GPT-6 Astra and Claude 5.1 reset the frontier in one week

Two frontier models land within days of each other

OpenAI's GPT-6 Astra is the week's dominant story, arriving on the API with a broader capability set for reasoning, coding, computer use, and document creation — spreadsheets and presentations from your own templates included. The tradeoff is control: custom reasoning effort, temperature, top_p, and logprobs are gone, only preset effort levels remain, and tool calling now requires the Responses API. Astra also carries misalignment monitoring that can pause a conversation when agent instructions look misread, and the ChatGPT-side rollout is limited to a set of organizations for now, with wider availability promised over the coming days.

Anthropic answered with Claude Fable 5.1 and Claude Mythos 5.1, launched for coding and knowledge work with a 1M-token context window on by default, 128k max output tokens, and cache reads cut to $0.25 per MTok. Adaptive thinking is now the only thinking mode, tool_choice types any and tool return 400 errors, and the models require 30-day data retention with no zero-retention option. For enterprises that need the opposite, Anthropic announced Enterprise Frontier Safeguards, which stores misuse-monitoring data in customer-controlled cloud infrastructure under the customer's own keys, with zero data retention on Fable 5 and 5.1 until it rolls out this fall.

The API grows up around long-running agents

Astra's real unlock is infrastructure for work that outlives a single request. The Responses API gained async tool results and mid-turn steering, including reasoning-effort changes mid-conversation and in-flight WebSocket updates, and error handling got sharper: 429 slow_down for traffic that ramps too fast versus 503 server_is_overloaded for genuine capacity pressure, both respecting Retry-After. Anthropic's ant CLI 1.30.0 brought the same discipline to configuration, with ant apply managing agents, environments, skills, and deployments from repo files and writing a claude-lock.json lockfile for CI.

The official SDKs kept pace, with Node, Python, and Go all adding Astra support and a canonical security model page. Node separately stopped Responses WebSockets from following redirects, and Go surfaced a new compute_units field in usage responses.

Coding agents race to adopt the new models

Claude Code shipped a dense run of updates, most of it hardening. It made Fable 5.1 the default Fable model and added a Containment Escape rule so auto mode stops approving cloud credential fetches and cross-tenant reach, then closed several sandbox permission holes and added a fullscreen /diff panel. Organizations can now push HTTP/SSE MCP servers to every user via managedMcpServers, and there's more room for command output with configurable bash and task output limits up to 128K characters plus a /skill-doctor command. Older fixes restored macOS 12 launch and Bash execution on Macs with relocated task directories.

Codex spent its week wiring Astra in: API configuration without changing the default model, then Bedrock model picker support and making Astra the bundled default. It also removed the deprecated codex mcp-server command and binary, so integrations launching it must move to the Codex app server, while vim mode gained undo/redo and the plugin CLI learned remote marketplaces. Its iOS app added a Priority view and cross-host attachments.

Healthcare, ads, and the enterprise surface

OpenAI pushed hardest into regulated data: ChatGPT can now connect to trusted healthcare data sources for clinician context, with a Healthcare Public Data plugin aggregating nine read-only public sources and Epic EHR access for HIPAA-enabled workspaces. On the commercial side, ChatGPT Ads crossed a $1B annualized run rate and expanded self-service to 40+ countries. Business and Enterprise tiers both picked up Zendesk and OneNote plugins plus external Site sharing.