Claude Haiku 5.5: price, 1M context and how it compares with Haiku 4.5

On October 7, 2026, Anthropic released Claude Haiku 5.5, the new small, fast model in its Claude lineup. Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, a tenth of Haiku 4.5's price. It also has a 1-million-token context window and 128,000-token maximum output. It follows Opus 5.5 (September 22) and Sonnet 5.5 (September 28).

This story is based on Anthropic's launch page, API release notes, pricing and model documentation, the Haiku 5.5 system card and the Claude Code changelog, checked October 12, 2026.

The short version

• Model ID: claude-haiku-5-5, for classification, extraction, routing and subagents.
• Specs: 1M-token context, 128K max output, adaptive thinking with effort (default medium), knowledge cutoff June 2026.
• Price: $0.10 / $0.50 per million tokens up to 100K-token prompts; $0.50 / $2.50 above that.
• Where: Claude API, Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry. Default Haiku in Claude Code 2.1.293 on the Anthropic API.
• Benchmarks: large gains over Haiku 4.5, all vendor-reported.

What is Claude Haiku 5.5?

According to the October 7 release notes, Haiku 5.5 is "the most capable model tuned for high-volume and latency-sensitive work." Its launch page calls it Anthropic's fastest model at standard speed. It's the first Haiku with an adjustable effort setting (low to max); Claude decides how much to think, and effort steers it. Our reasoning models explainer covers why that matters for cost, since thinking tokens are billed as output.

How much does Claude Haiku 5.5 cost?

Prices are from Anthropic's pricing page, in US dollars per million tokens.

Last verified: October 12, 2026

Three details change the real bill:

1. The 100K threshold covers the whole request. Prompt length "counts all of its input tokens, including cache reads and cache writes," and a request over 100,000 tokens "pays the higher prices even when part of its prompt is a cache hit," per the pricing page. Long-document jobs cost five times the headline rate.
2. The tokenizer is newer. Haiku 5.5 uses an updated tokenizer like Sonnet 5.5's and Opus 5.5's and uses "slightly more tokens per task," Anthropic says. Its pricing page says the newer tokenizer produces "approximately 30% more tokens for the same text" than the one Haiku 4.5 uses. Recount prompts with the token-counting API.
3. "75% cheaper" is Anthropic's average. The launch page estimates about 90% savings on prompts up to 100K tokens and about 50% above that, weighted by its own traffic (about 90% of Haiku 4.5 requests were under 100K). It already accounts for the tokenizer.

For how Haiku 5.5 fits against Sonnet, Opus and Fable, see which Claude model you should use. For why list prices mislead, see token pricing isn't the story.

Haiku 5.5 vs Haiku 4.5: what changed

Last verified: October 12, 2026 (context windows from Anthropic's context-window docs; status from the deprecations page)

It isn't a drop-in swap. Per the release notes, budget_tokens, temperature, top_p, top_k and assistant-message prefill can each return a 400 error. The older computer_20250124 tool is rejected. Structured outputs aren't available for Haiku 5.5 on Bedrock. Adaptive thinking is on by default, so responses can start with thinking blocks.

Benchmarks (vendor-reported)

All numbers below are vendor-reported: Anthropic published them on its launch page on October 7, 2026, and says the system card describes the methods. The launch page doesn't give the effort level or harness for each score, and the system card says it states the effort setting for some evaluations but not all. The page doesn't say who produced the GPT-6 Luna figures.

OSWorld here is an "offline subset," so don't compare it with full-suite scores elsewhere. Anthropic still points complex agentic coding to Sonnet 5.5 and Opus 5.5. Our benchmarks explainer covers what Terminal-Bench measures and its limits.

What the system card says

Anthropic published a roughly 140-page system card on October 7. It says Haiku 5.5 "does not cross any new RSP thresholds" under Anthropic's Responsible Scaling Policy. Its cyber capabilities "significantly outperformed Claude Haiku 4.5" but fell short of Opus 5.5. Anthropic calls it its "most robust Haiku-class model yet to prompt injections."

The card is also candid about weak spots. Haiku 5.5 "over-refused more than any other model we tested in our automated behavioral audit." It "hallucinated more than other recent models, and about as much as Claude Haiku 4.5." On some mental-health tests it "assisted more often than Haiku 4.5, with the clearest regression when thinking was disabled," and Anthropic recommends developer safeguards. Our guide to reading a system card explains how to weigh findings like these.

Where is Haiku 5.5 available?

• API and clouds: the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Claude on Google Cloud and Claude in Microsoft Foundry, per the release notes. On Bedrock and Google Cloud, the provider sets the price, and regional endpoints cost 10% more than global ones.
• Claude Code: the changelog entry for 2.1.293, published to npm on October 7, reads: "Added Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API." Claude Code's model docs show the haiku alias still resolving to Haiku 4.5 on Bedrock, Google Cloud, Claude Platform on AWS and Microsoft Foundry. Pin the full model ID if you want 5.5 there.
• Claude apps: the Help Center announced Haiku 5.5 on October 7 but doesn't say which plans get it.

Anthropic suggests Haiku 5.5 as a subagent for Opus 5.5 or Sonnet 5.5; our model routing explainer covers when that saves money.

What's confirmed, reported and uncertain

• Confirmed (Anthropic docs): release date, model ID, 1M context, 128K max output, prices and the 100K threshold, breaking changes, platform availability, Claude Code 2.1.293 default, Haiku 4.5 status.
• Vendor-reported: all benchmark scores, the "75% cheaper" average, speed claims and customer results.
• Reported: Reuters, on October 7, described it as the third Claude 5.5 model in a month, launched "ahead of a planned IPO."
• Uncertain: which Claude app plans include Haiku 5.5, when Claude Code's cloud-provider aliases will move to it, and when Haiku 4.5 will actually be retired.