Claude Haiku 5.5 Is Out. Claude Code Needs v2.1.293 to Use It.

The common thread this Thursday is that new things arrive faster than the channels that carry them. Anthropic shipped Claude Haiku 5.5 on Wednesday, but the Claude Code build that understands it, and the overnight build that fixes a hook problem, both sit ahead of stable. VS Code and Copilot CLI meanwhile changed where policies and settings are read from, which is easy to miss when nothing looks broken. Treat today as the day to check which Claude Code build each machine runs before switching to Haiku 5.5, reread any prompt hook that is meant to block something, and confirm your sandbox policy and Copilot settings still land where you think.

Top Stories

1. Claude Haiku 5.5 is out, and Claude Code needs v2.1.293 to use it

Anthropic introduced Claude Haiku 5.5 on Wednesday 7 Oct as its fastest and cheapest small model, aimed at high-volume work such as summaries, compaction, classification and subagents. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, rising to $0.50 and $2.50 above that, and Anthropic says it runs around 75% cheaper than Haiku 4.5 on average. It has a 1M token context window and is the first Haiku with an adjustable effort setting, which defaults to medium. Anthropic also halved cache-read pricing on Sonnet 5.5.

Claude Code v2.1.293, published to npm at 1:18 PM ET Wednesday, added claude-haiku-5-5 and made it the default Haiku on the Anthropic API, and the model configuration docs say to use v2.1.293 or later with it. GitHub made Haiku 5.5 generally available in Copilot at 4:12 PM ET Wednesday for Pro, Pro+, Max, Business and Enterprise users. The rollout is gradual across VS Code, Copilot CLI, the cloud agent and other clients, at provider list pricing, so some of you will not see it yet.

Practical dev impact: On v2.1.293 or later the haiku alias resolves to Haiku 5.5 on the Anthropic API, and prompt hooks default to the background model that the same ANTHROPIC_DEFAULT_HAIKU_MODEL variable controls, so pin that variable if you want hook judging to stay put through the upgrade. Machines on npm stable (v2.1.285 at 6:51 AM ET today) do not have the model. Copilot Business and Enterprise admins should check the model policy now, because new models are enabled automatically unless the global default is off.

2. Claude Code’s hook fix was on npm next, not latest or stable, at 6:51 AM ET

Anthropic published v2.1.294 to npm at 11:42 PM ET Wednesday, and its first line fixes prompt and agent hooks written as instructions, such as “Block commands that…”, allowing what they should block. It also changes how Stop and SubagentStop prompt hooks written as instructions, such as “Carry on if the build is broken”, are judged, so Claude is less likely to stop early. At 6:51 AM ET today npm next pointed to v2.1.294 while latest was v2.1.293 and stable v2.1.285, so neither main npm channel has the fix. Anthropic’s native installer is ahead, with its latest pointer on v2.1.294 and stable on v2.1.286.

Practical dev impact: A prompt or agent hook is a model call that answers ok true or false, and one phrased as an order may have been approving the very commands it names. Our advice, not Anthropic’s, is to rewrite those hooks as explicit evaluations that say when to answer ok false, the way the docs’ own examples are phrased, until v2.1.294 reaches your channel. Keep any block that truly matters in a command hook, which the docs recommend for production over the experimental agent hooks.

3. VS Code v1.141 makes the old sandbox policy overridable in the Copilot harness

VS Code v1.141 went out as stable at 7:01 AM ET Wednesday. The release notes say the Copilot harness, which runs in a separate agent host process, may already be your default and is gradually becoming the default for enterprise users with VS Code experiments enabled. Inside that agent host, the legacy ChatAgentSandboxEnabled device policy now supplies an overridable default rather than a mandatory requirement, and the legacy sandbox settings and policies are deprecated. Local behavior is unchanged. To require sandboxing and block bypass, Microsoft points admins to unified managed settings with sandbox.enabled true and sandbox.allowBypass false, a control it labels Preview. Sandboxing now covers Windows, macOS and Linux, and the release adds Chat: Open Worktree Cleanup for reclaiming disk from inactive agent worktrees.

Practical dev impact: An enterprise that forced sandboxing through the old device policy should assume it is now a default that users on the Copilot harness can switch off. Add the Preview managed settings before relying on it again, and keep Microsoft’s own caveat in mind that sandboxing is a layer of protection, not a standalone security boundary.

4. Copilot CLI v1.0.93 stops reading user settings from config.json

GitHub published Copilot CLI v1.0.93 at 10:06 AM ET Wednesday, and npm latest now points to it. The notes say user settings are read only from ~/.copilot/settings.json and user-setting keys left in ~/.copilot/config.json are ignored. GitHub’s config directory docs now say the same thing, and describe config.json as a home for internal state such as authentication and installed plugins. The release also adds an enterprise permissions.limitTo setting that keeps agent network requests inside managed domains, and MCP configuration changes now apply between turns. On the pre-release channel, v1.0.94-0 lets /model discover models from a running local Ollama, which GitHub says does not turn on offline mode or disable telemetry.

Practical dev impact: A dotfiles repo or provisioning script that still writes preferences into config.json stops taking effect on v1.0.93, so move those keys to settings.json. Try local model discovery only on a pre-release test machine.

OpenAI’s Codex v0.161.0, now npm latest, adds /mcp login with a server name for signing in to MCP servers mid-session, while Gemini CLI shipped only a nightly in this window.

Practical Impact Analysis

Most of today’s work is about where a change actually lands, because almost nothing here breaks loudly. Haiku 5.5 is cheap enough that teams may point subagents and compaction at it right away, but on Claude Code that only works on v2.1.293 or later. A mixed fleet will run the new model on some machines and the old one on others, and the model judging prompt hooks can change in the same upgrade.

The hook fix is the more urgent read, because a hook that approves what it should block fails quietly. Rephrasing those prompts takes minutes and helps on every channel today, while waiting for v2.1.294 protects only machines that update. VS Code and Copilot CLI teach the same lesson, since in both the old place still exists but no longer carries its old weight. An admin who checks only that sandboxing was configured, or a developer who checks only that a preference was saved, can be wrong without seeing an error.

Tutorial

Check the channels first, since these tags move without notice, then list the prompt and agent hooks in your Claude Code settings and compare the keys in Copilot’s old and new files. The install lines are commented out so you can run the checks safely and uncomment one on purpose, on the channel it names.

#!/usr/bin/env bash
# 1. Read the channel pointers and local versions.
npm view @anthropic-ai/claude-code dist-tags --json
npm view @github/copilot dist-tags --json
claude --version   # Haiku 5.5 needs v2.1.293 or later; the hook fix is v2.1.294
copilot --version

# 2. List prompt and agent hooks so you can reread how each is phrased.
for f in ~/.claude/settings.json .claude/settings.json .claude/settings.local.json; do
  [ -f "$f" ] && echo "$f" && jq -c '.. | objects | select(.type == "prompt" or .type == "agent") | {type, prompt}' "$f"
done

# 3. Compare top-level keys in Copilot's old and new files.
[ -f ~/.copilot/config.json ] && jq -c 'keys' ~/.copilot/config.json
[ -f ~/.copilot/settings.json ] && jq -c 'keys' ~/.copilot/settings.json

... click "Show full code" below to expand

#!/usr/bin/env bash
# 1. Read the channel pointers and local versions.
npm view @anthropic-ai/claude-code dist-tags --json
npm view @github/copilot dist-tags --json
claude --version   # Haiku 5.5 needs v2.1.293 or later; the hook fix is v2.1.294
copilot --version

# 2. List prompt and agent hooks so you can reread how each is phrased.
for f in ~/.claude/settings.json .claude/settings.json .claude/settings.local.json; do
  [ -f "$f" ] && echo "$f" && jq -c '.. | objects | select(.type == "prompt" or .type == "agent") | {type, prompt}' "$f"
done

# 3. Compare top-level keys in Copilot's old and new files.
[ -f ~/.copilot/config.json ] && jq -c 'keys' ~/.copilot/config.json
[ -f ~/.copilot/settings.json ] && jq -c 'keys' ~/.copilot/settings.json

# 4. Optional canaries. Leave stable fleets (npm stable was v2.1.285 at 6:51 AM ET) alone.
# npm latest channel today, adds Haiku 5.5:
# npm install -g @anthropic-ai/claude-code@2.1.293
# npm next channel today, adds the hook fix; test machine only:
# npm install -g @anthropic-ai/claude-code@2.1.294

At 6:51 AM ET today npm showed Claude Code stable v2.1.285, latest v2.1.293 and next v2.1.294, and Copilot CLI latest v1.0.93 with prerelease v1.0.94-3. Native Claude Code installs follow their own channel pointers, which can sit a build apart from npm, so trust claude --version over the tag and do not mix npm and native installs on one machine. Any config.json key you set yourself, rather than state such as authentication, installed plugins or trusted folders, belongs in settings.json now. If jq rejects settings.json, it likely has comments, which Copilot allows, so open it in an editor instead. We tested the jq filters on a sample settings file, not on a live install.

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.

I use Claude Code and GitHub Copilot CLI and want to review my setup after this week’s changes. Claude Code supports the new Claude Haiku 5.5 from v2.1.293, and v2.1.294 fixed prompt and agent hooks written as instructions, such as “Block commands that…”, allowing what they should block, while the npm stable channel was still on v2.1.285 at 6:51 AM ET on 8 October 2026. Copilot CLI v1.0.93 reads user settings only from ~/.copilot/settings.json and ignores user-setting keys left in ~/.copilot/config.json. I will paste my Claude Code hooks configuration, the output of claude --version, and the top-level keys from both Copilot files below. Rewrite each prompt or agent hook as an explicit evaluation that says exactly when to return ok false, tell me which hooks should become command hooks because the block really matters, and list which Copilot keys look like user settings I should move into settings.json. Do not invent setting names or versions I did not give you, and ask me for anything you need instead of guessing.