Mistral Large 4 opens its weights and leads open coding benchmarks
Mistral Large 4, Claude Code v2.1.316 and Google Fitbit Edge, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
Mistral
Mistral Large 4
DEPLOYif you run coding agents or security tooling
WHAT SHIPPEDMistral released the weights of Mistral Large 4, a 1-trillion-parameter multimodal model, after three weeks of additional reinforcement learning. It leads open-weight coding on FrontierCode 2.0 with 19.7% and Terminal-Bench 4.0 with 41%, and scores 51% on the Artificial Analysis Cyber Index.
THE CATCHThe model uses 47% fewer tokens per task than GLM-5.3 on the Artificial Analysis Coding Agent Index while scoring 48%, so the cost advantage shows up at scale, not in a single benchmark number.
IF YOU RUNa security team that needs to audit model architecture and deploy on your own infrastructure
SOURCEMistral, Mistral Large 4 (Le Chonk) Built for Coders, Trusted by Defenders
Anthropic
Claude Code v2.1.316
WATCHif you run Claude Code behind a gateway
WHAT SHIPPEDClaude Code v2.1.316 added pooled group budgets to the Claude apps gateway, CLAUDE_CODE_SUBAGENT_EFFORT, and guardrail.input_tagging for Bedrock upstreams. It deprecated fast mode for Opus 4.8 as of October 30, 2026, with removal on November 30, 2026.
THE CATCHFast mode requests on Opus 4.8 will return an error after November 30, 2026 and will not fall back to standard Opus 4.8, so any pinned fast mode usage breaks rather than slowing down.
IF YOU RUNOpus 4.8 in fast mode and have not migrated to Opus 5.5
SOURCEAnthropic, Claude Code v2.1.316 release notes
Google Fitbit Edge
WATCHif you track workouts and want larger display
WHAT SHIPPEDGoogle Fitbit Edge is on shelves now with the largest, brightest tracker display and a new screen design, physical side buttons, and a wrist-contoured shape with accessories.
THE CATCHThe announcement names no battery life, sensor set, or price, so the hardware decision rests on the display and buttons alone.
IF YOU RUNa fitness routine where at-a-glance readability during workouts matters
SOURCEGoogle, Google Fitbit Edge: On shelves now
The Deploy Log
Models and APIs
| DEPLOYif your agency requires GovCloud inference endpoints ChatGPT on Bedrock GovCloud. The ChatGPT desktop app is now available on Amazon Bedrock in AWS GovCloud (US), routing Codex model inference through Amazon Bedrock Mantle in us-gov-west-1 with IAM governance and zero data retention support. Amazon Web Services |
Tools and products
| DEPLOYif you run LFM3 locally on Linux Ollama v0.40.13. Ollama v0.40.13 brings Liquid AI LFM3 models, including text-to-speech in lfm3-audio and lfm3-omni, to Linux and Windows in addition to Apple silicon. Ollama |
| DEPLOYif you deploy new Vercel projects Node.js 26 LTS. Node.js 26 LTS is generally available for builds and Vercel Functions, with HTTP/2 by default for built-in fetch via Undici 8, the Temporal API, and V8 14.6, and it is the default for all new projects. Vercel |
| DEPLOYif you already call mistral/mistral-large-4 on AI Gateway Mistral Large 4 Gateway. Mistral Large 4 on Vercel AI Gateway now uses updated weights with lower pricing for input, cached input, and output tokens, and the existing model ID string mistral/mistral-large-4 serves the new version without changes. Vercel |
| DEPLOYif you maintain a static webhook allowlist GitHub webhook IP ranges. GitHub added three webhook delivery IP ranges to the Meta API hooks field: 9.234.98.160/28, 9.234.98.176/28, and 9.234.106.48/28, for github.com across all plans including GitHub Enterprise Cloud. GitHub |
| DEPLOYif you use Codex TUI daily Codex rust-v0.166.0. Codex rust-v0.166.0 added keyboard-driven transcript selection with word, line, and page motions, integrated search, GPT-6.1 Terra in bundled, Vertex AI, and Amazon Bedrock model catalogs, and codex queue --resume. OpenAI |
| WATCHif you want to trial a 1M-token model Ling 3.1 Flash 1019. Ling 3.1 Flash 1019 from InclusionAI is available on Vercel AI Gateway free until November 3, 2026, with a 1M-token context window for coding, multi-step reasoning, and tool-using agents. Vercel |
| WATCHif you use geolocation for access controls Vercel IP geolocation. Vercel IP geolocation now resolves city-level data for nearly all requests and postal-code data for 90% of traffic, with more frequent refreshes, enabled for all projects with no code or configuration changes. Vercel |
| DEPLOYif you use Vercel Flags in functions Vercel Flags SDK. Vercel Flags SDK initialization on cold starts dropped from over 70 ms to 10 ms at p75, roughly 7x faster, and usage reporting no longer keeps functions running after the response, with update @flags-sdk/vercel to 1.5.3 or later. Vercel |
| WATCHif you build on langchain-core message merging langchain-core 1.7.2. langchain-core 1.7.2 added add_notification_metadata and get_notification_metadata and fixed merge_content so it no longer changes the first message's content. LangChain |
Left out
- Google Workspace Gemini file editing: duplicate of the Google Gemini file editing item with the same availability facts
- vLLM v0.33.0rc2: release note carries only a ROCm GEMM bugfix with no user-facing availability or deployment fact
- langchain-core 1.7.3: release note contains only a version bump with no feature or fix detail
- Claude Code v2.1.314: the publisher's page could not be read when this edition was built
- Vercel AI Bots rules: the publisher's page does not show it available yet
- Sheets cell citations: it did not carry a number that moved
- AgentCore Session endpoints: it did not carry a number that moved
- Cloudflare private origins: it did not carry a number that moved
- Gemini Slides editing: it did not carry a number that moved
- Gemini file editing: it did not carry a number that moved
The pattern
Mistral Large 4 opens its weights with a 1-trillion-parameter multimodal model that leads open-weight coding on FrontierCode 2.0 at 19.7% and Terminal-Bench 4.0 at 41%, and Vercel immediately serves the updated weights at lower token pricing through the same mistral/mistral-large-4 model ID. The same week, Claude Code v2.1.316 deprecates fast mode for Opus 4.8 with a hard error after November 30, 2026, and Codex rust-v0.166.0 adds GPT-6.1 Terra to bundled, Vertex AI, and Amazon Bedrock catalogs. The pattern is not that one model wins; it is that model access is now a routing and migration problem. A security team can audit Mistral Large 4 on its own infrastructure, but a team pinned to Opus 4.8 fast mode breaks rather than slows down, and a team on Vercel AI Gateway gets Ling 3.1 Flash 1019 free until November 3, 2026 with a 1M-token context window. The cost and capability signal sits in token efficiency, Mistral Large 4 uses 47% fewer tokens per task than GLM-5.3 on the Artificial Analysis Coding Agent Index while scoring 48%, and in cold start latency, Vercel Flags SDK initialization drops from over 70 ms to 10 ms at p75.
The infrastructure layer is tightening around these model changes. Node.js 26 LTS becomes the default for all new Vercel projects with HTTP/2 by default via Undici 8 and V8 14.6, while GitHub adds three webhook delivery IP ranges to the Meta API hooks field for github.com across all plans including GitHub Enterprise Cloud. ChatGPT on Bedrock GovCloud routes Codex model inference through Amazon Bedrock Mantle in us-gov-west-1 with IAM governance and zero data retention support, and Ollama v0.40.13 brings Liquid AI LFM3 models including text-to-speech in lfm3-audio and lfm3-omni to Linux and Windows in addition to Apple silicon. Google Fitbit Edge arrives on shelves with the largest, brightest tracker display and physical side buttons but no battery life, sensor set, or price, so the hardware decision rests on the display and buttons alone. Across all of these, the hard tokens are dates and version numbers: October 30, 2026, November 30, 2026, November 3, 2026, v2.1.316, v0.40.13, rust-v0.166.0, 1.7.2, 1.5.3, 26 LTS, 14.6, 8, 1019, 4.8, 5.5, 6.1, 3.1, 4, 5.3, 2.0, 4.0. The message is that every deployment choice now carries a migration deadline or a version pin, and the only safe position is to read the release notes as infrastructure, not as announcements.