OpenAI launched GPT-Rosalind-6 for biology research under a controlled-access program, and the FT reports it has cut its fundraising target to about $20B amid an AI stock selloff. Google opened Android Studio to more than 40 coding agents, including Claude and Codex, via the open ACP protocol. On safety, a trojanized Claude Code build slipped into the Cursor Marketplace, while Anthropic is extending text watermarking to older models ahead of the EU AI Act deadline.
New models
1. OpenAI releases GPT-Rosalind-6 for biology research, for approved customers only
Per its Oct 30 system card, GPT-Rosalind-6 shares weights with GPT-6 Astra and clearly beats version 5.5 on medicinal chemistry (49.6% vs 28.9%) and genomics (37.7% vs 16.7%) benchmarks. OpenAI rates it High for biological/chemical risk, so it is offered only through a trusted-access program with monitoring and required human review before outputs inform experiments.
2. Moonshot closes the Kimi K2 series to new users, steering them to K3.1 Pro
From Oct 31, new API users can no longer access K2 models; existing users keep access until the end of December 2026, when the whole K2 series is retired. Moonshot recommends migrating to kimi-k3.1-pro (1M-token context, native vision), so K2 developers should plan the move early.
3. Android Studio opens up to more than 40 coding agents, including Claude and Codex
At the Android Developer Summit (Oct 31), Google introduced “Bring Your Own Agent” in Android Studio Rabbit 2: agents connect via the open Agent Client Protocol (ACP), read project structure, run apps on the Emulator and read Logcat to verify their own work. The Android CLI also gained official Claude Code and Codex plugins, a sign that major IDEs are moving to swappable agents.
4. Trojanized Claude Code build spread through the Cursor Marketplace
A fake version 2.1.313 of the Claude Code extension was live on the Cursor Marketplace for about 12 hours on Oct 29, adding an attacker SSH key, opening a reverse SSH tunnel and sending host details to a remote server; Cursor estimates at most 5,000 users were affected. Anthropic says its systems were not compromised and shipped a clean 2.1.314; affected users should remove the rogue SSH key and rotate credentials.
5. Anthropic extends text watermarking to older Claude models ahead of the EU AI Act deadline
According to an email to API customers (reported by The Decoder on Oct 31), the imperceptible word-choice watermark will cover Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, applied at model level so it also works on AWS, Google Cloud and Microsoft Foundry. Caveat: on Opus 4.6 and Sonnet 4.6 it only applies at default sampling settings – changing temperature, top_p or top_k can avoid it.
6. OpenAI reportedly cuts its fundraising target to about $20B
Per the FT (cited by CoinDesk late Oct 30), OpenAI lowered its target from $30B at a $1.4T valuation to about $20B at $1.25T, as the Nasdaq 100 fell roughly 10% in three weeks. It suggests caution over AI spending is spreading to private capital markets, just as Anthropic prepares to IPO in November.
7. Apple acquihires agentic data-engineering startup Ascend.io
The deal surfaced in Apple's mandatory EU Digital Markets Act disclosures (9to5Mac, Oct 30): Apple gets the right to hire certain staff and a non-exclusive license to the IP of Ascend, whose agents automate routine data-engineering work. Ascend.io says it is ceasing operations.