Google ships Gemini 3 Flash as default across Search and Gemini app

Google released Gemini 3 Flash on May 17, making it the new default model in the Gemini app and AI Mode in Search. The model hits 90.4% on GPQA Diamond and uses 30% fewer tokens on average than Gemini 2.5 Pro, pushing frontier reasoning into the free tier. Developers can access it via Google AI Studio, Vertex AI, Gemini CLI, and Android Studio.

Gemini 3 Deep Think upgraded, now opening API access to researchers

Google's May 17 Deep Think update sets a new mark of 48.4% on Humanity's Last Exam (without tools) and 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation. Until now Deep Think was locked to Google AI Ultra subscribers in the Gemini app; for the first time, scientists and enterprises can register for early API access. Early testers include a Rutgers mathematician who used it to catch a logical flaw that passed human peer review.

Anthropic partners with Gates Foundation for $200M, pledges to stay ad-free

Anthropic announced a $200 million partnership with the Gates Foundation on May 14, aimed at expanding AI access to underserved communities. The same week, Anthropic published a public commitment that Claude will remain ad-free, arguing advertising incentives are incompatible with a genuinely helpful assistant — a direct contrast to OpenAI's simultaneous rollout of a self-serve Ads Manager for ChatGPT.

OpenAI adds personal finance to ChatGPT via Plaid integration

OpenAI launched a personal finance experience inside ChatGPT on May 15, letting users connect bank and investment accounts through Plaid to get budgeting help and spending analysis. The feature is rolling out gradually to eligible users and explicitly cannot move money, place trades, or give regulated financial advice. It extends ChatGPT's push into high-frequency consumer use cases after the Ads Manager and Deployment Company launches earlier this month.

GPT-5.5 Clears Human Baseline on OSWorld-Verified Desktop Agent Benchmark

GPT-5.5 scores 78.7% on OSWorld-Verified—a benchmark measuring autonomous operation of real desktop environments—surpassing the 72.4% human expert baseline and edging out GPT-5.4's prior 75.0%. The model also hits 84.9% on GDPval, which tests knowledge-work task completion across 44 occupations, and 98.0% on Tau2-bench Telecom without prompt tuning. The primary caveat: Claude Mythos Preview still leads on SWE-bench Pro (77.8% vs. 58.6%), so GPT-5.5's lead is specific to computer-use and workflow automation rather than deep software engineering.

Cursor Composer 2.5 Matches Frontier Coding Models at a Fraction of the Cost

Cursor shipped Composer 2.5, built on Kimi K2.5 and trained on 25× more synthetic coding tasks than its predecessor, with reported benchmark parity against Claude Opus 4.7 and GPT-5.5 on standard coding evaluations. The release marks the first time a dedicated coding IDE model—rather than a general-purpose frontier model—has reached top-tier SWE-bench-class performance, at substantially lower inference cost. Independent score verification is still pending; Cursor's self-reported numbers should be treated with that caveat until third-party runs are published.

FORGE gives agents self-evolving memory with zero weight updates

FORGE, posted to arXiv on May 18, lets agent populations broadcast and absorb experience across instances via a shared memory substrate—no gradient steps required. The system sidesteps the usual fine-tuning bottleneck by treating memory as a mutable population resource rather than baked-in weights. If the approach generalises, it could cut the cost of adapting deployed agents to new tasks significantly.

Replay buffers beat on-policy sampling for LLM RL post-training

A paper highlighted in Air Street's May 2026 State of AI analysis shows that strict on-policy data collection is suboptimal when generation cost is high, and that a well-designed replay buffer—balancing staleness variance, sample diversity, and compute—can cut inference compute during RL post-training without hurting final policy quality, in some cases improving it. The result ports two decades of mainstream RL practice directly into the LLM training stack and challenges a core assumption of current RLHF and RLVR pipelines.

Georgia governor signs AI chatbot safety bill into law

Georgia Governor Brian Kemp signed an AI chatbot safety bill this week, making Georgia one of a growing number of states enacting legislation specifically targeting companion and commercial chatbot interactions. The bill joins similar laws advancing in Oklahoma and Hawaii, which both cleared their respective chambers this month. This wave of state chatbot laws is accelerating as Congress has so far declined to pass any comparable federal standard.

House bill proposes standalone federal moratorium on state AI laws

Rep. Baumgartner introduced H.R. 5388, the American Artificial Intelligence Leadership and Uniformity Act, which would establish a temporary federal moratorium preempting state laws that restrict AI models and systems engaged in interstate commerce. The bill has been referred to the House Committee on Science, Space, and Technology. It arrives as states continue passing enforceable AI rules and the Senate has twice rejected similar moratorium language inserted into budget reconciliation bills.