Claude Haiku 5.5 was released on October 7, 2026, with a documented 1M-token context window, adaptive thinking and input pricing starting at $0.10 per million tokens. GitHub says Anthropic’s newest lightweight model is generally available in Copilot, while AWS has announced availability on Amazon Bedrock and Claude Platform on AWS. The release targets developers handling frequent, latency-sensitive tasks rather than establishing a new benchmark result. GitHub announcement, AWS announcement, model overview.
The practical questions are how its pricing changes with prompt length, which request settings need attention and what the advertised speed and savings actually establish.
A large context window with conditional output limits
The model overview describes Haiku 5.5 as built for classification, routing, extraction and subagent tasks. GitHub also identifies quick edits and terminal tasks as intended uses. Those descriptions establish the workload focus; they do not provide independent measurements of accuracy or response time.
The overview lists these specifications:
| Specification | Documented value |
|---|---|
| Context window | 1M tokens |
| Maximum output | 128k tokens |
| Message Batches API maximum output, beta | 300k tokens |
| Thinking | Adaptive; on by default |
| Knowledge cutoff | June 2026 |
The 300k-output limit is conditional: the overview specifies the Message Batches API and the output-300k-2026-03-24 beta header. It should therefore be distinguished from the ordinary 128k maximum. Model overview.
For integration planning, context capacity and output capacity answer different questions. The documented 1M-token window does not mean every request can produce a response of that length, and the extended batch output should not be assumed to apply to other request paths.
Pricing changes above 100,000 prompt tokens
The overview gives two input and output pricing tiers, with both determined by prompt length:
| Prompt length | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Up to 100,000 tokens | $0.10 | $0.50 |
| Over 100,000 tokens | $0.50 | $2.50 |
The distinction matters for the headline starting price. A model can accept a 1M-token context window while charging higher rates for prompts above 100,000 tokens. The output tier is also tied to prompt length, rather than to the output length shown in the table. Model overview.
AWS relays Anthropic’s claim that Haiku 5.5 costs around 75% less than Haiku 4.5 for most tasks. It also attributes to Anthropic the description of Haiku 5.5 as the fastest, most efficient model in the Claude 5.5 family. The supplied announcement excerpt provides no workload methodology or measured latency results for those claims, so they remain vendor assertions rather than independently established comparisons. AWS announcement.
There is a further migration consideration: the overview says Haiku 5.5 uses the newer tokenizer shared with Claude 4.7 and later models. The same text counts as approximately 30% more tokens than on Haiku 4.5. That makes prior token counts an unreliable basis for estimating the new model’s bill or proximity to the pricing threshold. Developers evaluating the savings claim should account for both the applicable rate and the changed token count.
Request settings need attention
Adaptive thinking is enabled by default, and the overview identifies effort as the control for thinking depth. It does not provide an executable request example or measurements showing the effect of different effort settings, so no particular latency or quality trade-off can be quantified from this evidence.
The overview instructs developers to omit temperature, top_p and top_k: a non-default value for any of them returns a 400 error. Existing request configurations therefore warrant review before changing the model identifier.
Thinking blocks have an account boundary as well. According to the overview, they work only in the account that produced them or in a linked account. Workflows that move those blocks between accounts need to respect that restriction. Model overview.
Availability and the remaining evaluation gap
Alongside Copilot and the AWS offerings, the model overview lists the Claude API, Google Cloud and Microsoft Foundry. It also states that retirement will occur no sooner than October 7, 2027; that is an earliest retirement commitment, not a fixed retirement date. Model overview.
The release provides concrete specifications and migration constraints for evaluating Haiku 5.5. What the supplied evidence does not establish is how quickly or accurately it performs on a developer’s own workload. The useful next step is to assess those results alongside prompt-length pricing and the tokenizer change before treating the advertised savings as an application-level outcome.
