Stay ahead on AI.
All in ONE place.

OpenAI's latest models show major gains in low-resource programming languages like COBOL, Fortran, and Visual Basic. GPT-6 Astra achieves an 87.4% average test pass rate on MultiLangBench, up from 26.8% for GPT-5.6 Sol, across 20 languages including COBOL rising from 10.9% to 72.1%. On COBOL migration to languages like Python or Rust, GPT-6.1 Sol reaches 86.6% accuracy at low reasoning effort and 97.9% at higher effort, versus just 5.4% for GPT-5.6 Sol, at 39% lower cost.

Gemini Notebook is rolling out new co-creation features globally that give users more creative control over Slide Decks and Video Overviews. Instead of jumping straight from a prompt to a final output, users can now review and steer content topics, visual style, and format mid-process. A new Review toggle lets users examine outline or producer notes in the studio panel and request changes before anything renders. The update is rolling out to all Gemini Notebook users worldwide over the upcoming week.

Three new OpenAI image model references — aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha — briefly appeared in OpenAI's API before being pulled, suggesting an accidental rollout. The models then surfaced in blind arena tests generating images, with red running faster at lower quality and green slower but sharper. Lumina Bench speculated these are likely GPT-Image-3. No documentation or pricing exists yet, and the blue variant hasn't been tested.

Anthropic is committing at least 5% of its shares to the new Anthropic Public Benefit Fund, a wholly-owned LLC designed to grow alongside the company. An additional 2% of shares will be added at the board's discretion for every 50% increase in company value above a set baseline. Combined with co-founder pledges of 80% of their equity and an employee equity donation match program, roughly 20% of Anthropic shares will be dedicated to public benefit. The fund may support nonprofit donations, direct distributions, and labor market impact initiatives.

DeepSeek has released V4.Flash-1017, a reinforcement learning-powered coding model offering stronger benchmark performance, faster task completion, and shorter outputs than its predecessor. The model supports end-to-end project development, including game creation and 3D modeling, and is optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. DeepSeek also opened its Search and Fetch APIs at $0.75 per 1,000 calls, with free access for API users until October 31, 2026. The model is available on Hugging Face and ModelScope.

Meta has removed language from its commercial terms that experts said allowed the company to scrape news sites and other web content to train its AI models. Cornell law professor James Grimmelmann and Nieman Lab had flagged the clause as potentially superseding robots.txt blocking instructions. Meta VP Andy Stone and spokesperson Daniel Roberts denied any policy change, stating the consolidated crawling terms only applied to ads and business products. Meta says content owners can still use robots.txt to signal their preferences.

Major companies including Yelp, Delta Air Lines, JPMorgan Chase, and Charles Schwab are blocking AI agents from OpenAI, Meta, and Google from accessing their websites and services. Yelp requires agents to pay for its data licensing program, Delta is preventing ticket bookings, and the financial firms cite security concerns. Users of Meta's Muse, xAI's Grok, and OpenAI's Dots have also reported blocks on Amazon, Cars.com, and Amtrak, raising questions about how useful AI agents can truly be.

Perplexity has open-sourced two new Decision Models: pplx-decider-flash-v1.2 at 4B parameters priced at 1 cent per million tokens, and pplx-decider-v1.2 at 27B parameters at 2 cents per million tokens. The 4B model ranks first on Hugging Face's Decision Index among 4B models, while the 27B ranks second overall. Both are available via the Perplexity Decisions API and as open weights on Hugging Face. CEO Aravind Srinivas noted plans to further reduce costs through input caching and serving optimizations.

Anthropic researchers gave Claude plain-language specifications for eight quantum circuit design tasks spanning chemistry, optimization, physics simulation, and number theory. Using a closed-loop process—proposing designs, receiving automated scores based on qubit and gate counts, then iterating—Claude Mythos 5 beat the best published benchmarks in all eight cases, sometimes by an order of magnitude or more. Key wins include making a quantum optimization algorithm's costliest step roughly 12x cheaper and devising a novel modular inversion approach for elliptic-curve calculations. The work was led by Andi Gu through Anthropic's STEM Fellows Program.