Changelog — October 30, 2026

GPT-6.1 Terra (gpt-6.1-terra) was released on 2026-10-26 on the Responses and Chat Completions APIs, described by OpenAI as "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing." OpenAI's pricing page lists it at $1.00 input, $0.05 cached input and $5.00 output per million tokens, rising to $2.00, $0.10 and $7.50 once a request passes 272K input tokens. It sits between GPT-6 Luna ($0.10 and $0.50) and GPT-6.1 Sol ($2 and $10), and prices are unchanged for the other GPT-6 models. The pricing page gives no context window. OpenRouter also lists a gpt-6.1-terra-pro, but it is on neither OpenAI's pricing page nor its changelog, so we have not tracked it.

Also. On 2026-10-28 OpenAI said eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in free API credits as a subscription benefit. On 2026-10-27 Google released the Gemini API CLI (gemini-api) in beta, which it describes as a stateless, zero-dependency command-line client for coding agents, scripts and CI that works with the Interactions API to generate text and media, analyse files and run Gemini managed agents. xAI's Grok Bot docs overview was updated on 2026-10-28 without naming a new release.

OpenAI pricing → Affects: /chatgpt/pricing/, /chatgpt/api-vs-chat/, /tools/model-pricing/, /tools/cost-calculator/

2.1.314 adds permissions.blockWritesOutsideWorkingDirectories: Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode. It also brings idle compaction to VS Code and Claude Desktop's Code tab (a conversation over 200K tokens left idle is summarised, and the summary is used on your next message if the prompt cache has expired), and from 2.1.314 idle compaction also runs for subscribers with telemetry off and on Bedrock, Vertex, Foundry and LLM gateways once the one-hour prompt cache is on. Fixes include Bedrock sessions with many images failing every turn, Foundry sessions failing when the resource has no deployment for the default model (it now falls back to an older model of that family), and claude ultrareview and claude remote-control now refusing an untrusted folder at once when CI is set.

A reversal. 2.1.312 stopped sending the claude_code.safety_event OpenTelemetry event that 2.1.307 had added, so do not build monitoring on it. 2.1.312 also adds oidc.user_id_claim to the Claude apps gateway, makes Bedrock, Vertex, Foundry and Mantle use the shorter system prompt for model names Claude Code doesn't recognise, such as custom deployment names (set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 to keep the full prompt), and adds CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1 to turn off automatic model switches after a flagged message. Prompt and agent hooks now run at most 24 at a time per session.

Other changes to check. 2.1.311 makes ultrathink run its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic's API, and on Bedrock, Mantle, Vertex and Foundry loads model settings from a signed file on downloads.claude.ai (CLAUDE_CODE_MODEL_CATALOG=off turns it off). 2.1.314 changes FORCE_HYPERLINK=false (and no or off) to turn hyperlinks off instead of on, and makes the file tools ask before editing the local copy of your organisation's server-managed settings. 2.1.313 does not appear in the changelog.

Changelog → Affects: /claude-cowork/setup/, /claude-cowork/faq/effort-levels/, /claude-cowork/faq/claude-code-undocumented-configuration/, /commands/

Usage and Cost Analytics API (beta, 2026-10-29). It reports token usage and cost for Claude Console and Claude Enterprise organisations, with filtering and grouping by model, workspace, API key and the user or service account behind a request. It uses the header analytics-usage-cost-2026-08-24 and the endpoints /v1/organizations/analytics/usage and /v1/organizations/analytics/cost. The older Usage and Cost API, and the usage_report, cost_report, user_usage_report and user_cost_report endpoints of the Claude Enterprise Analytics API, are now "legacy" but remain supported.

Larger requests, and a Claude Tag API. On 2026-10-28 the maximum request size for the Messages and Token Counting APIs on the first-party Claude API rose from 32 MB to 128 MB. On 2026-10-23 a Claude Tag API (beta, Claude Enterprise only, header experimental-claude-tag-2026-10-21, endpoints under /v1/organizations/claude_tag/) arrived for reading and changing Claude Tag settings on Slack channels and workspaces; it needs an Admin API key with the read:claude_tag or write:claude_tag scope. The notes announce no deprecations, and the Claude apps release notes list nothing new since Claude Duet on 2026-10-21.

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

2026.10.8 and 2026.10.9 are patch releases with no new features, and the first one needs a Gateway restart. 2026.10.8 patches the secret egress proxy and its exported trust bundle for the eMudhra certificate issue ("Restart the Gateway to apply the change"; apps with their own trust stores need their own updates). Sandboxed child agents can no longer launch a CLI model outside their sandbox, and agents can no longer bypass browser-tool restrictions through the generic Nodes tool: use an API-backed child model or change the sandbox policy if you need a host CLI. POSIX exec allowlists no longer automatically trust test -v or [ -v, so some legitimate commands may need an allowlist entry or approval.

Migrations to check. Relative session-store paths could create a second database and break follow-up chats; if you already have split stores, stop the Gateway, run openclaw doctor --fix and restart. Out-of-process clients of the optional standalone loopback browser API that use trusted-proxy authentication now need an explicit gateway.auth.password. A new opt-in key, plugins.entries.anthropic.config.managedNativeTools, exposes Agent and Workflow native tools to managed Claude CLI turns; the default exposes none. 2026.10.9 updates fast-uri, MSAL Node and Nodemailer for URI-parsing, credential-routing and IMAP sender-spoofing issues, hides recognised credentials in MCP command-line inspection, applies workspace-only limits to Gateway-built delegation and media tools, and says malformed schema identifiers that used to be accepted may now fail validation. A 2026.10.10 beta line (beta.4 and beta.5) is on the beta channel.

Release notes → Affects: /openclaw/, /openclaw/security/, /openclaw/configuration/, /openclaw/troubleshooting/

v0.0.140 sends telemetry about completed operations by default. Ordinary installer and CLI operations attempt to send it to NVIDIA's fixed TEST receiver (production delivery is not enabled), and CI operations do the same with a source-run label. Set NEMOCLAW_DISABLE_TELEMETRY=1 in each shell before installing or using the CLI to turn it off. The release also bounds local-inference onboarding requests with NEMOCLAW_LOCAL_INFERENCE_TIMEOUT (60-second minimum), makes a piped nemoclaw <name> agent message finish within its host-side deadline and be at most 4 MiB, and lets destroy clean up on the sandbox's recorded gateway. The installer now tells a timed-out release lookup, an unavailable reference and a Git verification failure apart; it warns that a managed install or upgrade can remove the existing CLI source checkout before a failed lookup, in which case you fix the lookup and rerun the installer.

v0.0.139 (2026-10-26) makes nemoclaw backup-all report directory and file counts for each native archive, makes nemoclaw host probe list detected GPUs with index, name and memory, supports the Qwen vLLM profile for a 64 GB DGX Spark (automatic selection accepts hosts with at least 60 GB of unified memory), and makes nemoclaw <name> channels remove whatsapp succeed without changes when WhatsApp is not configured.

Release notes → Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/local-gpu/