Jonathon's AI Wiki

What's New in Claude Code — Week 40 (September 28–October 4, 2026, in progress)

7 min read

Source: raw/anthropic-watch-claude-code-tag-v2-1-284.md (Claude Code v2.1.284, published 2026-09-28) and raw/anthropic-watch-skills-commit-8a1541c4a3ffa5a20a5a91de0dcf3f0bab1d1ef4.md (anthropics/skills, 2026-09-29 02:20 UTC), with the same-day SDK releases raw/anthropic-watch-anthropic-sdk-python-tag-v1-9-0.md and raw/anthropic-watch-anthropic-sdk-typescript-tag-sdk-v0-129-0.md.

Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.284 · Date: 2026-09-28

This week is in progress. It covers one release so far, v2.1.284, which adds Claude Sonnet 5.5 as the default Sonnet model on the Anthropic API and makes auto mode the default permission mode for interactive sessions on every plan and provider; full model coverage is in Claude Sonnet 5.5. The bundled claude-api skill moved to Opus 5.5 as its default and Sonnet 5.5 as its current Sonnet a few hours later. Later W40 releases will be added here as they are ingested. Continues Week 39.

Key Takeaways

  • Claude Sonnet 5.5 (claude-sonnet-5-5) is the default Sonnet model on the Anthropic API — 1M context, 10 per Mtok, $0.20/Mtok cache reads (v2.1.284, 2026-09-28). The release note scopes the default to the Anthropic API; Bedrock, Vertex and Foundry defaults are not stated. See Claude Sonnet 5.5.
  • Auto mode is now the starting permission mode everywhere. Interactive terminal and VS Code sessions start in auto mode when no permission mode is configured, on every plan and provider; permissions.defaultMode still overrides it (v2.1.284). v2.1.283 had done this only for third-party providers and telemetry-off installs. See Auto Mode.
  • Ultracode is its own toggle in /effort (Tab, or /effort ultracode [on|off]): it no longer forces xhigh effort and stays on at any effort level (v2.1.284). New keybinding actions effortSlider:decreaseEffort, increaseEffort and toggleUltracode. See Dynamic Workflows.
  • Safety-related model switches in sessions pinning an Opus model via ANTHROPIC_DEFAULT_OPUS_MODEL or modelOverrides: on the Anthropic API, the API now picks the model to switch to for each kind of flag, not the pinned model (v2.1.284).
  • {"decision":"block"} from Elicitation and ElicitationResult hooks now declines the MCP elicitation, as exit code 2 does; it was being ignored (v2.1.284). See Hooks.
  • The claude-api skill makes Opus 5.5 the default model, adds Sonnet 5.5 as the current Sonnet, adds a “Migrating to Claude Opus 5.5” section, and ships build-eval and hillclimb flows (skills 8a1541c4).
  • SDKs added claude-sonnet-5-5 the same day — Python v1.9.0, TypeScript sdk-v0.129.0, and the Bedrock, Vertex and Foundry TypeScript packages. See Anthropic SDK Releases Aug–Sep 2026.

Release table

VersionDate (UTC)Shape
v2.1.2842026-09-28Claude Sonnet 5.5; auto mode default for all; Ultracode toggle; /mcp reconnect all

Models and defaults

  • Claude Sonnet 5.5 added, default Sonnet on the Anthropic API (v2.1.284).
  • Auto mode default on every plan and provider when no permission mode is set (v2.1.284).
  • Ultracode decoupled from effort level (v2.1.284).
  • Safety-related model switch target chosen by the API per kind of flag when an Opus pin is set (v2.1.284).
  • The notice shown when a Sonnet model’s safeguards flag a message now explains why and offers editing and retrying (v2.1.284).
  • The Explore subagent no longer switches to Opus on the Claude API when the session runs an unrecognized model ID, such as a custom model behind a proxy; Explore inherits that model (v2.1.284). See Subagents.

New commands and features

  • /mcp reconnect all retries every MCP server that failed to connect or needs authentication (v2.1.284). See MCP.
  • /rate-limit-options appears in /help and the command menu for claude.ai subscribers (v2.1.284).
  • Auto mode’s prompt before a read outside the working directories gains “Yes, but ask again next time” (v2.1.284).
  • Gateway spend limits show dollar amounts in /usage and the status line (e.g. “500.00 spent this month”); rate_limits.spend_limit gains used_usd, limit_usd and period (v2.1.284). See Claude Apps Gateway.
  • Gateway: startup warnings when a managed policy’s availableModels is empty or omits the starting model; auth: { google: {} } for telemetry export to Google Cloud’s OTLP endpoint; certificate client authentication (private_key_jwt) with the identity provider (v2.1.284).
  • /claude-api hillclimb no longer spends rounds on prompt rewordings too small for the eval to measure (v2.1.284).
  • Monitor event rows show what each event printed instead of repeating the description (v2.1.284). See Monitor Tool.
  • Claude Tag: model family choices such as “Opus (latest)” for a thread, a channel default or a DM, so the choice follows the newest model in that family (v2.1.284). See Claude Tag.

Hooks, MCP, plugins and skills

  • Elicitation / ElicitationResult {"decision":"block"} honoured (v2.1.284).
  • A failed hook’s stderr is no longer dropped from the debug log when it also wrote stdout, and failed hooks log their status code (v2.1.284).
  • MCP tool calls in a resumed session wait up to 10 seconds for a still-connecting server instead of failing with “No such tool available” (v2.1.284).
  • claude mcp add refuses when managed settings restrict MCP servers to plugins (v2.1.284).
  • Plugins from marketplaces, claude.ai and npm no longer pre-approve their own tools via allowed-tools under managed allowManagedPermissionRulesOnly; only official Anthropic sources, or sources managed settings vouch for, keep that pre-approval (v2.1.284). See Plugins and Marketplaces.
  • The non-interactive first turn still waits up to 2 seconds for MCP servers named by --allowedTools or an mcp_tool hook, even when CLAUDE_CODE_MCP_STARTUP_WAIT_MS is 0 (v2.1.284).
  • /loop self-paced mode writes each status update, and the outcome when it stops, as visible text instead of only in reasoning (v2.1.284). See Scheduled Tasks.
  • Rules symlinked into .claude/rules from outside the project now trigger the external-imports approval prompt (v2.1.284).
  • Auto-memory: invisible characters and tags imitating Claude Code’s own markup are neutralized in MEMORY.md and recalled notes (v2.1.284). See Auto Memory.

The claude-api skill update (skills 8a1541c4, 2026-09-29)

  • Claude Opus 5.5 (claude-opus-5-5) becomes the default model across SKILL.md, models.md and every language’s examples, with Opus 5 as the previous Opus; a “Migrating to Claude Opus 5.5” section is added to shared/model-migration.md.
  • Claude Sonnet 5.5 (claude-sonnet-5-5) is the current Sonnet, with Sonnet 5 as the previous one; pricing, model-support callouts and platform availability updated.
  • Build-eval and hillclimb flows under shared/evals/ (build-eval, eval-hillclimb, cost-hillclimb, eval-audit), with a report schema, a lite HTML report builder and a runner scaffold.
  • A preserved-thinking-migration subcommand with probe scripts, and a note that drop_block can raise token usage; cost-optimize updated for preserved thinking and per-model cache rates, pointing cache misses at cache diagnostics.
  • Streaming client tools default to eager_input_streaming; model access failures documented as a 404 and beta gating as a 400, not a 403.
  • ant apply recommended for version-controlled Managed Agents resources. See Managed Agents.

Breaking and behaviour changes

  • /recap declines with a short notice when relayed from a chat thread, routine or webhook (v2.1.284).
  • Retries after a dropped connection mid-response share one budget with the request’s other retries, so a failing request gives up sooner (v2.1.284).
  • Artifact publishing refuses a file on a network share unless it is on a mapped drive added with --add-dir (v2.1.284).
  • VS Code: CLAUDE_CONFIG_DIR in claudeCode.environmentVariables applies only when it is an absolute path (v2.1.284).

Notable fixes

  • “Prompt is too long” persisting after compaction: Claude Code now compacts once more, keeping less of the recent conversation (v2.1.284).
  • An overloaded or server error right after a thinking block is retried instead of ending the turn (v2.1.284).
  • ANTHROPIC_FOUNDRY_RESOURCE was interpolated into the Foundry endpoint host unvalidated; non-plain values are now refused (v2.1.284).
  • Claude Desktop behind a Claude apps gateway offered no 1M context option; the gateway now marks each 1M-capable model (v2.1.284).

Try It

  • If you don’t want to start in auto mode, set permissions.defaultMode in user or managed settings.
  • Pinned Opus via ANTHROPIC_DEFAULT_OPUS_MODEL? A safety-related switch may now land on a model you didn’t pin; check the notice when it happens.
  • Using Ultracode at lower effort: turn it on with /effort ultracode on and set effort separately.
  • Update the bundled skill’s examples in your own code to claude-opus-5-5 / claude-sonnet-5-5 if you copied model IDs from the claude-api skill.

Open Questions

  • Is Sonnet 5.5 the default Sonnet on Bedrock, Vertex and Foundry? v2.1.284 says “on the Anthropic API” only. The TypeScript Bedrock, Vertex and Foundry SDKs added the model ID the same day.
  • What else ships in W40? This digest covers v2.1.284 only; later releases this week are not yet ingested.
  • What does the Opus 5.5 migration section say? The skills commit names the section but not its contents.

Graph View

Backlinks