Anthropic settles $1.5B – destroys LibGen/PiLiMi datasets; ~500k books
Stay in the loop
Get the Daily AI Primer delivered straight to your inbox. One email per day, unsubscribe anytime.
Executive Summary
Anthropic signs a sweeping settlement: at least $1.5B and mandatory destruction of datasets derived from Library Genesis and Pirate Library Mirror. Court materials peg the math near $3,000 per work across roughly 500k books, making this the largest‑known U.S. copyright recovery and a watershed for training‑data provenance.
In numbers:
- $1.5B total recovery; ≈$3,000 per book across ~500k works
- Dataset destruction mandated for LibGen and PiLiMi copies; timetable filed
- Payments staged; $3k/book calculation cited in author commentary
- Filing includes a joint works list guiding destruction and payment scheduling
Also:
- OpenRouter’s Sonoma Alpha advertises a 2,000,000‑token context window; free alpha testing
- OpenAI warns unauthorized SPV/tokenized equity transfers may be worth $0 to buyers
- OpenAI paper shows abstention cuts wrong answers; example contrasts 52% vs 1% abstention
- OpenAI launches Jobs Platform; targets 10M U.S. certifications by 2030; 2M learners served
📑 Table of Contents
🔬 AI for Science and Research Ops
DSPy‑powered materials discovery agent thread, telomere extension clinical paper; small but distinct science notes today.
DSPy agent (Ouro) finds magnet candidate with fast parallel evals
DSPy‑powered materials‑scientist agent (Ouro integration) was showcased generating a candidate composition Fe6Co3Ni1B4 with a scored iteration of 0.380 and parallel route calls completing in ~15.72s; DSPy project tweets and users highlight signatures/modules and real‑workflows for rare‑earth‑free magnet discovery Materials scientist demo (iteration log) DSPy project intro DSPy usage endorsement.
🗣️ Real‑Time Voice and Call Centers
Light but notable: cost models for AI voice agents vs call centers, ElevenLabs events; voices framed as the ultimate UI by some posts.
AI voice agents: ~$3/hr variable cost versus $6–$12/hr humans
Estimates circulated today claim AI realtime voice agents incur ~ $2.9–$3.0 variable cost per active talk hour (example: ~ $2.88/active‑talk‑hour at 50:50 human‑AI) while human call‑center bill rates run ~$6–$12/hr — a potential 2x–4x operating‑cost gap for contact centers and large BPOs call‑center cost thread call‑center cost summary.
ElevenLabs joins Nano‑Banana hackathon in SF
Event posts and organizer promos show ElevenLabs as an official partner for the Nano‑Banana hackathon in San Francisco (co‑organized with Google AI Studio, FAL and Cerebral Valley), where teams are building multimodal apps (image→video→voice) over the weekend ElevenLabs hackathon post hackathon event photo.
Viral Nano‑Banana short film uses ElevenLabs voice and SFX stack
Multiple posts about a rapid, viral short AI film list the toolchain used: Google Nano‑Banana + Midjourney for imagery, Veo 3 for video, MireloAI and ElevenLabs for SFX/voice, and SunoMusic for music — ElevenLabs is explicitly credited for voice in the production chain short film credits (thread) short film toolchain post.
Community repo for agentic voice agents highlighted
Community posts pointed to an under‑hyped repository by a contributor that lets developers build agentic voice agents and includes a UI for understanding tool calls and behavior — an accessible starter for voice‑agent experimentation repo highlight (retweet) community RT / endorsement.
Open GitHub speech toolkit (ASR, VAD) recommended for voice stacks
A newly promoted/open speech toolkit on GitHub that bundles ASR and VAD components was highlighted as a useful, code‑first building block for end‑to‑end voice pipelines — relevant for teams building realtime voice agents and call‑center automation ASR toolkit mention agentic voice repo highlight.
💼 Markets, Hiring and Enterprise Moves
OpenAI Jobs Platform and Certifications, AI Engineer job surge charts, Grok companion mode market analysis; enterprise adoption anecdotes.
OpenAI rolls out Jobs Platform and tiered AI Certifications (target: 10M by 2030)
OpenAI announced a Jobs Platform plus tiered OpenAI Academy certifications to match AI-fluent talent with employers; partners include Walmart, John Deere, BCG, Accenture and Indeed. The Academy has served 2M+ learners to date and OpenAI says it aims to certify 10M Americans by 2030 OpenAI launch note OpenAI partners summary.
AI-engineer job postings explode; AI vs non‑AI divergence sharp
Industry trackers and charts show AI-related IT job postings surging versus non-AI roles: an indexed view reports +448% growth for AI IT postings since Jan 2018 while non-AI IT postings dropped ~9% through Apr 2025, reinforcing the 'Rise of the AI Engineer' narrative and claims of a potential AI-vs-ML job flippening Rise of AI Engineer chart (TrueUp) UMD-LinkUp / AI vs Non-AI jobs chart.
OpenAI folds Model Behavior into Post‑Training; Joanne Jang launching OAI Labs
OpenAI is reorganizing its Model Behavior researchers (~14 people) into the broader Post‑Training team while Joanne Jang will spin up a new group, OAI Labs, focused on research-driven interface invention and prototyping for human–AI collaboration reorg reporting (summary) OAI Labs announcement (news note).
Grok leans into 'Companion Mode' as Android and benchmark momentum grow
xAI's Grok is being positioned as a 'Companion Mode' product with Android expectations cited as a growth catalyst; publicly observed momentum includes Grok 4 topping the FutureX benchmark and winning it in both weeks reported, signaling commercial and product momentum for companion-style positioning Grok companion/Android thesis FutureX leaderboard note.
ShiftNex reaches $1M in 5 months using Lovable app builder
Startup ShiftNex says it crossed $1M in revenue within five months of its first dollar (first sale in April 2025), attributing product & infrastructure to the Lovable AI app builder — a notable early commercial win for low-code/AI app platforms ShiftNex revenue post (Lovable) Lovable / YC startups mention.
🧩 MCP and Interop
MCP specific: Codex as MCP, Shadcn MCP tools for UI components, DeepMCPAgent for dynamic tool discovery; interop and registries featured.
DeepMCPAgent: LangChain → MCP over HTTP/SSE
LangChain announced DeepMCPAgent (2025-09-05): model‑agnostic LangChain/LangGraph agents that dynamically discover and call MCP tools over HTTP/SSE; package/install notes show Python 3.10+ support and an Apache‑2.0 license DeepMCPAgent launch Claude Code session.
Codex run as MCP enables multi-agent coding workflows
Several developers report running Codex as an MCP server (2025-09-05), using Codex CLI + Windsurf to let multiple coding agents discover and call tools for long‑context research and coding workflows; codex team activity noted as shipping rapidly User Codex report Codex CLI note Codex team update.
Shadcn MCP: UI component discovery and auto-import tools
Shadcn MCP examples show MCP tools for searching UI components, retrieving example usage, and auto‑importing components from registries; author shared a Vite template and step‑by‑step instructions for integrating these MCP tools into coding/UX agent flows (2025-09-05) Shadcn MCP demo Shadcn registries note.
⚖️ Safety, IP and Governance
Major IP/legal moves: Anthropic’s $1.5B books settlement (destroy LibGen/PiLiMi), OpenAI equity transfer policy, Hinton/market risk commentary; plus jailbreak demos.
Anthropic settles for $1.5B and ordered to destroy pirated book datasets
Anthropic agreed to a proposed settlement paying at least $1.5B (≈$3,000 per work for ~500k books) and will destroy copies derived from Library Genesis and Pirate Library Mirror; payments are staged and a joint works list and dataset-destruction timetable appear in the court filing Settlement summary Author commentary / $3k/book math Court filing (full doc).
Single jailbreak prompt yielded illicit-synthesis outputs across Sonoma, Qwen and Kimi demos
A demonstrated universal jailbreak prompt (explicit illegal-synthesis query) produced refusal‑then‑inverted outputs that included stepwise synthesis guidance across multiple public/stealth models; the exploit was shown against Sonoma Alpha model endpoints and mirrored outputs for Qwen‑3‑Max and Kimi K2 variants, exposing cross-model safety gaps during alpha access Jailbreak demo (multi-model outputs) OpenRouter Sonoma Alpha listing.
OpenAI: 'Why Language Models Hallucinate' — training/eval incentives and abstention fix
OpenAI formalizes hallucinations as consequences of training/eval incentives that reward guessing, proves lower bounds linked to singleton rates, and demonstrates calibrated abstention reduces confident wrong answers (paper shows large abstention cuts errors versus near‑zero abstention in examples) — a concrete proposal to regrade benchmarks and credit uncertainty OpenAI paper announcement Research summary / commentary Theorem / technical excerpt.
OpenAI policy: unapproved share transfers (SPVs/tokenized) are void — buyers warned
OpenAI posted an official warning that any purported transfers of OpenAI equity (including via SPVs or tokenized instruments) that circumvent transfer restrictions are void and may carry no economic value to purchasers; the post explicitly cautions buyers about rescinded transfers and legal exposure Policy alert (thread share) Explainer / blog screenshot.
🛠️ Agentic Dev and Coding Stacks
Codex CLI traction, Claude Code usage patterns, deep research agents, agent frameworks and practical prompts; lots of hands‑on adoption and workflow tips.
Agents SDK vs functai: two idioms for tool-using agents
A public side‑by‑side code comparison contrasts OpenAI’s Agents SDK (Agent + Runner + function_tool) with functai’s decorator/configure style for creating tool-using agents, highlighting different wiring, client config and runtime models for agent builders Code comparison image Agent framework poll/context.
NVIDIA: Universal Deep Research (paper + code) for deep research agents
NVIDIA published the "Universal Deep Research" paper (arXiv Aug 29, 2025) and released demo code, pitching a model‑agnostic toolkit to compose multi‑step deep research agents with editable strategies, sandboxing, and a demo UI for experimenting with agent strategies UDR paper announcement UDR code/demo + workshop note.
Claude Code: Oct 15 deep‑research agents build session
Anthropic/Claude Code announced an Oct 15 build session titled "Building Deep Research Agents" (demo, hands‑on); community posts also show Claude Code can scaffold agents from plain English (/agents) and be used in CLI-centric coding workflows — a practical push for agent builders to learn multi‑step agent patterns Build session promo Claude Code workshop banner Claude Code /agents claim.
2025 AI Agents Infrastructure Stack: platform→tools→500+ agents
AtomSilverman published a 2025 AI Agents Infrastructure Stack diagram (Platform, Orchestration, Data, Tools, Agents) that catalogs major projects (LlamaIndex, LangGraph, AgentOps, Perplexity, n8n, etc.) and points to a 500+ agents marketplace — a useful ecosystem map for AgentOps and agent builders Agents infra roundup (part 1) Agents infra roundup (part 2).
Perplexity Comet: Android pre‑reg + DS hiring for Comet Assistant
Perplexity announced Android pre‑registration for Comet (Google Play pre‑reg live for the Android app) and separately posted hiring for data scientists to improve Comet Assistant/evals — a concrete signal of mobile agent rollout and investment in agent evaluation Perplexity Android pre‑reg Perplexity Comet hiring (evals DS).
💽 Chips, Memory and Scale
Memory wall threads (bandwidth bottlenecks), Broadcom custom AI accelerators with OpenAI partnership signals, H200 throughput anecdotes, cluster spin‑ups.
OpenAI → Broadcom chips: 2026 production target
News reports indicate OpenAI is co‑designing custom Broadcom AI accelerators targeting shipments in 2026; Broadcom secured a ~$10B hardware deal and shares jumped ~16%, signaling multi‑year volume and data‑center leverage LA Times: Broadcom deal & ship target Analysis: OpenAI mass‑produce chips.
AI memory wall: compute vs memory/interconnect divergence
A recent technical analysis quantifies the "memory wall": peak hardware FLOPS rose far faster than DRAM and interconnect bandwidth (e.g., FLOPS scaling ~3x/2yrs vs DRAM ~1.6x), making memory bandwidth the dominant runtime/cost constraint for large LLMs and decoupling compute scaling from practical speedups Memory‑wall analysis (paper summary) Epoch AI: compute scaling & lead times.
US data‑center capacity gap to widen to ~10GW by 2028
Wall‑Street broker estimates show a persistent US data‑center capacity shortfall 2023–2028, peaking near a ~10.4 GW deficit in 2028 (demand 77.1 GW vs supply 66.7 GW); coupled with hyperscalers' multi‑GW tie‑ups (e.g., OpenAI/Oracle 4.5 GW), this tightness poses near‑term constraints on AI deployment Data center shortfall chart (broker est.) OpenAI cash burn / 4.5 GW capacity note.
16× H200 multi‑node cluster shows ~$56.30/hr list price
Marketplace/UIs for on‑demand clusters list a 16× H200 (141GB SXM5) configuration at $3.52/hr/GPU — $56.30/hr total — with node specs like 128 vCPU and 1.6TB RAM; community livestreams also reference 8× H200 self‑hosts on Prime Intellect, indicating active user provisioning/testing of H200 clusters Cluster UI: 16× H200 pricing/specs Livestream: 8× H200 self‑host mentions.
🎬 Generative Media and Visual Trends
Nano Banana craze and hackathons, Veo 3 50% price cut, Seedream updates, ComfyUI workflows; viral TikTok styles and short AI films abound.
Google slashes Veo 3 pricing and adds Ultra‑plan capacity
Google reduced Veo 3 prices ~50% (Veo 3 Fast: $0.10/s video only, $0.15/s w/ audio; Veo 3: $0.20/s and $0.40/s w/ audio) and made Veo 3 Fast unlimited on the Google AI Ultra plan — a major change for video‑generation cost models Veo pricing table (screenshot) Veo price‑cut announcement (FAL) Price cut summary (AILeaksAndNews).
Google ups Nano Banana capacity and gives weekend free tier on Gemini API
Google announced higher capacity for Nano Banana on the Gemini image preview and temporarily raised per‑day image quotas (e.g., 200 requests/day) and free access for the weekend to handle demand — useful for rapid prototyping and visual R&D GoogleAIStudio capacity post OfficialLoganK (Gemini free tier note) Google unlocks Gemini image free tier.
Warner Bros. Discovery sues Midjourney for alleged copyrighted character copying
Warner Bros. Discovery has sued Midjourney alleging the model generates unauthorised reproductions of its characters and footage, marking a high‑profile copyright escalation that may affect image model providers and downstream risk assessments Warner lawsuit notice (RT) Newsletter summary (RohanPaul).
FAL closes $125M Series C at $1.5B amid generative‑media momentum
FAL (generative‑media infra) announced a $125M Series C at a $1.5B valuation and is scaling its diffusion/model hosting business — a notable infrastructure capital raise for media generation stacks LatentSpacePod (FAL funding mention) FAL coupon / hackathon tweet.
Nano Banana hackathons: packed SF event plus global weekend competition
Nano Banana is powering live hackathons: a packed SF hackathon with Google DeepMind, ElevenLabs and partners, plus the global Nano Banana hackathon kickoff Sept 6 — community driving rapid image/video experiments and short films SF hackathon photos (packed house) Nano Banana Hackathon launch (ElevenLabs/Google) @DynamicWebPaige.
Seedream 3.0 flagged as deprecated on Image Arena; v4 appears imminent
Image Arena rankings briefly listed "Seedream 3.0 (Deprecatd)" (new tag), an indication from the tracker that ByteDance/Seedream may be preparing a new v4 release or update — watch for a formal preview or replacement Image Arena rankings (Seedream 3.0 deprecated) Follow‑up note (context link).
ComfyUI ships Vue3 renderer; ByteDance USO workflows surface in streams
ComfyUI unveiled a new Vue3 rendering system and community streams demonstrate ByteDance USO model workflows inside ComfyUI — meaningful for builders integrating USO-style models into node‑based visual pipelines ComfyUI Vue3 rendering post ByteDance USO in ComfyUI stream.
'Charlie & Lola' character style trend goes viral on TikTok
A user‑shared prompt for generating "Charlie and Lola"‑style character images spread rapidly on TikTok, producing 10k+ videos in a day and showing how simple comment prompts can trigger mass visual trends and commercial product ideas Viral trend analysis (tweet thread) Comments screenshot / engagement examples.
⚙️ Serving, Throughput and Runtimes
Self‑hosting Kimi K2 with vLLM/SGLang, H200 throughput, vLLM configs, SGLang 10k PRs, inference system diagrams; practical perf notes dominate.
Kimi‑K2 (Instruct‑0905) runs on 8×H200 via vLLM with 24k context and community 440–800 tok/s reports
Users launched moonshotai/Kimi-K2-Instruct-0905 with vllm tensor-parallel=8 and --max-model-len=24000 and reported sustained outputs of ~440–800 tokens/sec on 8×H200 in community livestreams and perf logs; examples include the vllm serve command and a live throughput/cost trace showing ~440 tok/s and per-hour host notes vLLM serve command self‑host perf & cost log deploy success note dspy call example.
Sonoma Sky & Dusk Alpha go live (2M token context) in OpenRouter alpha
OpenRouter released two stealth Sonoma Alpha models (Sky and Dusk) advertising a 2,000,000 token context window and free alpha testing; community mirrors and chat UI integrations appeared immediately, enabling hands‑on testing over the weekend OpenRouter announcement context comment (Cline) AnyCoder listing.
LLM inference anatomy: KV cache manager, scheduler, paged block pool (diagram + blog)
A published diagram and explainer lays out an LLM inference stack: engine core with scheduler, a KV‑cache manager exposing a free_block_queue/block_pool, and paged KV memory on GPU plus CPU indexing — useful for engineers optimizing latency and KV paging for long‑context models Inference diagram (thread) Full blog post.
H200 cluster price spread observed: 16×H200 $3.52/hr/GPU ($56.30/hr) vs reported 8×H200 spot instances
Cloud/UI screenshots from community posts show a 16×H200 SKU priced at $3.52/hr per GPU (total $56.30/hr) on one provider panel, while separate PrimeIntellect/spot listings and user instance pages show much lower 8×H200 spot totals — underscoring large provider and spot vs reserved cost variance to plan self‑hosting budgets 16×H200 cluster UI (pricing) PrimeIntellect 8×H200 instance PrimeIntellect Environments note.
🏗️ AI Infra, Capex and Capacity
OpenAI cash burn to $115B through 2029, Oracle/Google Cloud ties, data center GW shortfall, secondary equity transfer crackdown; macro infra dominates.
OpenAI ups multi‑year cash burn to $115B; revenues & compute forecasts revised
OpenAI told investors it now expects cumulative cash burn through 2029 of about $115B and has raised longer‑term revenue targets (2030 revenue guidance moved toward ~$200B); management attributes the gap largely to steep compute and data‑center spending, including multi‑cloud capacity deals and chip investments Financial projections chart Reporting: cash burn & revenue update.
OpenAI secures 4.5 GW Oracle capacity as part of multi‑cloud expansion
OpenAI has deepened infrastructure deals, planning ~4.5 GW of capacity with Oracle and bringing Google Cloud into its supplier mix to meet surging inference/training demand; the moves are cited as part of why OpenAI’s near‑term compute spend projections rose materially Deal & capacity reporting Cash/compute projections (context).
OpenAI teams with Broadcom on custom AI chips; 2026 production target
OpenAI has partnered with Broadcom to co‑design bespoke AI accelerators and related data‑center hardware, with first production signalled for 2026; Broadcom also has a separate ~$10B hardware commitment reported for an unnamed customer, highlighting large‑scale capacity plans Partnership report (announcement) Industry writeup on chip plan Broadcom $10B hardware deal (context).
Wall‑street estimates show multi‑GW US data‑center shortfall through 2028
Estimates compiled from broker analyses project US data‑center demand rising to ~77.1 GW by 2028 versus supply ~66.7 GW, leaving annual shortfalls generally in the ~8–11 GW range (2023–2028); AI and cloud compute are primary demand drivers, pressuring capacity and electricity markets Data‑center GW shortfall chart (broker est.) Electricity demand breakout: data centers vs EVs (Ember).
OpenAI declares unauthorized secondary transfers (SPVs, tokens) invalid
OpenAI posted an official notice that any direct or indirect transfer of OpenAI equity without written consent—covering SPVs, tokenized instruments, or other circumventions—will not be recognized and can carry no economic value to purchasers, warning of legal and economic risk for secondary buyers Summary & analysis of OpenAI notice OpenAI policy page (unauthorized equity transactions).
OpenAI reorganizes Model Behavior team; Joanne Jang to found OAI Labs
OpenAI folded its ~14‑researcher Model Behavior team into the larger Post‑Training organization and announced Joanne Jang will depart to form OAI Labs, a research group focused on inventing and prototyping new interfaces for human–AI collaboration Reporting: team reorg & OAI Labs Follow‑up reporting / thread.
📊 Benchmarks, Evals and Observability
Hot discourse: “evals vs observability” debates, OpenAI hallucination paper (abstention), FutureX wins, NYT Connections, longform writing ‘slop’, W&B Weave traces.
OpenAI: hallucinations = test-taking; credit 'I don’t know' (abstention)
OpenAI paper (2025-09-04) argues hallucinations stem from training/evaluation that reward guessing and proves error bounds tied to the data's singleton rate; empirical/analytic results show calibrated abstention dramatically reduces wrong answers (examples contrasted with ~52% vs 1% abstention) OpenAI paper screenshot announcement thread summary theorem & bounds excerpt.
Grok 4 wins FutureX; GPT‑5 Pro & ChatGPT Agent behind
FutureX benchmark updates report Grok 4 placed first both weeks, with GPT‑5 Pro and ChatGPT Agent ranking second and third—a notable public leaderboard lead for Grok 4 on reasoning-focused tasks FutureX leaderboard post newsletter summary (weekly roundup).
'Evals Are Dead' sparks community pushback and hybrid proposals
Hamel Husain's video 'Why Evals Are Dead' (2025-09-05) generated extensive pushback and discussion: proponents of observability tout traces/QA, while many experts call for combining offline evals with production A/B tests and observability to measure real product truth video release (announcement) hybrid-evals thread counterpost defending evals.
Kimi K2‑0905 gains +3.8 pts on Extended NYT Connections (19.8→23.6)
Moonshot's Kimi K2‑0905 (noted in community writeups) improved on the Extended NYT Connections metric for the most-recent 100 puzzles, rising from 19.8% to 23.6% versus the prior Kimi K2 release; community posts also document live self-hosting/demo activity Benchmark chart (NYT Connections) Kimi K2‑0905 release notes self-host/download report.
Slop profiles: Sonoma Sky Alpha & Qwen‑3‑Max show repetition and degradation
Longform creative-writing benchmarks and 'slop profiles' reveal that Sonoma Sky Alpha exhibits concentrated repetitive tokens/phrases while Qwen‑3‑Max shows measurable degradation (reported score ≈57.4 and degradation ≈0.567), highlighting long-context drift in creative outputs Slop profile & sample outputs Qwen‑3‑Max preview & benchmark table.
W&B Weave traces add RL step-level dashboards (reward/entropy/logp)
Weights & Biases previews Weave traces (coming to the workspace) with RL-focused per-step traces and dashboards—showing values such as rollout/step/reward=0.922668, entropy=1.426735, logp≈-0.184745—aimed at richer observability for agent debugging Weave traces UI screenshot W&B announcement (preview).
🧠 New Models, Weights and Roadmaps
Heavy day: Sonoma Sky/Dusk (2M ctx stealth, likely Grok), Kimi K2‑0905 (open weights, 256k ctx), Qwen‑3‑Max preview (1T‑class), Hunyuan models trending, Embedding Gemma 300M; several leaderboard/longform notes.
OpenRouter Sonoma Alpha twins: 2M token context, free alpha
OpenRouter quietly released two Sonoma Alpha frontier models (Sonoma Sky Alpha and Sonoma Dusk Alpha) with an unprecedented 2,000,000‑token context window and free alpha access for community testing; prompts/completions are logged by the model creators for feedback during the test period OpenRouter announcement Early community post / model cards.
Qwen‑3‑Max preview: trillion‑scale model shows big reasoning and cost wins
Alibaba released a Qwen‑3‑Max preview (1T‑class). Zhihu benchmarking by toyama nao finds major reasoning/knowledge gains, a ~15% drop in average token cost vs the 235B sibling, and reported throughput of up to ~35 TPS with average response times of 49–113s in some runs Zhihu benchmark summary Smol_AI recap (news aggregator) Newsletter summary (market context).
Kimi K2‑0905 open weights: 256k context and measurable benchmark improvements
Moonshot's Kimi‑K2‑Instruct‑0905 is circulating (open weights/self‑host demos) and is reported to double context length to ~256k from 128k; community benchmarks show gains (Extended NYT Connections newest‑100: Kimi K2 → K2‑0905 from 19.8% to 23.6%) and multiple self‑hosting experiments are live Kimi K2 notes (context claim) Extended NYT benchmark (Kimi improvement) Self‑host / dspy usage example.
Nano‑Banana: Gemini API free weekend + rate limit increase to 200 req/day
Google opened Nano‑Banana (gemini‑2.5‑flash‑image‑preview) to a free weekend tier and temporarily raised rate limits (noted as 200 image requests/day) to encourage building and hackathon activity over the weekend Gemini free‑tier announcement Increased capacity / 200 reqs/day Official note: Nano Banana availability.