Changelog — September 30, 2026

Sonnet 5.5 is now the default Sonnet, and sessions start in auto mode. 2.1.284 adds claude-sonnet-5-5 (1M context, $2/$10 per MTok) as the default Sonnet model on the Anthropic API, and changes interactive terminal and VS Code sessions to start in auto mode when no permission mode is configured, on every plan and provider; permissions.defaultMode still overrides it. 2.1.285 extends that to claude -p and Python Agent SDK sessions on third-party providers or with telemetry off. If you rely on the old default, set permissions.defaultMode explicitly.

New controls in 2.1.285. allowedProviders is a managed setting that limits which API providers a machine may use. CLAUDE_CODE_DISABLE_WEB_FETCH turns off the WebFetch tool. claude --desktop opens the Claude desktop app on the current directory, and claude plugin configure <plugin> shows or saves a plugin's options. Background Bash and PowerShell commands now stop after a time limit (default 30 minutes, maximum 2 hours) and Claude is notified. 2.1.284 also adds /mcp reconnect all and makes Ultracode its own toggle in /effort. 2.1.286 is fixes only, including claude --resume losing turns after a batch of parallel tool calls and a retry on the previous model of the same tier when the API refuses the model your alias resolves to.

Releases → Affects: /claude-cowork/faq/claude-code-undocumented-configuration/, /claude-cowork/setup/, /claude-cowork/pricing/, /commands/

Claude Sonnet 5.5 launched on 2026-09-28 as claude-sonnet-5-5: 1M-token context, 128K max output, $2 input / $10 output per MTok, the same price as Sonnet 5, on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Anthropic calls it the second model in the Claude 5.5 family, a faster, lower-cost complement to Opus 5.5.

Code written for Sonnet 5 can break in five ways. Turn off up-front thinking with thinking: {"type": "between_tools"} (at high effort or below) instead of "disabled"; forced tool use (tool_choice types any and tool) returns a 400 error; thinking blocks work only in the account that produced them; the earlier computer_20251124 tool is not accepted on the Claude API and Google Cloud; and the advisor tool rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors. Anthropic publishes a migration guide with before-and-after requests.

Sonnet 4.5 is on the retirement path. On 2026-09-30 Anthropic announced the deprecation of claude-sonnet-4-5-20250929, with retirement on the Claude API scheduled for 2026-11-30, and recommends moving to Sonnet 5.5. Check any config or guide still naming it, including OpenClaw model lists.

API release notes → Affects: /claude-cowork/, /claude-cowork/vs-api/, /claude-cowork/pricing/, /tools/cost-calculator/

Updates are safer, with limits. New verified backups check that every agent database recorded at capture is present, and automatic rollback is available for eligible updates that fail before the new version is allowed to start. Database restoration needs a capture made while OpenClaw was stopped; a failed or unverified update keeps its snapshots and may leave OpenClaw stopped; and snapshots need substantial free disk space. Installs older than June 2026 must upgrade through 2026.9.5 first. After a manual upgrade or a custom container entrypoint, run openclaw doctor --fix before starting.

Removals to plan for. Tasks and TaskFlow panels, commands and APIs are gone; use subagent controls, ACP sessions, background processes, automation history or Lobster pipelines. Tool Search now uses separate search, describe and call tools, and its JavaScript tool tool_search_code is removed (Code Mode replaces it); normal updates migrate the settings, but an unmigrated manual upgrade can stop the Gateway from starting. The default allowance for pending unauthenticated connections per IP rose from 32 to 128, which administrators can lower. The release also adds OpenAI's Agents API runtime and Sign in with ChatGPT (Beta).

Release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/troubleshooting/, /openclaw/security/, /openclaw/faq/openclaw-rollback-after-update/

Stable promotion of the 1.4.1-rc.2 release candidate. Gmail and Google Calendar can now be activated when the operator supplies the Google OAuth client through the web UI instead of environment variables; before, sign-in completed, activation failed and the failure revoked the new credential, so every retry re-consented and failed again. Administrator configuration now takes precedence over deployment (environment or config.toml) client material. The WASM runtime moves to Wasmtime and WASI 48.0.3, fixing recent guest fuel-accounting vulnerabilities and a WASI filesystem timestamp panic, so operators running untrusted extensions should update.

Releases → Affects: /ironclaw/, /ironclaw/setup/, /ironclaw/security/

OpenAI released GPT-6.1 Sol (gpt-6.1-sol) on 2026-09-29 for complex coding and professional work at a lower cost than GPT-6 Astra. It is priced like GPT-6 Sol: $2 input, $0.10 cached input, $2.50 cache write and $10 output per million tokens for prompts up to 272K tokens, with a 1.05M context window and 128K max output. reasoning.effort supports low, medium (default), high, xhigh and max; use the Responses API for tool calling. The same day OpenAI added computer use to the Agents API (an OpenAI-hosted browser, with website approvals and sign-in handled by your application) and Ultrafast mode for GPT-6 Astra (service_tier: "ultrafast"). Our cost calculator now lists GPT-6.1 Sol and treats GPT-6 Sol as the previous tier.

Cost calculator → Affects: /tools/cost-calculator/, /chatgpt/pricing/