GPT-6 Astra arrives across models, APIs, and SDKs

The Astra platform moment

OpenAI's week was dominated by a single story: GPT-6 Astra arriving in ChatGPT with improvements across coding, research, computer use, and multi-step work. The model can generate documents, spreadsheets, and presentations from templates, then adapt when requirements change mid-task. Access begins with a limited set of organizations, with broader rollout planned over the coming days. The ChatGPT release also introduces new safety monitoring that can pause or stop conversations if agent instructions appear to be misread.

The API release of GPT-6 Astra for general use is the more consequential story for developers building on the platform. OpenAI has removed custom reasoning effort controls — only preset levels are supported — and dropped custom temperature, top_p, and log probabilities. Tool calling now requires the Responses API, a clear directional signal that the older Chat Completions path is no longer the target for frontier agent work. The model also ships with misalignment monitoring built in, a safeguard that fits the longer-running agentic workloads Astra is designed for.

Support landed across the stack within hours. The Node SDK v7.10.0, Python SDK v3.8.0, and Go SDK v3.56.0 all added gpt-6-astra support, and Codex followed with Bedrock model picker support and a fix making Astra the bundled default in the CLI model picker.

Steering long-running agents

Alongside the model itself, the Responses API gained controls clearly designed for Astra's longer agentic loops. Async tool results and mid-turn steering let the model keep working while your application runs function or custom tools, then receive results as they become available. Developers can send additional instructions over WebSockets while a response is in progress, and adjust reasoning effort mid-conversation — a meaningful upgrade for anyone building agents that need to course-correct without restarting a turn. The API also separated rate-limit from overload errors with distinct slow_down and server_is_overloaded codes, giving client code clearer retry semantics under load.

Claude Code hardens auto mode

Anthropic spent the week on security and reliability. Claude Code v2.1.257 added a Containment Escape rule that stops auto-approving cloud credential fetches, egress evasion, and cross-tenant reach — a meaningful hardening for unattended or lightly supervised sessions. It also made Claude Fable 5.1 the default Fable model with 1M context. The follow-up v2.1.260 closed several sandbox permission holes, including Edit/Write/Read rules with parentheses being dropped as invalid, which could leave read-only folders writable. A new fullscreen /diff panel shows uncommitted changes as Claude edits, and the v2.1.259 release added org-managed MCP servers and fixed concurrent sessions silently reverting each other's config state. Smaller fixes covered macOS 12 launch regressions and bash output limits in v2.1.258 and v2.1.261.

Healthcare and enterprise reach

OpenAI pushed hard into regulated verticals. ChatGPT now connects to EHR and healthcare data, with healthcare plugins for Enterprise and Edu adding read-only access to Epic records and public health data, plus a Healthcare Public Data plugin for clinicians. The Zendesk and OneNote plugins landed in ChatGPT and Codex for Business, Enterprise, and Edu accounts. Separately, ChatGPT Ads passed a $1B run rate with expansion to 40+ countries.