then synthesizing their feedback with Opus 4.5 design doc workflow. They repeat this 5–10 times as the spec evolves, treating agents less as single oracles and more as parallel reviewers whose disagreements surface edge cases and design risks before implementation starts.

Despite good results on bug fixes and small features, one engineer reports that LLMs remain poor at generating complex UI integration tests: they default to extremely verbose Playwright tests rather than concise parameterized cases, and are prone to hallucinating fixes when given only a failing stack trace ui tests reflection. They note they haven’t yet given their agent structured visibility into the browser (for example via a Playwright MCP server), so the current loop still requires them to manually inspect failures and explain the actual UI state back to the model.

Ben Goodger, now leading engineering for ChatGPT Atlas at OpenAI, argues that AGI needs a place to act, and that the web browser is the natural "body" where it can click, type, and manipulate real services via an Agent-to-User Interface (A2UI). Goodger quote He contrasts today’s chat-centric UIs with a future where agents use the browser as an environment, similar to how humans use it, making A2UI to agents what chat UI was to chatbots. A2UI framing

Video loads on view

In the longer interview, he talks through practical constraints like private data, untrusted content, and the need for a clear "stop button", which are exactly the concerns infra and security leads will need to resolve before letting browser‑embedded agents operate autonomously. YouTube interview For builders, the takeaway is that UX standards for action-taking agents will likely coalesce around the browser surface, not bespoke native apps, so investing in browser-centric tool APIs, sandboxing and permission models now is a future-proof bet.

Will McGugan ran the same agents through multiple CLIs and his ACP-based Toad terminal, surfacing very different UX trade-offs across Claude Code, Gemini CLI, OpenCode and AmpCode, following up on Toad update about its minimal UI and VS Code integration. Toad overview Claude’s own terminal sticks to scrollback mode: markdown doesn’t stream, large chunks appear only after a delay, and a color-handling bug wraps fractal output badly; even file autocomplete feels laggy because directory scans kick in only after the first character. Claude comparison

Video loads on view

Gemini’s CLI shows paragraph-level markdown streaming and permission prompts, but strips ANSI color once commands exit, so colorful TTY output that exists internally never reaches clients like Toad. Gemini comparison OpenCode uses an alternate-screen TUI similar to Toad and exposes agent “thoughts”, yet its file picker only shows directories until you type a character, flickers as its dialog width resizes with path length, and reduces ls output to a single column, suggesting PTY handling issues. OpenCode comparison AmpCode adds an ASCII sphere animation and a scrollbar, but keeps everything monochrome, lets its file picker flicker as it resizes, and even plays sounds when agents finish—details McGugan calls out as distracting for terminal users. AmpCode review

Video loads on view

By contrast, Toad leans on the Agent Client Protocol to share backends with these tools while exploring a different UX stance (alternate screen, richer layout, VT Code support) and McGugan’s thread effectively sets an emerging checklist for agentic CLIs: streaming markdown, stable layouts, color-preserving PTYs, low-latency autocomplete, and configurable affordances like sounds and scrollback. (Toad overview, VT Code note)

Several builders are explicit that the most useful interface for coding and automation agents is the terminal, not the traditional chat UI: Addy Osmani calls out Claude Code, Gemini CLI and Codex CLI as the real "power" interfaces for file cleanup, refactors, automation and bulk ops, flatly stating that "claude, gemini, codex > UI" for this work. CLI advocacy Thdxr keeps doubling down on a TUI-first OpenCode workflow, adding commands like opencode --port=4096 plus a one-shot "Open WebUI" action to pop up a browser view connected to the current terminal session, arguing that people who pit GUIs vs TUIs "can't do either". OpenCode webui command

Video loads on view

Warp pushes this further into automation by exposing its agent directly from GitHub Actions: a simple warp-agent-action step lets the agent label underwritten bug reports as "needs info" with full codebase context, moving agent UX into CI logs where developers already live. Warp GitHub setup McGugan’s broader point that "the agentic coding CLI is an entirely new class of app" and that ACP lets UI authors like him experiment at the presentation layer while reusing the same agent brains[ t:265|Toad positioning ] suggests an emerging standard: serious users will expect rich, scriptable CLIs and TUIs, with chat surfaces layered on top rather than the other way around.

MiniMax’s M2.1 model has been added to LMArena’s Code Arena, so builders can now pit it head‑to‑head against other frontier coding models in live web‑app tasks and vote in "Battle Mode" arena announcement. Following up on subagent usage, where M2.1 impressed with interleaved coding and sub‑agent orchestration, today’s move puts those claims under community scrutiny on shared benchmarks that test planning, scaffolding and debugging. Early users also highlight that M2.1 "brought major upgrades to design and visual quality" design comment and call it "a beast in design" when paired with Factory AI’s stack design praise, with at least one polished project explicitly "Built with M2.1" shipping already built with m2-1.

Video loads on view

If you care about which model to standardize on for coding agents, Code Arena now lets you see how M2.1 actually behaves on multi‑step, real‑world tasks instead of just reading benchmark tables, and gives you a concrete way to compare its reliability and speed to GPT‑5.2‑Codex or Claude Opus in a shared environment code arena page.

New commentary on NVIDIA’s NitroGen vision‑action foundation model fills in how it was trained: the team built an internet‑scale dataset with 40,000 hours of gameplay across 1,000+ commercial games, automatically extracting controller actions from videos that display on‑screen input overlays, and then wrapped everything in a Gymnasium‑style "universal simulator" benchmark so different agents can be compared on cross‑game generalization NitroGen intro NitroGen training. Following initial launch, the key new number is that NitroGen reports up to 52% relative success gains when transferring to unseen games versus training from scratch, suggesting a real shared game understanding rather than per‑title memorization dataset explanation.

Video loads on view

For AI engineers interested in embodied agents or robotics‑adjacent work, this matters because NitroGen is both open and built around messy real‑world data instead of clean simulator logs, and the universal simulator plus dataset give you a ready‑made benchmark to test your own policies or fine‑tuned variants. It also sets a useful pattern: open multi‑game data, a standard control API, and published transfer numbers, all packaged on an accessible site rather than hidden in internal infra NitroGen site.

Alibaba’s Tongyi‑MAI team’s first image model, Z‑Image Turbo, is now ranked #1 among open‑weights text‑to‑image models on the Artificial Analysis Image Arena, beating FLUX.2 [dev], HunyuanImage 3.0 and Qwen‑Image while running as a 6B‑param model that fits in ~16GB VRAM. The model is priced at about $5 per 1,000 images on Alibaba Cloud (vs ~$12 for FLUX.2 dev and ~$20+ for several rivals) and ships under Apache‑2.0 for unrestricted commercial use, making it an unusually cheap, permissively licensed choice for teams that want a deployable open image backbone rather than a research toy model overview. For AI engineers this means you can realistically host it on a single consumer‑class GPU, experiment with prompts and fine‑tuning locally, and still match or exceed the aesthetic quality of the current open‑weights field without getting locked into a restrictive license or high per‑image API costs.

AllenAI (AI2) quietly put two significant multimodal models into public demo on Hugging Face: Molmo 2, described as their latest state‑of‑the‑art multimodal model, and SAGE‑MM, a "Smart Any‑Horizon" long‑video reasoning agent Molmo2 mention SAGE-MM mention. Molmo 2’s space supports true multi‑image/multi‑modal input, so you can probe how it handles cross‑image reasoning, document understanding and grounded visual QA rather than single‑frame captions alone Molmo demo. SAGE‑MM, by contrast, targets long‑horizon video understanding: its demo takes extended clips and lets the model answer questions that require tracking entities and events over time, which is exactly the failure mode of many current "video" models that mostly hallucinate from a few frames SAGEMM demo. For engineers and researchers, these spaces are a low‑friction way to benchmark your own prompts, compare them to Gemini 3’s or OpenAI’s multimodal behavior, and decide whether these open projects are mature enough to plug into your own agent or analytics pipelines.

vLLM released 0.13.0 with new engine features like compile_ranges for selective kernel compilation, PrefixLM support for Flex/TritonAttention, and CUDA graphs for 3D Triton attention, plus runtime niceties such as xxHash-based prefix caching, chunked prefill for all pooling tasks, and improved ModelRunner (min‑p sampling, NaN logits detection) vLLM release thread. Following earlier diffusion cache work on TeaCache/Cache‑DiT diffusion cache, this release also adds large‑scale serving hooks—Mooncake KV connectors, /reset_prefix_cache, KV events, failure‑recovery configs, NIXL handshake checks, and an external launcher mode for orchestrators api changes. Hardware support now includes NVIDIA Blackwell Ultra SM103 (GB300) via CUDA 13, while DeepSeek‑specific optimizations (DeepEP CUDA graphs, DeepGEMM fused layouts, expert init, group_topk) report ~4–5% throughput gains and up to ~10% faster TTFT on DeepSeek‑V3.1 workloads deepseek tuning.

vLLM published an official serving recipe for Xiaomi’s MiMo‑V2‑Flash MoE model (309B total, 15B active), including a concrete vllm serve XiaomiMiMo/MiMo-V2-Flash command with --tensor-parallel-size 4, --tool-call-parser qwen3_xml, --reasoning-parser qwen3, and higher --gpu-memory-utilization (up to 0.95) to expand KV cache recipe tweet. The guide recommends tuning --max-model-len (commonly 65,536, max 128k) and --max-num-batched-tokens (e.g., 32,768 for throughput vs 16k/8k for latency/activation headroom), and enabling a "thinking" mode via enable_thinking: true in chat_template_kwargs, giving practitioners a ready-made profile for balancing context length, latency, and tool use.

It complements MiMo’s open‑weights release and technical report on GitHub, which detail the model architecture and benchmarks for reasoning, coding, and agents GitHub repo.

MorphLLM reports that its WarpGrep code search tool is now about 40% faster after using reinforcement learning to cut "wasted turns" and deploying fused MoE kernels on NVIDIA B200s to make forward passes more efficient warpgrep update. A follow‑up notes they had to roll back to investigate an issue rollback note, which is a reminder for runtime engineers that aggressive RL‑driven loop pruning and low‑level kernel fusion can materially improve latency, but need strong regression tests and monitoring before broad rollout.

Epoch AI’s latest census of large AI supercomputers estimates the US controls about 74.4% of total available AI compute, versus 14.1% for China and 4.8% for the EU as of March 2025. compute share thread That’s roughly 5.3× China’s capacity and 15.5× the EU’s, even though the data only covers an estimated 10–20% of global aggregate performance.

For engineers and AI leads, this clarifies why most frontier training runs and the most capable public models are still coming from US labs: they’re literally sitting on almost three-quarters of the known GPU-scale infrastructure. Commentators are already reframing the "AI race" as a race about infrastructure—compute, power, and space—rather than algorithms alone. infra framing For non‑US teams, it underlines the value of model‑distillation, clever inference optimization, and hardware diversity to stay competitive without US‑scale clusters.

A long Grok‑assisted roundup sketches where the major AI accelerators and hyperscaler data centers actually stand heading into 2026, and when they’ll matter in practice. chip and dc thread NVIDIA’s GB200/GB300 (Blackwell Ultra) is already shipping in volume since mid‑2025, with GB300 racks landing in Azure (4,600+ GPUs earmarked for OpenAI), Google Cloud, and others; most of that is new capacity that will start showing up in user‑visible products through 2026.

On the custom‑silicon side, Google’s TPU v7 Ironwood has been generally available since November 2025 for high‑volume FP8 inference, with multi‑datacenter deployments like New Albany, OH and plans to expand sales into customer data centers in 2026. AMD’s MI300X/MI355X (CDNA4) are rolling out in full‑rack systems at IBM Cloud and providers like Crusoe and Vultr, with a MI450 family planned for 2026, while Intel’s Gaudi3 entered volume production in H2 2025 with OEM boxes from Dell, Supermicro, and HPE.

The same overview tracks cloud‑vendor chips: Amazon’s Trainium3 is live in Trn3 UltraServers with Trainium4 in the works (aiming at Nvidia‑compatibility), Microsoft’s Maia 100 is deployed with Maia 200 delayed into 2026, and Qualcomm’s AI200 targets data‑center inference in 2026 alongside Huawei’s Ascend 950/960 roadmap. chip and dc thread On top of that, labs and partners are planning multi‑GW AI campuses—OpenAI/Microsoft’s Stargate, xAI’s Colossus, Anthropic’s AWS Rainier buildout, Google’s TPU‑heavy sites, and Tesla’s 200 MW training center in Texas—adding roughly 7 GW of new capacity by the end of 2025 with more in 2026.

For AI engineers and infra leads, the point is: most of the hardware that will power 2026–2028 frontier models is only now reaching racks. If you’re planning model upgrades or cost curves, this argues for a noticeable shift in available flops, memory bandwidth, and price/performance over the next 12–24 months, but not an overnight transformation—all of this silicon still has to be wired into power‑hungry, physically limited campuses.

According to The Information, OpenAI’s compute margin on paying users climbed to about 70% in October 2025, up from ~52% at the end of 2024 and roughly 35% in January 2024. margin summary The gains reportedly come from cheaper rented compute, inference‑side efficiency work, and a higher‑priced subscription tier that soaks up more of the fixed infra costs.

For AI product teams, this is a reminder that serving economics can move fast when you aggressively optimize kernels, caching, batching, and model mix—even without changing your list prices. It also hints that OpenAI has more room to cut API/ChatGPT prices or cross‑subsidize new features if competition forces it, though the same report suggests that on total computing costs (including training) Anthropic may currently be more efficient overall. margin summary If you’re building on their APIs, this kind of margin trajectory is exactly what makes volume discounts, cheaper “Flash” tiers, or more generous rate limits plausible over the next year.

Bloomberg reporting says OpenAI has gone into internal “Code Red” mode multiple times, most recently right after Google’s Gemini 3 launch, redirecting teams from side bets back onto the core ChatGPT stack. bloomberg summary In the latest sprint they focused on lower latency, higher uptime, and tighter eval loops so quality drops are caught before users see them, then shipped GPT‑5.2, GPT‑5.2‑Codex, and a rebuilt ChatGPT Images experience with up to 4× faster generations.

Code Red here is less about existential panic and more an operating model: leadership temporarily pauses agents, ads, and other projects to tune the serving layer and product fundamentals when a competitor steps up. For teams outside OpenAI, this is a useful signal that even at frontier scale, competitive pressure is showing up as ops and infra discipline (latency, reliability, eval pipelines), not only as bigger models. The piece also notes that OpenAI’s next big internal focus is the training engine itself—algorithms plus infrastructure for large runs—with discussions of up to $1.4T in infrastructure over eight years, which will shape how often they can afford these big Code Red pushes. bloomberg summary

A new breakdown of OpenAI’s Stargate buildout lists at least nine major US data center sites either operating, approved, or in fast build across Texas, Ohio, Michigan, New Mexico, and Wisconsin, many in the 1–1.5 GW range. stargate site list Abilene, TX is already partially live on Oracle Cloud with NVIDIA GB200 racks and ~600 MW planned; Lordstown, OH and Milam County, TX are each slated to scale to about 1.5 GW in roughly 18 months via SoftBank partners.

Following Michigan stall, where a $10B, 1 GW Oracle‑backed Michigan site briefly looked shaky, Michigan’s Saline Township has now approved a 1.4 GW facility with construction expected to start in early 2026, alongside land purchases near Grand Rapids. stargate site list The same summary notes additional Oracle‑run sites in Shackelford County (TX), Doña Ana County (NM), and Wisconsin, plus international expansions in Norway, Patagonia (Argentina), and UAE/South Korea partnerships.

Taken together, this looks like a deliberate plan to secure 10+ GW of AI‑grade power and space by the late 2020s, spread across multiple utilities and financing structures. For anyone betting on long‑term OpenAI APIs or model hosting, the message is clear: they’re not just renting GPUs for the next quarter, they’re building a continental‑scale power and cooling footprint that will set the ceiling on how often they can train and refresh frontier models.

Building on the Epoch AI compute‑share chart that put the US at 74.4% of measured AI capacity and China at 14.1%, some analysts argue the “AI race” is now primarily about who can secure power, land, and chips rather than who has the fanciest architecture. compute share thread One thread sums it up bluntly as an "AI race is now infrastructure" problem: compute, power, and space decide who can scale to frontier models. infra comment This framing lines up with the parallel data points on China’s rapidly growing power generation and US‑based projects like Amazon’s 2.2 GW Indiana campus and OpenAI’s multi‑GW Stargate network, even though those specific facilities were detailed on earlier days. For AI leaders, it’s a reminder that policy, energy strategy, and real‑estate decisions are now as central to long‑term model strategy as optimizer tricks or new attention variants. If your organization doesn’t control its own large‑scale power and compute, your leverage will increasingly come from choosing the right partners, negotiating long‑term capacity, and optimizing workloads to ride on top of whatever those infrastructure players can actually deliver.

OpenBMB and Tsinghua researchers released an interpretability study showing that a remarkably sparse subset of neurons (under 0.1 percent of all units) carries strong predictive signal for when an LLM will hallucinate across multiple domains, dubbing them H-Neurons paper overview. By selectively suppressing or activating these neurons, they can flip the model between refusing bad prompts and over-complying with invalid premises, misleading context, skeptical users, or even harmful instructions, with qualitative examples like cat feathers, Marie Curie’s field, second-guessing correct answers, and weapon-creation requests paper overview.

The work further traces H-Neurons to pretraining rather than later alignment or RLHF stages, suggesting that the next-token prediction objective itself bakes in an over-compliance tendency that later safety fine-tuning only partially counteracts paper overview. For engineers and evaluators, this offers a concrete mechanistic target for reducing hallucinations (suppressing or regularizing a tiny neuron subset) and a new way to frame benign failure modes like helpfulness-over-truth as an internal circuit, not just a surface behavior.

Chinese researchers have unveiled a prototype analog AI accelerator that reportedly solves certain complex mathematical workloads up to 1,000× faster than leading digital processors like Nvidia GPUs, while consuming far less power by encoding computation directly into the physics of the chip instead of sequences of digital operations analog chip clip.

Video loads on view

A longer explainer notes that the Peking University team’s design targets niche classes of scientific and AI problems (e.g., structured linear algebra) rather than acting as a drop‑in GPU replacement, trading generality for extreme throughput and energy efficiency on those patterns YouTube report. For AI engineers and infra leads this reinforces a likely future where some high‑volume kernels (matrix solves, certain transforms) migrate to specialized analog hardware, forcing model and workload designers to think in terms of hardware‑aware decompositions rather than assuming everything runs on standard CUDA stacks.

A separate line of Chinese work is pushing photonic AI accelerators that replace electrons with photons, using optical interference to perform matrix operations for specific generative workloads at over 100× the speed and far lower energy than Nvidia GPUs on those targeted tasks photonic chips thread.

Unlike general‑purpose GPUs, these chips (examples include ACCEL and LightGen) are highly specialized: they excel at fixed dense linear transforms common in image and video generation but are hard to reprogram, require analog calibration, and only help if models and runtimes are co‑designed to offload the right layers to optics analysis article. For AI leaders the takeaway is that, if photonics matures, the bottleneck may shift from raw GPU supply to how well your model architecture, compiler, and serving stack can match workloads to a heterogeneous mix of digital, analog, and optical accelerators.

PJ Accetturo took six Hollywood cinematographers with no AI background and walked them through a repeatable framework to make a $1M‑prize AI short, then published the exact prompts and process. film workflow thread The core stack combines shot ideation, style locking, and iteration into a loop that feels like traditional pre‑pro: block out beats, generate boards and looks, then refine shots until they feel cinematic instead of "AI-ish". film workflow thread Follow‑up clips show how quickly they picked up the tools, with viewers calling out how strong the emotional micro‑expressions and camera language have become in the latest models. (followup reply, emotional gesture comment)

Video loads on view

For AI‑curious filmmakers, this is more than another showcase: it’s a teaching blueprint you can borrow. You can plug your own tools into the same loop—concept → boards → shot recipes → polish—and onboard non‑technical collaborators by handing them a proven set of prompts instead of making them "learn AI" from scratch.

Creators are sharing a step‑by‑step workflow for Kling 2.6’s Motion Control that transfers motion from a reference clip onto an edited still, turning random footage into precise, directed shots. motion control guide The pattern: grab a key frame, run it through an image model to swap in your subject, then feed both still and reference video into Motion Control with a short prompt or JSON block to lock style and timing. motion control guide A second tester shows that even chaotic body movements are tracked convincingly, underscoring how much of the "magic" now lives in pre‑ and post‑Kling setup rather than raw text prompting. prompt test

Video loads on view

For teams already storyboarding in stills, this effectively turns Kling into a motion layer you can bolt onto existing concept art or photos. It also hints at a new kind of reusable recipe: save your frame‑editing prompt plus a Motion Control JSON once, then swap in new reference clips to rapidly prototype alternate performances or memes at near-zero extra prompting cost. kling referral

A detailed Freepik Spaces walkthrough turns one selfie into a full mini episode of Pawn Stars using a node graph that chains Nano Banana Pro images into Veo 3.1 animations. pawn stars tutorial The flow: generate a 3×3 grid of character stills, split each cell into its own node via an "extraction" prompt, optionally touch up problem shots, then send each through Veo with motion and dialogue prompts or JSON start/end frame controls to get consistent, talking shots. pawn stars tutorial The thread also shares screen captures of the prompts inside node ALT text, making it effectively a copy‑pasteable template for other mock TV formats. spaces link

Video loads on view

Because Spaces lets you wire image and video nodes visually, this becomes a reusable scene factory: keep the grid‑extraction and Veo nodes, swap in a new reference face or show concept, and you’ve got a working pipeline for explainer skits, fake reality TV, or internal training content without re‑engineering the stack each time. If you’re evaluating which "AI canvas" to learn, this is a good example of the level of structure you should expect. Freepik plans

Creator @fofrAI is popularising deeply structured JSON prompts that read more like a shot spec than a sentence, with nested fields for subject, pose, clothing, background, vibe, constraints, and an explicit negative_prompt list. portrait example One example describes a pop‑art fashion portrait down to eyeliner style, glove material, portal layout, lighting, and what must be preserved (neon outlines, circular cutouts, blue wall) across generations. portrait example A second thread shows how to mutate a JSON prompt by prepending a short instruction like "generate a new image with significantly different nouns" while keeping the original aesthetic and mood, effectively turning it into a series template. mtg card mock

He applies the same pattern to more mundane scenes, like a POV phone gallery in a cozy living room with specific UI text, background TV timecodes, and plushie placement, which is crucial if you’re trying to keep continuity across shots or brand assets. phone gallery spec For teams, this suggests a new abstraction layer: instead of one‑off prompts buried in DMs, you can maintain versioned JSON "style docs" in git, diff them, and let both humans and agents reliably generate on‑brand imagery from the same spec. profile outlines

Developer @ryolu_ shipped a full karaoke web app—word‑synced lyrics, YouTube search or URL paste, Japanese furigana, Korean romanization, and on‑the‑fly translations—entirely from a phone by leaning on Cursor cloud agents. karaoke demo The code, backend, and UI were orchestrated through remote agents while away from a laptop, turning the phone into what the author calls an "ideas expression device" rather than a consumption screen. cursor-built note Others are already asking to test it live, suggesting that this kind of agent‑assisted development is crossing from gimmick into practical weekend shipping. swyx comment

Video loads on view

For AI engineers, the interesting part isn’t karaoke itself—it’s that a reasonably polished, multi‑language, media‑heavy app can now be pushed to production with almost all of the heavy lifting handled by agents driving git, package managers, and hosting. If you’re still only using coding models as glorified autocomplete, this is a strong nudge to experiment with cloud agents plus mobile workflows.karaoke site

On top of the tutorial thread, @techhalla turned the Pawn Stars Space into something others can clone directly, framing Freepik Spaces as "the best way to work with AI" when combining NB Pro for stills and Veo 3.1 for motion. spaces link The Space wires together grid generation, cell extraction, minor touch‑ups, and multiple Veo video nodes so you can drop in your own face or concept and immediately get similar TV‑style scenes without rebuilding the graph. pawn stars tutorial

Video loads on view

For AI teams, this highlights an emerging pattern: Spaces or canvases as distribution for multi‑model workflows. Instead of sharing isolated prompts, creators are starting to share whole node graphs that you can fork, parameterise, and slot into your own production stack, which is much closer to how engineers think about reusing code than how artists have historically traded presets. Freepik plans

Ethan Mollick points out that Midjourney’s product direction is drifting further from "type a perfect prompt, get exactly that" and more toward tools that help you explore, curate, and vary large grids of options. midjourney stance In his own wizard‑themed project he rapidly iterated through many moods and styles, selecting interesting branches rather than trying to coerce the model into matching a single textual spec. wizard gallery This is a useful mental model shift for art directors and designers: Midjourney is optimised for interactive search through visual space, not contract‑style instruction following. That means the right workflow looks more like: loose prompt → generate many → pick promising frames → riff with variations and region edits, instead of spending all your time wordsmithing one massive prompt hoping for a perfect, one‑shot render.

One parent reports that his 2‑ and 5‑year‑old daughters are "huge fans" of fully AI‑generated music, especially a track titled "Poop & Boobies" that he admits is annoyingly catchy. kids song clip The same service can spin up more niche genres like post‑disco on demand, suggesting that generative music is already good enough to hold kids’ attention without any brand behind it. post disco track

Video loads on view

This is anecdotal, but it’s a useful signal if you’re thinking about AI audio products: kids are a natural fit for hyper‑personalised, infinite catalogs, and they don’t care that a model wrote the song. The real work for builders will be on controls, safety, and filtering, not raw audio quality.

DeepWiki’s MCP server is getting more attention as a practical way to give agents recoverable "memory" over large codebases: you can ask natural‑language questions about a GitHub repo and get structured answers with file/line references, and now a Codemap view that traces concepts like dialog systems or popover UIs across files. repo Q&A demo

For engineers wiring MCP into Claude Code, Codex, or other agentic CLIs, DeepWiki effectively is your long‑term repo memory: it indexes code once, then serves fast, navigable retrieval for design questions ("where is the dialog system implemented?"), change impact analysis, or onboarding, without you having to reinvent RAG or embeddings for every new project. codemap screenshot

Oracle and the LangChain community are pushing a very opinionated take on agent memory with the new AI Developer Hub notebook that implements six distinct persistent memory patterns—semantic, episodic, conversation history, working, procedural, and cross‑session—on top of Oracle AI Database for scalable context management. Oracle hub overview

For AI engineers, this is a concrete, production‑minded reference for how to separate short‑term execution state from long‑term cross‑session memory, wire that into LangChain agents, and layer RAG plus evaluation/observability on top. GitHub notebook It’s especially relevant if you’re trying to convince infra or data teams that agent memory can live in a real database with schemas and SLAs instead of ad‑hoc JSON blobs scattered across Redis or S3.

LangChain community contributors released LangAlpha, a LangGraph-based multi‑agent system that takes a high‑level equity query and orchestrates agents over market data, news, and financial statements to spit out what they call "institutional‑grade" research reports in minutes. LangAlpha summary

For AI leaders and quant/product teams, this is a concrete pattern for verticalized research agents: instead of a single chat model, LangAlpha routes subtasks like data gathering, factuality checking, and drafting across specialized agents, then composes the result into a structured report artifact you could plausibly drop into an analyst workflow or client deck.

Researchers from UIUC, Stanford, Harvard and others lay out a clean taxonomy for how agentic AI systems should adapt: A1/A2 adapt the agent itself (using tool outcomes or its own outputs), while T1/T2 adapt the tools, including memory, either independently or under supervision from a frozen agent. adaptation thread The paper argues that most real systems should combine them—for example, frozen-but-improving retrievers (T1), retrievers and memories shaped by agent preferences (T2), and a reasoning model fine‑tuned on execution results (A1). tool taxonomy They explicitly classify adaptive memory as a T2‑style tool, since the agent’s outputs are written back into memory and then change future behavior, and provide both an ArXiv paper and an "Awesome" repo cataloging adaptation strategies for builders. If you’re designing long‑lived agents, this gives a shared language and checklists for deciding when to spend compute fine‑tuning the model versus upgrading retrievers, memories, and other tools instead. (ArXiv paper, GitHub repo)

A UCSC/CMU team proposes a multi-agent framework that combines dynamic routing, bidirectional feedback, and parallel agent evaluation to analyze long, messy documents like SEC 10‑Ks, reaching 92% factual coverage and 94% compliance accuracy versus ~71%/74% for static baselines. paper summary The system routes subtasks to specialized agents at run time, lets downstream "QA" agents request upstream revisions, and runs multiple agents in parallel on high‑ambiguity spans, cutting revision rates by 74% and redundancy by 73% in their 10‑K benchmark.

For AI engineers building serious doc-understanding agents, this is a concrete recipe: shared memory + feedback loops + selective parallelism beat rigid role pipelines in both accuracy and human-rated coherence.

New charts show US VC fundraising for 2025 on track to be the weakest in a decade, with only about $45B in new fund commitments by Q3—roughly a 75% drop from the 2022 peak. vc fundraising thread Dealmaking, however, remains near 2021 highs at ~$330B deployed over the last four quarters, so managers are investing off older funds and reserving scarce "dry powder" mostly to defend existing portfolios. vc fundraising thread

For AI founders this means fresh lead investors are harder to find, new fund formation is way down, and capital is concentrating in bigger, later‑stage rounds. Early‑stage AI teams should expect tougher terms, more structured deals, and investors who are pickier about clear differentiation and paths to revenue, even while top AI infra and application companies can still raise sizable growth rounds.

Aaron Levie lays out a very specific picture of where value accrues in enterprise AI: LLM labs will ship "generally capable college student" models, while application companies own the layer that bundles tools, private data, orchestration, and UX into domain specialists. levie app thesis He highlights context engineering, multi-call DAG orchestration, human-in-the-loop GUIs, autonomy sliders, and deep system integration/change‑management as the ingredients needed to turn raw models into deployed "professionals" inside enterprises. levie app thesis

For AI leaders this is a playbook: focus less on chasing each new frontier model, and more on vertical agents that plug into CRMs, ERPs, ticketing, and custom data, with opinionated interfaces for specific roles (e.g., underwriters, sales ops, finance). It also suggests that generic "chat with your data" wrappers—thin prompts over APIs—will get squeezed, while teams that invest in integration, routing, memory, and process change have room to build durable products and GTM.

A new visualization of "largest potential IPOs" pegs SpaceX at ~$1.5T, OpenAI at ~$830B, and Anthropic at ~$230B, with the top 10 prospective listings totaling about $3.6T in aggregate value. ipo chart post That follows earlier reporting that OpenAI has already explored private funding at roughly an $830B valuation, 830B valuation underscoring how AI labs have moved into the same valuation tier as the biggest non‑public tech companies.

For AI leaders and investors, this potential IPO pipeline matters in two ways. First, public listings at these scales would recycle huge amounts of capital back to LPs, helping unclog the current VC fundraising slump and setting reference prices for late‑stage AI valuations. Second, the relative sizes—OpenAI almost 4× Anthropic and ~0.5× SpaceX—shape expectations about eventual consolidation, ecosystem power dynamics, and which partners or vendors feel "too strategic" to ignore.

New November 2025 numbers from Similarweb show ChatGPT.com still capturing 66.8% of US web visits among AI chatbot and tool sites, with roughly 1.0B monthly visits. traffic stats Google’s Gemini site ranks second at 11.2% and 168M visits, followed by Claude (3.6%), Grok (3.5%), Character.ai (3.0%), OpenAI.com (2.9%), and Perplexity (2.2%), with DeepSeek, Microsoft Copilot, and newcomer Polybuzz rounding out the top ten.

For product leaders, this is a reminder that ChatGPT remains the primary discovery channel for mainstream users, even as Gemini grows quickly and Grok/Perplexity carve out niches. Distribution strategy in 2026 likely means meeting users where they already are (web, mobile, and embeddings into Google and Microsoft surfaces), rather than expecting them to switch wholesale to smaller standalone tools—unless those tools can ride enterprise rollouts or deeply vertical workflows.

Several consumer startups are reframing AI "assistants" as personalized talent agents that continuously match people with opportunities across jobs, fundraising, and even dating. Examples include Jack & Jill for candidates and companies, Boardy AI for investors and founders, and Known AI for relationships, all of which run multi‑party interviews and matchmaking over time. talent agent thread The thesis is that once these agents aggregate proprietary supply (candidate pools, cap tables, dating profiles), they start to look less like tools and more like two‑sided marketplaces with AI in the middle. marketplace followup

Video loads on view

For builders, this highlights a playbook: use agents to bootstrap liquidity and data on both sides of a market before layering in pricing, reputation, and routing logic. It also hints at defensibility: whoever’s AI ends up "owning" the ongoing relationship with candidates or founders can become the default discovery surface, even if the underlying models are commoditized.

Public.com quietly rolled out Generated Assets, an AI feature that turns natural‑language prompts like "autonomous vehicle companies with >25% YoY revenue growth" into investable stock baskets you can backtest and buy. generated assets demo Behind the scenes, multiple "analysis agents" research and screen thousands of securities, then assemble and periodically rebalance a custom index, with performance compared against benchmarks such as the S&P 500. generated asset page

Video loads on view

The product comes with strong disclaimers that outputs are not personalized advice and that users remain responsible for portfolio fit, but it’s a concrete example of LLMs and tool‑use escaping the chatbot and surfacing directly in retail investing workflows. For AI teams building in finance, it’s a signal that regulators will likely scrutinize how agentic research is disclosed, how backtests are presented, and how conflicts between "what tests well" and suitability for individuals are handled.

A short field note from an engineer who watched Runloop’s recent demo boils the enterprise pitch down to three words: predictability, auditability, templatization. runloop summary Teams can define Blueprints for sandboxes so every agent run happens in a known, repeatable environment that IT can approve, while the Deep Agent harness stays fully open—every prompt, tool call, and execution loop is inspectable, and traces can be logged to LangSmith and S3 to satisfy audit and retention requirements. runloop summary This is exactly the kind of framing enterprise buyers want to hear as they evaluate agentic systems: not "look what the model can do", but "show us the blast radius, the logs, and the rollback story." If you are building internal agent platforms, adopting similar patterns—templateable environments plus first‑class observability—will likely matter more for rollout success than squeezing a few more benchmark points out of your base model.