🧭 Claude Code v2.1.316 — Subagent Effort Controls, Pooled Group Budgets, and 34 Bug Fixes

Claude Code v2.1.316 shipped on October 31, and it is a notably operator-heavy release. The biggest new capability is CLAUDE_CODE_SUBAGENT_EFFORT, an environment variable that lets you pin the effort level agents spawn at — useful when you want a fast, cheap sweep pass before a deeper review, or when you are running many parallel subagents on a tight budget. Alongside that, the Claude Apps Gateway now supports managed.policy_mode: layered, which lets administrators stack multiple policy files rather than choosing one: a top-level org policy is applied first, then a team-level policy is merged on top, giving enterprises a way to enforce baseline guardrails while allowing per-team customisation without forking configurations.

Highlights for developers and operators

• CLAUDE_CODE_DISABLE_ULTRATHINK — add this to your environment to globally suppress the ultrathink keyword. Helpful in production pipelines where you want deterministic, fixed-effort invocations and cannot control what users type.
• Pooled group budgets in the Apps Gateway — one monthly token budget shared across a group, with individual usage still tracked. The pool refills at month-end. Groups with fewer than 5 members are not eligible (de-anonymisation guard).
• CLAUDE_CODE_DISABLE_LOOPED_CROSS_CERT_DROP — a workaround for enterprise networks whose TLS inspection proxies send looped system cross-certificates that previously caused handshake failures.
• Sign-in to Apps Gateway from VS Code chat panel — no longer requires a terminal session; the IDE integration now handles the OAuth flow inline.
• Cloud sessions: form card support for MCP server forms — MCP servers can now surface structured input forms in cloud sessions, replacing the previous free-text prompt workaround.
• Code Review analytics page — a new findings-by-severity card on the analytics dashboard lets teams track review-flagged issues over time without exporting data.

Opus 4.8 Fast Mode: 30-day warning

This release also formally marks the start of the 30-day deprecation window for fast mode on Opus 4.8. After November 30, requests to claude-opus-4-8 with speed: "fast" will return a hard error — there is no silent fallback to standard Opus 4.8. If your pipeline uses Opus 4.8 fast mode, migrate to claude-opus-5-5 (which has its own fast mode) before month-end. The model itself (claude-opus-4-8 without fast mode) is unaffected and remains active until at least May 28, 2027.

🧭 Sonnet 4.5 Retires in 30 Days — Everything That Breaks and the Four Steps to Fix It

As of today, claude-sonnet-4-5-20250929 (and its short alias claude-sonnet-4-5) has exactly 30 days before it is retired from the Claude API, Claude Platform on AWS, and Microsoft Foundry. After November 30, requests to those model IDs will return a hard error — there is no grace-period fallback. Anthropic deprecated this model on September 30, satisfying its 60-day minimum notice commitment. The recommended migration target is claude-sonnet-5-5.

What breaks if you do nothing

• Any code passing model="claude-sonnet-4-5" or model="claude-sonnet-4-5-20250929" to the Anthropic SDK, Bedrock API, or Foundry will receive a 404/model-not-found error.
• Automated pipelines, CI/CD jobs, and scheduled cron tasks that hard-code the model string will silently start failing at midnight UTC on December 1.
• Third-party integrations (LangChain, LlamaIndex, custom wrappers) that pin to the old model ID will break even if your own code is updated — check upstream dependencies.

Four-step migration

1. Audit: Go to Claude Console → Usage → Export. The CSV breaks down calls by API key and model — find every place claude-sonnet-4-5 appears.
2. Swap: Replace claude-sonnet-4-5-20250929 with claude-sonnet-5-5 everywhere. If you use an alias (the short form without a date), update that too.
3. Test: Run your evals on Sonnet 5.5. It is significantly more capable than Sonnet 4.5 — but watch for output-length changes and differences in instruction-following strictness. Your system prompts may need minor tuning.
4. Verify partner platforms: Amazon Bedrock and Google Cloud (Google-managed) run their own retirement schedules. If you use those platforms, check their model lifecycle pages separately — the November 30 date applies only to Anthropic-operated endpoints.

🧭 DeepSeek Harness Launches as Open-Source Claude Code Rival — V4-Pro Also Hits the API

DeepSeek today released two interrelated products aimed squarely at the professional developer segment where Anthropic's Claude Code is strongest. The first is DeepSeek Harness (DSH), an open-source autonomous coding agent framework designed to match Claude Code's task-delegation model — complete with a plugin/extension ecosystem, a CLAUDE.md-equivalent project context file, and a hooks system for pre/post tool-call automation. The second is DeepSeek V4-Pro reaching general availability on DeepSeek's API, a 1.6-trillion parameter Mixture-of-Experts model with a 1M-token context window, priced at $0.66/M input tokens (off-peak), which undercuts comparable Claude tiers significantly on list price alone.

What makes DeepSeek Harness different from earlier DeepSeek agents

• Proactive workflow monitoring: Unlike conventional coding assistants that respond to individual prompts, DSH is designed to monitor development group chats and file systems continuously, proposing solutions without waiting for direct invocation.
• Open-source plugin directory: Developers can write and share DSH plugins in TypeScript or Python — a direct mirror of the Claude Code Mods architecture that Anthropic introduced in early October.
• MCP-compatible tool layer: DSH ships with an MCP-compatible tool interface, meaning tools already written for Claude Code can be wrapped for use in DSH with minimal porting effort.
• $15 billion in new funding: DeepSeek's Harness team was established earlier in 2026, supported by a ~100 billion yuan ($15B) funding round intended to fund sustained competition with US frontier labs.