Claude Haiku 5.5 vs Haiku 4.5: price, the new 1M window, and code review
Claude Haiku 5.5 vs Haiku 4.5: price, the new 1M window, and code review
Haiku 5.5 does the Haiku job — fast, scoped, high volume — at a much lower bill, with a context window five times larger.
On October 7, Anthropic released Claude Haiku 5.5 and priced it far below Haiku 4.5. The list price for a short prompt drops by 90%. The average bill, after a tokenizer that spends more tokens on the same text, falls by about 75%. The context window moves from 200,000 tokens to 1 million. And Haiku gets an effort dial for the first time.
That combination is what changes code review. A cheap model that can hold the diff and the files around it is a model you can run on the whole queue. Haiku 5.5 is available in ThinkReview on the Lite tier, next to Haiku 4.5.
What the tokens cost
Anthropic's comparison table, per million tokens. Haiku 5.5 has two rates. The cheaper one applies to prompts up to 100,000 tokens, which Anthropic says covered about 90% of requests to Haiku 4.5. The higher one applies above that line.
| Price per 1 million tokens | Haiku 5.5, up to / over 100k | Haiku 4.5 |
|---|---|---|
| Input tokens | $0.10 / $0.50 | $1.00 |
| Output tokens | $0.50 / $2.50 | $5.00 |
| Cache reads | $0.01 / $0.05 | $0.10 |
| Cache writes | $0.125 / $0.625 | $1.25 |
Source: Introducing Claude Haiku 5.5. Cache-write rates in that table match the 5-minute cache write on the model overview. The 1-hour cache write is $0.20 / $1.00 on the same split.
The row that compounds is cache reads: $0.01 instead of $0.10 on a short prompt, paid again each time a review reuses the same repository prefix. Above 100,000 tokens the cut is 50%, so a long review is half the Haiku 4.5 price.
Anthropic's average is about 75% less to run, and the footnote matters. Haiku 5.5 uses the newer tokenizer shared with Sonnet 5.5 and Opus 5.5, so the same text counts as about 30% more tokens than on Haiku 4.5. A 90% list-price cut and a 30% token increase land together near that 75% figure. Recount a typical diff in the new tokenizer. A prompt that sat safely under 100,000 tokens on Haiku 4.5 can step into the higher band.
The new 1M context window
Haiku 4.5's context window is 200,000 tokens. Haiku 5.5's is 1 million. Max output goes from 64,000 tokens to 128,000.
A 200,000-token Haiku fills up once the diff is large and the surrounding files come with it. The review drops context, or it splits into passes that cannot see each other. A 1 million token window holds the diff, the files it touches, and the repository-level context ThinkReview already attaches, in one pass. Max output of 128,000 tokens lets a wide review list its findings without stopping at 64,000.
The 100,000-token price step sits inside that window. A review that uses a large share of the million tokens pays $0.50 / $2.50, half of Haiku 4.5, on a prompt the old model could not accept. Most reviews stay under the step. The new tokenizer's extra 30% is why a diff that used to sit just under 100,000 tokens is worth recounting.
Thinking tokens count toward the output limit. A max_tokens value sized for a short Haiku 4.5 reply can stop after the thinking block, before the comment. Leave room for the think.
Adaptive thinking, and what it changes in a review
Haiku 5.5 is the first Haiku with an adjustable effort setting. Adaptive thinking is on by default: the model decides when to think and how much, and effort steers the depth. The API default is medium. Lower effort thinks less, and on a simple request the model can skip thinking. Higher effort spends more tokens and more time.
Haiku 4.5 used a manual thinking budget (budget_tokens). That field returns an error on Haiku 5.5. Effort replaces it. A review queue is mixed, so the dial is the part that changes daily work:
- A one-line fix needs a fast pass: is the change what it claims, and is there an obvious break. Low effort keeps that pass cheap and short.
- A multi-file bugfix needs the model to follow a call, check a null, and write a comment a human will trust. Medium, the default, is the setting Anthropic ships for that.
- A dense change — a migration, a permission check, a parser — can take a higher effort and still stay on Haiku's token price. The alternative on Haiku 4.5 was a fixed budget, or a jump to Sonnet.
Reviewing every pull request gets cheaper twice: the token price is lower, and a trivial hunk can finish without a long reasoning trace. Use effort to get more from a scoped review. When the review is an open-ended design judgment, move to Sonnet 5.5.
Where the scores moved
These rows are from Anthropic's launch table.
| Evaluation | Haiku 5.5 | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| GDPval-AA v2.1 | 1620 | 735 | 1840 |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 70.6% |
| FrontierCode 1.1 (Main) | 46.4% | — | 52.1% (Xhigh) |
| Humanity's Last Exam, no tools | 45.9% | 10.2% | 56.9% |
Source: Introducing Claude Haiku 5.5. No Haiku 4.5 FrontierCode score was published. Sonnet's FrontierCode figure is at Xhigh effort.
FrontierCode, which scores whether a code change would be merged, puts Haiku 5.5 at 46.4%, near Sonnet 5.5 at Xhigh effort and ahead of GPT-6 Luna at 42.4%. GDPval-AA more than doubles. Terminal-Bench is the long multi-step job, and Sonnet 5.5 still leads it.
Box measured 11 points above Haiku 4.5 at about half the latency, including on weekly recurring reviews. Asana saw over a 30% cut in task latency and up to 2.5× faster inference per agent turn.
Which Haiku to review with
Use Haiku 5.5 on the reviews you want for every pull request: a defined change, a checkable result, a comment in the same sitting. The 1M window is why a large diff can stay on this model. The price is why the small diffs can too. Adaptive thinking is why a mixed queue can share one model.
Keep Haiku 4.5 for a pinned comparison on old evals. For new reviews, set the model id to claude-haiku-5-5, drop the old thinking budget, and recount tokens before you trust a Haiku 4.5 cost estimate.
In ThinkReview, both models are on Lite (and plans that include Lite models). Open Settings or Model selection, choose Claude Haiku 5.5, and run the review. The catalog is in Model Selection.
When the diff needs a judgment held across an open design, switch to Sonnet 5.5. When it needs the surrounding repository and a fast, cheap pass, Haiku 5.5 is the model this release was priced for.
Compare them on a real diff. Install ThinkReview or manage your models in the portal.
Prices, context limits, and benchmark figures reference Anthropic's Claude Haiku 5.5 announcement, October 7, 2026, the what's new page, and the model overview.