What changed

Anthropic’s Claude Code changelog lists Claude Haiku 5.5 in version 2.1.293, dated October 7. The model ID is claude-haiku-5-5, and it is now the default Haiku model on the Anthropic API. Haiku is Anthropic’s small, fast model tier, meant for cheap high-volume jobs.

Haiku 5.5 can read up to one million tokens at once. Tokens are the small chunks of text models process and bill for. Anthropic’s standard API rates are $0.10 per million input tokens and $0.50 per million output tokens. Input is the text you send; output is the text the model generates.

For prompts over 100,000 tokens, those rates rise to $0.50 for input and $2.50 for output. Sending a large document can therefore increase the rate for both the prompt and the response.

The model also shipped elsewhere the same day. GitHub says it is generally available in Copilot for Pro, Pro+, Max, Business, and Enterprise, with a gradual rollout. GitHub describes it as built for subagents, quick edits, and terminal tasks. A subagent is a helper that the main agent hands a smaller job to.

Vercel also added it to AI Gateway as anthropic/claude-haiku-5.5. Vercel says it is the first Haiku with effort levels from low to max, which control how much the model thinks and how many tokens it spends.

Plan controls and thinking effort affect usage

GitHub says early testing showed Haiku 5.5 matching Claude Sonnet 5 on many coding tasks while using fewer tokens and steps. Results on your own tasks may differ, so compare quality and cost before moving more work to it.

Copilot bills Haiku 5.5 usage at the provider’s list price. Check your plan’s usage allowance and spending settings before running repeated jobs. Business and Enterprise administrators can turn it off through a model policy. New models are on by default unless an admin has changed that setting.

Vercel notes that at the two highest effort levels, thinking cannot be turned off. It also says Haiku 5.5 ships with biology and cybersecurity safeguards, so it may refuse some requests in those areas. Your app needs to handle those refusals without repeatedly retrying the same request.

What builders should do next

Pick one small, repeatable job you already run on a larger model, such as summarizing diffs or classifying support messages. Use a representative set of examples without private customer data or credentials. Run the same inputs through your current model and Haiku 5.5. Compare four things:

  • Correctness: how many outputs you would accept without editing.
  • Corrections: how many you had to fix by hand.
  • Cost: the charge or credits used across those runs, including retries and correction prompts.
  • Failure recovery: whether refusals or bad outputs break the job or fail cleanly.

Switch only if correctness holds and the cost drops enough to matter. Keep the larger model for work where one wrong answer is expensive. In Claude Code, you can check your installed version with claude --version. In Copilot, look for the model in the picker. If it is missing, check rollout status, your plan, and any organization model policy.

For an app using the API, check Anthropic’s Haiku 5.5 migration guide before changing the model ID. Request parameters and thinking behavior can differ from Haiku 4.5.

End of article