Claude Haiku 5.5: pricing, benchmarks and OpenCode setup

Claude Haiku 5.5 is Anthropic's fastest and cheapest current model, released on October 7, 2026. It costs $0.10 per million input tokens and $0.50 per million output tokens, but only while a request's prompt stays at or under 100,000 tokens. It has a 1M token context window and up to 128K output tokens.

Quick answer for OpenCode users: use it as a small_model or for subagents with short prompts, and keep a larger model for long sessions. The sections below show the price tiers, the published benchmarks and three ways to connect it.

Model IDs and specifications

Source: Anthropic's Claude Haiku 5.5 model page (opens in a new tab). Haiku 5.5 uses the newer tokenizer introduced with Claude Opus 4.7, so the same text counts as roughly 30% more tokens than on Haiku 4.5. Compare token counts, not only per-token prices.

Claude Haiku 5.5 API pricing

Prices are USD per million tokens on the Claude API. The tier is chosen per request, by the total prompt length: uncached input, cache reads and cache writes all count toward the 100,000-token threshold.

A request over the threshold pays the higher prices for all of its tokens, including the part that is a cache hit. Earlier requests in the same session keep the price they were billed at. Requesting US-only inference with inference_geo multiplies every category by 1.1. See Anthropic's pricing page (opens in a new tab) for the full list.

How it compares with other Claude models

Anthropic estimates Haiku 5.5 costs about 75% less to run than Haiku 4.5 on average. The per-token price is 90% lower below the threshold and 50% lower above it; the extra tokens from the new tokenizer account for the gap.

The 100K tier in practice

Each example is a single request at Claude API prices.

Moving from 99K to 101K prompt tokens makes the request about five times more expensive. A cache hit lowers the input cost but does not move the request back below the threshold.

Coding agents build long prompts: the system prompt, tool definitions, file contents and the conversation history are all resent on each call. A long OpenCode session can pass 100K tokens well before the 1M context window fills, and automatic compaction only runs when the context is full. Check the token counts in your provider's usage records before assuming the lower tier.

Claude Haiku 5.5 benchmarks

The scores below are from Anthropic's launch announcement (opens in a new tab). They are vendor-reported, and the announcement does not state the effort level or harness for most Haiku 5.5 results; the system card (opens in a new tab) has the methodology. This site has not run these benchmarks.

How to read them for coding work:

• The gap to Sonnet 5.5 on Terminal-Bench (39.2% vs 70.6%) is the most relevant number for long terminal-driven tasks. Haiku 5.5 is not a drop-in replacement for a larger model there.
• Against Haiku 4.5 the improvement is large on every listed task, so it is worth re-testing workloads that previously failed on Haiku.
• A benchmark score does not include your repository, rules or tools. Run the small evaluation with the same task on each candidate before changing your default model.

Use Claude Haiku 5.5 in OpenCode

OpenCode reads its model catalog from models.dev, which lists Haiku 5.5 under Anthropic, Zen, GitHub Copilot, Amazon Bedrock and Google Vertex AI. If /models does not show it, update OpenCode first. Pick one route:

Connect the provider with /connect, then select the model with /models. To make the choice persistent, the most useful setup keeps a larger model as the default and gives Haiku 5.5 the lightweight work:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "anthropic/claude-sonnet-5-5",
  "small_model": "anthropic/claude-haiku-5-5"
}

small_model handles lightweight tasks such as title generation (see model options). You can also assign Haiku 5.5 to a specific agent, for example a read-only explore or review subagent. Keep the provider prefix the same as the provider you connected; the Copilot ID uses a dot (claude-haiku-5.5), while Anthropic and Zen use hyphens.

Haiku 5.5 rejects some temperature, top_p and top_k values with a 400 error. If an agent configuration sets them for this model, remove them. See Anthropic's migration guide (opens in a new tab) for details.

Through GitHub Copilot

GitHub made Claude Haiku 5.5 generally available in Copilot on October 7, 2026 for the Pro, Pro+, Max, Business and Enterprise plans. It is billed at provider list price under usage-based billing (GitHub changelog (opens in a new tab)). Rollout is gradual. On Business and Enterprise, new models are enabled automatically unless an admin has turned off the default or disabled this model in the Copilot model policy.

In OpenCode, run /connect, choose GitHub Copilot and complete the device login described in the provider guide. If github-copilot/claude-haiku-5.5 is missing from /models, check your plan, your organization's model policy and whether the rollout has reached your account.

Troubleshooting

• Model not found: the prefix does not match the connected provider, the ID uses the wrong separator, or OpenCode's catalog is out of date.
• Bill higher than expected: check the prompt token count per request. Requests above 100K tokens are billed at the higher tier.
• 400 error after switching from another model: remove sampling parameters from the agent or provider options.
• 401, 403 or 429: see the model connection checklist.

Related guides

• Claude models in OpenCode: every current Claude model ID and price.
• Claude Haiku 5.5 on Amazon Bedrock: inference profiles, regional prices and IAM.
• Third-party API setup: connect a gateway through an OpenAI-compatible endpoint.
• Free AI coding agents: what is free in the client and what is billed by the model provider.