OpenAI DevDay 2026: GPT-6.1 Sol, Astra Ultrafast, the Agents API and Codex Cloud
OpenAI made more than 20 announcements at DevDay on September 29. The one developers will feel most is GPT-6.1 Sol (gpt-6.1-sol), released one week after GPT-6 Sol. OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one fifth of Astra's price: $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens, with a context window of roughly 1.05M tokens. It is available in the API and in ChatGPT Work and Codex for paid plans. For latency-sensitive work, Astra Ultrafast is a premium inference tier that serves GPT-6 Astra at up to 300 tokens per second, about six times the standard rate.
On the platform side, the Agents API is now in public beta. It exposes the managed Codex harness (sessions, orchestration, context compaction, recovery, code execution, file editing, MCP connections, agent delegation and computer use via an OpenAI-hosted browser) behind one endpoint. A partnership with AWS adds Bedrock Managed Agents, which runs the same harness with inference on Bedrock so data stays in AWS. Codex Cloud adds reusable remote dev environments with repos, dependencies and access settings preconfigured. There is also a desktop code-review experience (GitHub GA, GitLab in preview) and Codex Security Cloud for scheduled repository scanning.
The Codex CLI kept shipping alongside: 0.158.0 added MCP server OAuth client secrets (codex mcp add --oauth-client-secret), bearer-token secure WebSockets and terminal input approval on by default for elevated permissions. 0.159.0 introduced an opt-in instant_interrupt for steering the model mid-response, better Mermaid rendering and Windows sandbox fixes, and 0.159.1 added GPT-6.1 Sol to the bundled and Amazon Bedrock catalogs. OpenAI also previewed a Decisions API for classification and routing, covered in this week's AI Dev Patterns digest.
Read more — Runtime Wire
Claude Code 2.1.284–2.1.288: Sonnet 5.5 as Default Sonnet, Claude Mods, and Provider Lockdown
Claude Code had a busy week. 2.1.284 made Claude Sonnet 5.5 (claude-sonnet-5-5) the default Sonnet model, with a 1M-token context window at $2/$10 per million tokens and $0.20 per million for cache reads. The same release added /mcp reconnect all for retrying every failed MCP server at once, /rate-limit-options for subscribers, and private_key_jwt certificate client authentication on the Claude apps gateway. 2.1.285 added an allowedProviders managed setting to restrict which API providers a fleet can use, a CLAUDE_CODE_DISABLE_WEB_FETCH switch, claude plugin configure <plugin>, and claude --desktop to open the desktop app on the current directory. It also limited background-command time limits (30 minutes by default, 2 hours max) to unattended sessions.
The headline feature of 2.1.287 is Claude Mods: plugins can now change deeper behavior than commands and hooks allowed. The first built-in mod, "You should know", runs a side agent that flags things you or Claude might be missing (/plugin enable cc-plugin-you-should-know@builtin). Interactive sessions now start in auto mode when no permission mode is configured, and Ultracode became its own toggle instead of forcing xhigh effort. 2.1.288 (October 2) added --max-findings <n>|all to /code-review, Up-arrow recovery of prompts cleared with Ctrl+C, and re-authentication prompts when an MCP server asks for more OAuth scope.
The bug-fix lists matter as much as the features for anyone running Claude Code unattended. Non-interactive sessions and subagents now continue from partial responses after mid-stream API timeouts, long conversations auto-compact instead of failing with "Prompt is too long", and LSP calls time out after 60 seconds instead of hanging. Several safety regressions were fixed: dangerous rm commands had been losing their safeguards inside bash -c scripts and output redirections, PreToolUse and PermissionRequest hooks were skipped on serialization errors, and MCP errors and logs could leak credential values.
Read more — Anthropic
GitHub Copilot Adds Desktop Computer Use, Code-Defined Dynamic Workflows and a Code Review API
GitHub shipped three agent-focused Copilot changes on October 1–2. Computer use (public preview) lets Copilot CLI and the Copilot app on macOS and Windows operate desktop applications: it reads accessible content and screenshots, clicks controls, types, scrolls, drags and moves between apps. It is aimed at legacy and GUI-only tools that have no API or CLI. You turn it on with /computer on (/computer show and /computer off manage it). Copilot asks for approval before controlling each app, you can review or reset "always allow" apps, and organization policy can disable the feature entirely.
Dynamic workflows (public preview, all plans) are programs that combine deterministic steps (commands, tools, external services) with the work of one or more agents. Stages can run sequentially or in parallel, pass structured outputs between them, include verification steps, ask for user input, and pause at checkpoints for review. GitHub contrasts this with /fleet, where Copilot decides how to coordinate subagents: here the process is defined in code. Workflows are available in the Copilot app, in Copilot CLI behind --experimental, and in the Copilot SDK.
Finally, Copilot code review can now be requested through the REST and GraphQL APIs with a per-request effort level, so teams can call it from their own tools and pipelines. The "Default" effort level now maps to Balanced for new and existing repositories. Anyone who explicitly chose "Lite" keeps it, and the level remains configurable at enterprise, org, repo and user scope. GitHub also deprecated a set of older models across all Copilot experiences on October 2.
Read more — GitHub Changelog
Links & Sources
You May Also Enjoy
