/v1/complete), so today's update is worth reading before your next deploy. On the open-weights front, Mistral dropped a 1.05T-parameter MoE model ("Le Chonk") into public API preview on October 6, with open weights promised by end of month — keeping competitive pressure on closed-weights providers.completions.create() before next deploycurl -LsSf https://github.com/t8y2/dbx/releases/latest/download/dbx-installer.sh | sh
dbx connect postgres://localhost:5432/mydb
dbx mcp start --port 8765
anthropic.completions.create() / /v1/complete) is now explicitly marked deprecated in the TypeScript SDK. The endpoint still functions; no removal date has been announced. The Python SDK (v1.12.0, same day) does not include this deprecation notice — it may have been deprecated there in an earlier release.completions.create() deprecated — if your TypeScript code calls the legacy /v1/complete endpoint, plan the migration to messages.create() now.completions.create or client.completions — any match is code on a deprecation track. The Messages API (messages.create) has been the recommended path since 2023, so the migration surface should be small for modern codebases. If you call the raw HTTP endpoint at /v1/complete from a non-SDK HTTP client, the same applies. No removal date is announced yet; treat this as a "plan it this sprint, execute within a month" item. On the Python side: the Python SDK v1.12.0 released today doesn't include this notice — check whether your Python code uses the completions client as a separate action.anthropic.completions.create() or makes direct HTTP calls to /v1/complete.EffortModerate — requires finding all completions.create call sites and migrating to messages.create with updated parameter shapes.claude-haiku-4-5 to claude-haiku-5-5 and run your test suite. Key details before swapping in production: (1) Pricing tiers: ≤100K tokens: $0.10 input / $0.50 output; >100K tokens: $0.50 input / $2.50 output; Sonnet 5.5 cache-read price was also cut to $0.10/MTok today. (2) Effort levels: claude-haiku-5-5 is the first Haiku with adjustable effort (low/med/high/xhigh/max); med is default — combine with Claude Code v2.1.292's effort param on the Agent tool for per-sub-agent reasoning cost control. (3) Tokenizer change: Haiku 5.5 uses a slightly different tokenizer than Haiku 4.5 — expect marginally higher token counts per task; recalibrate cost models against real workloads before scaling. (4) Computer/browser use: now supported in beta (via SDK v1.12.0 / sdk-v0.132.0). (5) Context: 1M tokens; 128K max output (300K via output-300k-2026-03-24 beta header on Batches API). (6) Retirement commitment: not before October 7, 2027.| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 (Elo) | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1 | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1 (offline) | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity's Last Exam (no tools) | 45.9% | 10.2% | — | 56.9% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 (Main) | 46.4% | — | 42.4% | 52.1% |
claude-haiku-5-5 with typed computer and browser toolset tool calls. New model capabilities object fields: web_search and code_execution booleans. RBAC roles: display_name field added, name deprecated. /v1/models endpoint gets lifecycle_stage field and filter. url_sources in Managed Agents web_fetch tool config. include_default parameter for listing workspaces. line field on model objects. Agent tools limit newly documented as 256. TypeScript-only: Text Completions API marked deprecated (see [BREAKING] above); JSONL stream double-iteration now throws. Python bug fix: local edit retained after memory upload with lost response. Companion cloud SDK releases also published (Vertex, Bedrock, Foundry, Google Cloud, AWS — patch versions, no additional feature notes).anthropic==1.12.0 / @anthropic-ai/sdk@0.132.0 to use claude-haiku-5-5, gain model capability introspection fields for web_search and code_execution, and get the RBAC display_name field; TypeScript users must also check for completions.create() usage (deprecated) and JSONL stream double-iteration patterns.claude-haiku-5-5 model ID is now valid — use it. Computer/browser toolset typed tool calls are available if you build computer-use agents. (2) RBAC roles: switch from .name to .display_name on role objects in any code that reads role names; .name is deprecated and will be removed in a future version. (3) Model capabilities introspection: the capabilities object from /v1/models now includes web_search and code_execution booleans, useful for dynamic capability checking before constructing tool-use requests. (4) TypeScript only: completions.create() is deprecated (see [BREAKING] above); if you iterate a JSONL stream twice (e.g., for retry logic), it now throws rather than silently returning empty — surface bugs that were previously silent.claude-haiku-5-5 is now the default Haiku model on the Anthropic API. Critical bug fixes: (1) context compaction was causing Claude to retract already-completed work — fixed; (2) HTTP MCP connections leaked memory by keeping every request object in memory until connection close — fixed. New mod API: agentType field added to subagentStatusLine payload; isDeferred: false on $.tool.register lists a tool's schema in the prompt immediately rather than behind a ToolSearch step. Cloud session change: the v2.1.290 fix for sessions staying asleep after container restart was reverted (it caused regressions). Slack: channel rules raised from 20 to 50 per workspace; announcement channels joined without posting an intro message. Additional bug fixes: messages lost on background session swap with ←; /model effort wrapping past highest/lowest level; /tui disconnected in Chrome; SendMessage offered in sessions where it had been removed; claude purge silently stopping on file deletion errors; path-scoped rules and nested CLAUDE.md files not loading when files accessed via cat/grep in Bash; Windows: stopping a hook could kill an unrelated process with the same PID.claude-haiku-5-5 — update immediately.isDeferred: false on $.tool.register makes a tool schema immediately visible in the prompt without a ToolSearch step — use for tools that should always be available. agentType in subagentStatusLine lets status-line scripts distinguish custom subagent types from the main session. (4) Nested CLAUDE.md / path-scoped rules: the regression where these weren't loaded when files were accessed via cat/grep in Bash is now fixed — if you use nested CLAUDE.md files for domain-specific context, this was silently not applying.npm install -g @anthropic-ai/claude-code or equivalent.arXiv.org, huggingface.co, simonwillison.net, and paperswithcode.com remain egress-blocked from this network environment. Three candidate papers (arXiv:2610.07261 NP-Bench, arXiv:~2610.xxxxx KaliBench, arXiv:2608.26623 AgentJudgeBench) were located via secondary abstract mirrors but not read from their primary sources — excluded per non-negotiables. See Near-misses.
/mcp login <name> command for authenticated MCP server sign-in with stored credentials. Voice microphone and speaker device selection. Per-turn Cyber access program selection via codex exec or TypeScript SDK. Daybreak feature flag moved to opt-in. Bug fixes: approved filesystem escalation now correctly grants broader write while keeping read restrictions; launch permissions persist across reconnects; SQLite corruption detected earlier with backups preserved; retries honor server retry guidance./mcp login for authenticated MCP server sessions, and opens Bedrock GovCloud regions — update if you run Codex in Bedrock or against private MCP servers./mcp login <server-name> now handles sign-in and credential storage directly in Codex rather than requiring external setup. (3) GovCloud: if you operate in AWS GovCloud, Bedrock Mantle now supports those regions — previously required a workaround or wasn't supported at all. Note: rust-v0.162.0-alpha.18 was released the same day; stick to v0.161.0 for stable use.Generate Commit Message now follows .clinerules. Responses with no recognized finish reason get one continuation request instead of failing. Large MCP tool output previewed with a link to the remainder. Fixed 400 errors for Claude via custom Anthropic base URLs. Fixed rejected requests for reasoning-level-limited models (e.g., Kimi K3). Desktop: Mermaid diagrams now render inline with copy/download/fullscreen/pan+zoom controls. Cline Hub auto-retries for up to 60 seconds on connection drop. Local/self-hosted providers with no API key can start sessions again (previously failed with "Missing API key"). Windows CLI + Desktop: npx/uvx MCP server startup timeout raised from 3 to 10 seconds.npx/uvx-based MCP servers fail to start on Windows, the 3-second startup timeout was likely the cause — update to desktop-v0.0.44 or cli-v3.0.69 for the 10-second timeout. (2) Vultr users: the catalog refresh may have changed pinned model IDs for Vultr — check your model selection in settings after updating, and reselect if responses come from an unexpected model. (3) Custom Anthropic base URL users: the 400-error fix for Claude via custom base URLs is in v4.1.23 — update if you proxy Anthropic API calls through a custom endpoint.curl -LsSf https://github.com/t8y2/dbx/releases/latest/download/dbx-installer.sh | sh. Start the MCP server with dbx mcp start --port 8765, then add the endpoint to your Claude Code or Cline MCP config. Fastest value check: connect a database you already use locally and have your coding agent run a query without writing any MCP tool handler code. Verify connector maturity in the README before production use — some connectors (SQL Server full-schema introspection, Huawei GaussDB) are marked alpha.github.com/cursor/plugins/blob/main/docs/getting-started.md. Install the SDK: npm install @cursor/plugins-sdk. Fastest path to value: port an existing Claude Code mod to the Cursor plugin format and compare the ergonomics — the hook model is similar enough that the delta is instructive.lmarena.ai, swebench.com, huggingface.co/spaces/open-llm-leaderboard, and livecodebench.github.io remain egress-blocked this scan cycle. No confirmed leaderboard changes for October 7.
Standing context: The Haiku 5.5 launch page (fetched from anthropic.com) includes fresh benchmark numbers from Anthropic — see Model Releases for the full table (GDPval-AA v2.1 1620, AA-Briefcase 1578, OSWorld 2.1 72.4%, Terminal-Bench 4.0 39.2%, FrontierCode 1.1 46.4%). Third-party verification pending access to lmarena.ai.
/mcp login for authenticated MCP server sessions; Claude Code patches MCP tool path-scoped rules loading; dbx ships a first-class built-in MCP server for database access; Cline fixes MCP server startup on Windows; and Groq's MCP Connectors Beta (GroqCloud, recently live but undated) adds native MCP connectivity to the inference API. These aren't coordinated — they're independent teams reaching the same conclusion simultaneously.Nothing cleared the quality bar this period. arXiv, simonwillison.net, and the Hacker News API all remained egress-blocked. No qualifying threads found via web search.
CrewAI 1.15.24 (Oct 7) — crewai eval markdown brief output; Oracle integrations added; experimental job lifecycle and runner behind an opt-in flag; crewai eval --models can now swap models as well as roles; crewai eval exits with code 1 unless gate passed. github.com/crewAIInc/crewAI/releases/tag/1.15.24
LangGraph CLI 0.4.33 (Oct 7) — --image-uri option for deploying a pre-pushed image; langgraph deploy listeners list command; now rejects Git dependencies containing embedded credentials (security fix); dependency bumps: virtualenv 21.7.13, urllib3 2.8.0, pyjwt 2.15.0. github.com/langchain-ai/langgraph/releases
Sonnet 5.5 cache-read price cut (Oct 7, Anthropic) — Cache-read pricing for Claude Sonnet 5.5 cut from $0.20/MTok to $0.10/MTok, announced alongside the Haiku 5.5 launch. Update cost models if you run prompt-cached workloads on Sonnet 5.5. anthropic.com/pricing
GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash, and Grok 4.5 removed from all Copilot experiences. Audit any Copilot API integrations that pin specific model versions before October 19. [github.blog/changelog/2026-09-18-upcoming-deprecation-of-selected-github-copilot-models-in-mid-october/]
Mistral committed to open-weight release for Large 4; one secondary source cited October 27, but the official post was blocked this scan. Watch mistral.ai/news for the release.
The TypeScript SDK now marks /v1/complete as deprecated. No removal date announced yet. Watch the SDK changelog for the removal commit. [github.com/anthropics/anthropic-sdk-typescript]
Cheaper-tier 300 tok/s fast mode remains unshipped. GPT-6 Astra Ultrafast is live; Sol Ultrafast is not. Watch platform.openai.com/docs/changelog.
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.