Gemini 2.5 Pro and Flash Requests Reroute to Newer Models on Dec 1

Act before next issue

On our radar

• In review, not yet published

  None of these drafts has passed our source review, so we report no figures from them yet: GPT-6 Rosalind, ChatGPT Work and Codex sharing one usage allowance, Claude message threads, Gemini API Computer Use going GA, the ChatGPT share link takeover, Grok Bot's weekly limits, Claude Code's game cloning guardrail, LiteLLM's release cadence and LTS pilot, the VOLTA infostealer on OpenRouter, LongCat-2.5, GPT-6 Cyber, GPT-Image-2.6 Sunburst and Flare, the GPT-5.6 Sol promo end date, GPT-6 Astra's safety research underperformance, the Kimi K3.1 Pro weights, Tencent Hy4, Ripple-65B Preview, Inkling 2 Small, the ChatGPT Lockdown Mode bypass, the Haiou Codex adware, LLMjacking charges, Perplexity Teammates, Claude Personal Finance and ElevenLabs' 2 November changes.

• Two act drafts still failing review

  Our stories on Claude Code sending a user's account email in the User-Agent header and on the Mistral Large 4 weights and API price both failed source review on 6 November. They stay drafts until fixed.

• GitHub pull_request_target brownouts on 9 and 16 November

  GitHub Actions blocks pull_request_target by default in public repos from 2 December 2026, with brownouts on 9 and 16 November from 16:00 to 18:00 UTC. Held until a directory AI review or agent Action documents that it depends on pull_request_target.

• OpenAI models in Cursor: 12 November still unconfirmed

  The proposed 12 November cutoff for OpenAI models in Cursor still has no primary source. OpenAI's help page on it returned an empty body again on 6 November.

• Training and data claims with no policy text

  Reports that OpenAI re-enabled training on Codex chats, that Gemini trains on subscribers' chat history, and that Anthropic blocks users over work on training larger models rest on Hacker News titles or an X post. Held until a policy page or vendor statement is fetched.

• Research and acquisitions with no product change yet

  Anthropic's FlashProve provable inference prototype, OpenAI's activation classifiers for safety, OpenAI's deal for Pierre Computer Company, Meta buying Conductor and Baidu's ERNIE-5.5-Flash-Base checkpoint. Held until one changes a product, an API or a price.

• Narrow access and pricing changes

  OpenAI's 10% uplift on FedRAMP endpoints from 1 November and Daybreak Red arriving in AWS GovCloud through Bedrock. Held until either reaches readers outside government endpoints.

• Governance and litigation held at headline level

  Anthropic's revised S-1/A on its Responsible Scaling Policy, publishers suing Brave over selling search content to AI companies, China's first frontier AI export controls, the IETF training preference signal and Google turning off Search for schools that restrict its AI tools. Nothing read changes a reader's setup yet.

• Single reports and unopened posts

  An alleged leak of an Anthropic Mythos-class model on Hugging Face, a Codex user report that the agent destroyed systemd and containers on a reboot task, a Replit customer's account of an arbitration opt-out closing 8 November, Cursor's Build Live Apps changelog entry and two Claude Blog enterprise posts. Held until a primary source is read.