Claude Haiku 4.5 in Late 2026: When the Old Small Model Still Fits

The usual reason to keep an older small model is price. With Claude Haiku 4.5 that reason has gone: its successor is cheaper per token, has five times the context window and twice the output limit. So the interesting question about Haiku 4.5 in October 2026 is not whether it is good, but what would still justify running it.

The answer is narrower than most teams assume, and it is almost entirely about API behaviour rather than quality.

What Haiku 4.5 was built for

Anthropic released Claude Haiku 4.5 on 15 October 2025. The pitch was coding performance similar to Claude Sonnet 4 at one third of the cost and more than twice the speed, and Anthropic said it beat Sonnet 4 on some tasks such as computer use. The intended pattern was a larger model planning and Haiku 4.5 running sub-tasks in parallel underneath it.

Anthropic's published figures were 73.3% on SWE-bench Verified, averaged over 50 trials with a 128K thinking budget, and 41.75% on Terminal-Bench with a 32K budget (40.21% without thinking). Those are strong numbers for a small model, but note the thinking budgets: a 128K budget is not how most people run a cheap, fast model in production.

The specification is a 200K-token context window and 64K maximum output tokens, with text and image input. Thinking is the older manual mode, set with budget_tokens; the effort parameter is not supported. Anthropic lists its reliable knowledge cutoff as February 2025. It was released under ASL-2, a lower safety level than the larger models of that generation.

The price inversion

Haiku 4.5 costs $1 per million input tokens and $5 per million output, with batch at half that. The catalogue's :batch entry is the same model on that discounted asynchronous endpoint.

Claude Haiku 5.5 is priced at $0.10 input and $0.50 output for prompts up to 100,000 tokens, and $0.50 and $2.50 above that. It has a 1M-token context window and 128K output.

The headline is a ten-times difference, but adjust it before quoting it. Haiku 5.5 uses the newer tokenizer introduced with Claude Opus 4.7, and Anthropic says the same text produces roughly 30% more tokens on it than on Haiku 4.5. Large images can cost about two and a half times as many visual tokens, because Haiku 5.5 uses a higher-resolution image tier. Even after those adjustments the newer model is several times cheaper for text-heavy work. Tokenizer differences across models explains why per-token prices alone mislead.

What breaks when you move to Haiku 5.5

Anthropic's migration guide lists the changes. The ones that will actually bite:

• Thinking budgets return a 400. budget_tokens must become adaptive thinking, with effort as the control. Adaptive thinking is on by default, so responses can start with a thinking block even when you did not ask for one.
• Reading the first content block breaks. If your code treats content[0] as the answer, select blocks by type instead.
• Sampling parameters are rejected. Any non-default temperature or top_p, and any top_k, return a 400.
• Assistant prefill is rejected, even with thinking off. Classifiers that relied on prefilling an opening brace need structured outputs or a tool with enum fields.
• Small max_tokens values truncate. Thinking tokens count against the limit, so a 64-token cap tuned for a label classifier can stop before any text appears.
• Refusals. Haiku 5.5 runs safety classifiers that can return a refusal stop reason, with no server-side fallback.

Where Haiku 4.5 still earns its place

Given that list, there are a handful of honest reasons to keep it.

• Deterministic-leaning pipelines. If you rely on temperature 0 for repeatable extraction and have tests built around that, Haiku 4.5 still accepts it.
• Prefill-based formatting. Legacy prompts that steer output by prefilling the assistant turn keep working without a rewrite.
• Priority Tier commitments. Anthropic notes Priority Tier is not supported on Haiku 5.5, so a contracted capacity commitment on 4.5 needs separate planning.
• Frozen evaluations. A classification route that is validated, cheap in absolute terms and rarely touched may not be worth an engineering week to migrate.

Notice that none of these is about Haiku 4.5 being better at anything. If your workload is new, start on Haiku 5.5. When a cheap model is enough covers how to tell whether the small tier is right in the first place.

The retirement clock

Anthropic's deprecation page lists Haiku 4.5 as active, with retirement not sooner than 15 October 2026. That date has now passed, which means it can be deprecated at any point. Anthropic commits to at least 60 days' notice before retiring a publicly released model, so you will not lose it overnight, but you should not start anything new on it.

Bedrock and Google Cloud set their own lifecycle dates, so if you call it through a partner cloud, check that provider's schedule rather than Anthropic's.

A migration test plan

1. Count tokens for a sample of real prompts against Haiku 5.5 using the token-counting endpoint, and recompute the cost with the new counts and prices, including the over-100K tier.
2. Strip sampling parameters and prefill, switch to structured outputs where you were coercing format, and set effort to low for routes where Haiku 4.5 ran without thinking.
3. Raise max_tokens on short-output routes so thinking does not consume the whole budget.
4. Run both models over a few hundred labelled examples and compare accuracy and latency. For routing between them during the transition, model routing and fallbacks covers the pattern.
5. Add refusal handling before cutting over.

For most teams this is a one or two day job that ends with a lower bill. The exceptions are the frozen, prefill-dependent pipelines above, and even those should have a date.