GPT-6 goes to the life sciences as the SDKs converge on agents

GPT-6 goes to the bench

The week's biggest story is OpenAI's decision that its frontier reasoning model needed a domain-specific variant. GPT-6 Rosalind Research brings GPT-6 reasoning to multistep life sciences research with safeguards built for biological work — and access is gated accordingly: organizations and individual users need a separate Research approval, and prior Discovery approval doesn't count. Enterprise customers get it through a two-week Admin Preview that starts off by default, while approved Business customers can pick it up directly in Work and Codex. GPT-5.5 Rosalind Research stays available for existing workflows throughout, so nothing breaks for teams that don't migrate yet.

Anthropic is telling a quieter version of the same story: its scientists reported early progress from deep research partnerships with six drug developers, framing Claude as a tool for accelerating medicine discovery rather than a product launch.

Agents become the unit of work in the SDKs

The OpenAI SDKs converged on the same set of primitives across languages this week. The Node SDK led with scheduled agent automations and delegation controls — a client.beta.agents.automations namespace plus an automations agent-tool variant — and a breaking change to Live Decision events that will catch anyone reading item.content or item.role directly. It followed with agent session review and tool settings, context compaction, and experimental time awareness, and added typed Responses WebSocket warmup. The Python SDK landed the same foundation — delegation controls and prompt cache lifetimes first, then response tool usage models and Realtime typed references, followed by its own agent context compaction and session review settings. Go tracked in parallel, adding delegation controls and cache lifetimes and agent compaction, session review, and SIP DTMF delivery events, with image and web search usage reporting in between. Prompt-cache TTL is now something callers set explicitly rather than inherit: Python exposed the options, Go too, and Node clarified the breakpoint semantics.

All three SDKs also picked up the gpt-6.1-terra model identifier — the same addition in Python and in Go — and Go shipped a Live WebSocket guide. Python wrapped up with WebSocket behavior preservation and realtime auth validation.

Codex and Claude Code harden the local agent loop

Codex CLI added a built-in Vertex AI provider and GPT-6.1 Terra in the model catalogs, turned on memory v2 and mid-conversation reasoning-effort changes by default, and closed a marketplace URL validation gap that could send Git to a different host than the one checked. It then extended Terra to the bundled, Vertex, and Bedrock catalogs and gave transcripts keyboard-driven selection with integrated search.

Claude Code spent the week on trust boundaries and reliability. A write-block setting that refuses edits outside working directories in every permission mode came alongside idle compaction for long VS Code sessions, and a fix so the Write tool can no longer clobber a large CLAUDE.md after seeing only its opening lines. Startup freezes from pathological CLAUDE.md list items and context-limit stalls that now compact instead of stopping rounded out performance work. Opus 4.8 fast mode is deprecated with removal set for November 30 — migrate to Opus 5.5 to keep the speedup — and Claude Code gained subagent effort control and pooled group budgets for the apps gateway.

Product surface and platform plumbing

ChatGPT itself got more conversational infrastructure than model news: texting your dot via iMessage or RCS began a U.S. beta for Pro users, account switching arrived on iOS and Android, iCloud Mail and Calendar became connectable apps, and Free and Go users got more concise, less flattering GPT-6 replies. On the enterprise side, automatic Instant-to-thinking switching was retired in favor of explicit model-picker choices. Usage got more generous and more measurable: Pro and Business Premium plans doubled their Work and Codex allowances, eligible subscribers can claim up to $500 a month in promotional API credits — flagged for Business workspaces and Pro and Business subscribers separately — and the Claude Platform opened a Usage and Cost Analytics API in beta while raising the Messages API request ceiling to 128 MB. A Model Spec reorganized around five core values rounds out the week's policy-facing changes.