What's new in Claude Code
The latest Claude Code release is v2.1.305, released Oct 20, 2026.
October 2026: Claude Code adds Claude Mods, Haiku 5.5 default, hook failure blocking and idle compaction
So far in October 2026, Claude Code has added Claude Mods, which let plugins change deeper behavior, along with a new default Haiku model, stricter hook controls and idle compaction. Twelve releases have shipped so far, from v2.1.287 to v2.1.300, and the month is still in progress.
v2.1.287 introduced Claude Mods and an opt-in "You Should Know" side agent, enabled with /plugin enable cc-plugin-you-should-know@builtin. Later releases kept extending the mod hooks, including agent.spawn, prompt.change and $.ui.view. v2.1.293 made Claude Haiku 5.5 (claude-haiku-5-5) the default Haiku model on the Anthropic API. v2.1.295 added onFailure: "block" for command and HTTP hooks, and v2.1.299 added idle compaction for conversations of 200K+ tokens.
Other changes include --max-findings for /code-review, --marketplace for claude plugin install, and an effort parameter on the Agent tool. v2.1.291 fixed regressions from v2.1.288 and v2.1.290 that could lose final messages or drop permission answers in cloud sessions. Several sandbox and permission fixes also landed in v2.1.292.
- Claude Mods let plugins modify deeper behavior; opt-in "You Should Know" side agent via /plugin enable (v2.1.287)
- Claude Haiku 5.5 (claude-haiku-5-5) is the default Haiku model on the Anthropic API, with 1M context (v2.1.293)
- onFailure: "block" makes command and HTTP hooks block the action if they fail to start, time out or exit unexpectedly (v2.1.295)
- Idle compaction compacts a 200K+ token conversation after about 54 minutes idle; "idleCompaction": false turns it off (v2.1.299)
- -max-findings <n>|all for /code-review, and Up on an empty prompt restores a draft cleared with Ctrl+C (v2.1.288)
- claude plugin install --marketplace <source> and an effort parameter on the Agent tool for sub-agents (v2.1.292)
September 2026: Claude Code adds new default models, a /diff panel and AGENTS.md support
Claude Code's September 2026 releases changed the default models, added a side-by-side /diff panel and made AGENTS.md readable as project instructions. Twenty-seven releases shipped, from v2.1.257 to v2.1.286.
Model defaults moved three times: Claude Fable 5.1 (v2.1.257), Claude Opus 5.5 (v2.1.280) and Claude Sonnet 5.5 (v2.1.284), each with 1M context. Day to day, fullscreen mode gained a /diff panel for uncommitted changes, /cost and the status line now show a likely cause for prompt-cache misses, and ctrl+enter sends all queued messages at once. In projects with no CLAUDE.md, AGENTS.md is read instead. Plugin work included claude plugin eval and claude plugin configure.
Administrators got managed settings such as managedMcpServers, deniedModels, availableModelsMatch and allowedProviders, plus many Claude apps gateway options. Several releases fixed regressions affecting proxy and gateway setups, macOS 12 launch and session resume.
- Defaults changed: Claude Fable 5.1 (v2.1.257), Claude Opus 5.5 (v2.1.280) and Claude Sonnet 5.5 (v2.1.284), each with 1M context
- /diff opens a panel beside the conversation in fullscreen mode showing uncommitted changes (v2.1.260)
- AGENTS.md is read as project instructions when a project has no CLAUDE.md (v2.1.277)
- Send-now key (ctrl+enter or ctrl+x ctrl+s) interrupts the turn and sends all queued messages (v2.1.275)
- claude plugin eval runs a plugin's eval suite and produces scored JSON and HTML reports (v2.1.269)
- Managed settings deniedModels, availableModelsMatch and allowedProviders restrict which models and providers can be used (v2.1.283, v2.1.285)
Latest Claude Code releases
- v2.1.305 (Oct 20, 2026) — Claude Code 2.1.305 is out: tool search now retries cleanly on Vertex, Bedrock and gateways, worktrees start from the right branch, and screen reader mode finally reads search boxes correctly. Some ch
- v2.1.304 (Oct 19, 2026) — Claude Code 2.1.304 is here: a guided /auto-mode-setup that reads your project, a security check for Managed Agents, and a heads-up before resuming a conversation eats your usage limit. Plus a pile of
- v2.1.303 (Oct 18, 2026) — Claude Code 2.1.303 is here with bug fixes and reliability improvements to keep your workflow running smoothly. Curious what got sharper?
- v2.1.302 (Oct 16, 2026) — Claude Code 2.1.302 is here: pick your own auto mode classifier model, toggle idle compaction right from /config, and get a heads-up on standalone Claude Design's December 14, 2026 closure. Plus a bat
- v2.1.301 (Oct 15, 2026) — Claude Code 2.1.301 is here: smarter unattended retries, vim % bracket jumps, multi-line input, and richer telemetry on auto mode decisions. Curious what changed under the hood?
- v2.1.300 (Oct 14, 2026) — Claude Code 2.1.300 brings smarter prompt handling with virtual text that won't interfere with your submissions, plus new ways to customize agent views and control advisor tools directly. You can now
- v2.1.299 (Oct 14, 2026) — Claude Code 2.1.299 brings smarter memory management with idle compaction, so long conversations stay efficient without reprocessing, and gives you finer control over how Claude handles file reference
- v2.1.296 (Oct 9, 2026) — Claude Code 2.1.296 now syncs code settings across Claude Desktop and the gateway for more consistent behavior, while new controls let you manage window compaction and model selection for subagents wi
- v2.1.295 (Oct 8, 2026) — Hooks can now block actions when they fail, terminals can show what Claude Code is doing, and gateway upstreams get a time-to-first-byte limit. Ready to see what else changed?
- v2.1.294 (Oct 8, 2026) — Claude Code 2.1.294 strengthens your safety controls by fixing how instruction-based hooks are interpreted, so blocking rules actually work as intended. We've also improved when Claude decides to stop
Upgrading from v2.1.294? See every change since.