Claude Haiku 5.5: Price, Long-Context Tier and Who Upgrades

Claude Haiku 5.5 shipped on October 7, 2026 with a price list that has two columns, and the column you land in depends on how long your prompt is. Up to 100,000 tokens, input costs $0.10 per million tokens and output $0.50. Over 100,000 tokens, input costs $0.50 and output $2.50. Haiku 4.5, the model it replaces, charged $1 and $5 at every length. The API ID is claude-haiku-5-5, and Anthropic says it is live on the Claude API, Amazon Web Services, Google Cloud and Microsoft Azure from day one.

Two things changed at once, and they pull in different directions. The headline rate fell by 90% for the short prompts that make up most small-model traffic. At the same time the model gained a 1M-token context window, five times the 200K that Haiku 4.5 had, and a 128K output limit, double the old 64K. The long-context tier is the price of that extra room. A team that reads the first number and ignores the second will budget wrong.

This post does four things. It lays out both price tiers next to Haiku 4.5 and Sonnet 5.5. It prices two concrete requests, a 50,000-token prompt and a 200,000-token prompt, on each model. It separates what Anthropic measured from what you can expect, and labels every benchmark as the vendor's own. And it ends with a router for who should switch this week. For where Haiku sits in the whole lineup, our frontier model guide is the maintained reference; for the dated ledger of October releases, see the October 2026 release tracker.

Anthropic publishes two full rows for Haiku 5.5 on its pricing page, one for prompts up to 100,000 tokens and one for prompts over it. No other current Claude model is priced this way. Opus 5.5, Sonnet 5.5 and Fable 5.1 charge the same per-token rate across the whole 1M window. Haiku 5.5 is the exception, and the reason is plain enough: it is the first Haiku with a 1M window, and Anthropic has chosen to charge for using the far end of it.

The table puts the two Haiku 5.5 tiers next to Haiku 4.5. The short-prompt tier is a tenth of the old price on every line. The long-prompt tier is half the old input price and half the old output price, so even a 500,000-token request on Haiku 5.5 costs less per token than any request on Haiku 4.5 did. It just costs five times more than the same tokens would have cost under 100K.

One row is easy to misread. The cache-read price of $0.01 is 10% of the $0.10 input price, which is the standard Claude ratio. It is not the 5% ratio that Opus 5.5 and Sonnet 5.5 now carry, and not the 2.5% on Fable 5.1. Haiku 5.5's cache reads are cheap because its input is cheap, not because the multiplier moved. Section 04 works through what that means for a cached agent loop.

Anthropic's long-context pricing note states three rules, and each one matters for a budget. First, the prompt length that decides the tier is the whole input of the request: every token you send, including the tokens served from cache and the tokens written to it. A 95,000-token cached prefix plus a 6,000-token question is a 101,000-token prompt, and it pays the higher rates even though most of it was a cache hit. Second, each request is priced on its own. Third, requests that were billed at the low tier stay there; crossing the line later does not re-price them.

The practical consequence is that the line is not about a conversation's length. It is about the size of the single largest request in it. An agent that accumulates context and re-sends it every turn will cross 100K somewhere in the middle of a long task, and from that turn on its input costs five times more, cache hits included. A pipeline that sends one 40,000-token document per call never gets near it. Anthropic says prompts up to 100,000 tokens were about 90% of requests to the previous Haiku, which is its argument that most users will never see the second column.

One more thing moves the line without anyone noticing. Haiku 5.5 uses the newer tokenizer that arrived with Claude Opus 4.7, and Anthropic's model page says the same text counts as roughly 30% more tokens than it did on Haiku 4.5. A prompt that measured 80,000 tokens on Haiku 4.5 is about 104,000 tokens on Haiku 5.5, and it has crossed the threshold before you changed a word of it.

Rates per million are hard to feel. A single request is not. The table prices two requests, each with a 2,000-token answer, on Haiku 5.5, Haiku 4.5 and Sonnet 5.5, using the list rates above and no caching or batch discount. The arithmetic is tokens divided by one million, times the rate, summed for input and output. A third row shows the cliff itself: a prompt one token over 100,000.

Three readings follow. For the short request Haiku 5.5 is a tenth of Haiku 4.5 and a twentieth of Sonnet 5.5, and at a thousand requests the gap is the difference between six dollars and a hundred and twenty. For the long request Haiku 5.5 is still a quarter of Sonnet 5.5, so the long-context tier is not a penalty relative to the alternatives; it is a penalty relative to Haiku 5.5's own short-prompt rate. And the cliff row shows the shape of the risk: one token changes the input bill by a factor of five, with nothing in the response to tell you it happened.

Thinking tokens are the thing this table leaves out, and they are billed as output. Haiku 5.5 has adaptive thinking on by default at the medium effort level, so a request that looks like 2,000 output tokens may carry more. For a classification or extraction task, set effort low and the number stays close to the table. For a task that needs the model to reason, the output column grows and so does the gap to Haiku 4.5, which had no effort parameter at all.

Prompt caching on Haiku 5.5 uses the standard Claude multipliers: a 5-minute cache write at 1.25× input, a 1-hour write at 2× input, and reads at 10% of input. The minimum cacheable prompt is 512 tokens. Batch processing takes a flat 50% off input and output in both tiers. The figures below are the ones to copy into a cost model.

Work a cached loop through. An agent re-sends a 60,000-token cached prefix with a 3,000-token new turn, fifty times. On Haiku 5.5 each turn costs 60,000 tokens of cache read at $0.01 per million, which is $0.0006, plus 3,000 tokens of fresh input at $0.10, which is $0.0003: under a tenth of a cent for input, or about 4.5 cents across the fifty turns before output. On Haiku 4.5 the same loop was 45 cents. The saving is real, but notice what carries it: the base input rate, not the ratio.

The Sonnet 5.5 change is the one line in this launch that is not about Haiku. From October 7, Sonnet 5.5's cache reads cost $0.10 per million tokens instead of $0.20, which is 5% of its $2 input. Anthropic's estimate, stated in the launch post, is that this cuts the cost of Sonnet 5.5 on most agentic tasks by around 20%. The per-token input and output prices did not move, so the saving lands only where cache reads dominate the bill. Our cache-first agent architecture guide covers how to find out whether yours does.

The price is only half the upgrade. Anthropic's model overview lists a 1M-token context window, a 128K-token synchronous output limit, 300K output on the Batch API with a beta header, adaptive thinking with the effort parameter, and a reliable knowledge cutoff of June 2026. Haiku 4.5 had a 200K window, a 64K output limit, no effort parameter, and a reliable cutoff of February 2025, which is now twenty months stale. On context window, output limit and knowledge cutoff Haiku 5.5 now matches Sonnet 5.5 and Opus 5.5.

Two rows are migration work rather than marketing. The sampling row means a client that still passes temperature to Haiku can start receiving 400 errors the moment the model ID changes; Anthropic's migration guide lists the values that fail. The retirement row means Haiku 4.5's floor date is eight days away. Anthropic has not announced a retirement, and a floor is a commitment not to retire before that date rather than a plan to retire on it. But a team still on Haiku 4.5 in November is running a legacy model with no stated runway, and should treat this launch as the deadline it did not have last week.

The thinking row carries a quieter change. Haiku 5.5's thinking blocks work only in the account that produced them, or one linked to it, which matters for anyone who logs full conversations and replays them from a different account. The pattern is the same one we documented for the Sonnet 5.5 migration.

Anthropic's launch post calls Haiku 5.5 "the cheapest, fastest, and most capable small model we've ever released", and backs it with a benchmark table. Every figure in that table is Anthropic's own run, on Anthropic's chosen settings, with no independent replication on launch day. That does not make them wrong. It makes them a vendor's figures, and the honest way to read them is against the vendor's own larger models, which Anthropic helpfully includes.

Terminal-Bench 4.0 is the clearest example. Haiku 5.5 scores 39.2% where Haiku 4.5 scored 0.0% and GPT-6 Luna, the OpenAI model behind ChatGPT's Free and Go tiers, scored 16.4%. Sonnet 5.5 scores 70.6%. Anthropic's own commentary on that chart says Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding, and that Haiku 5.5 is suited to narrowly scoped tasks that were cost-prohibitive before: compaction, summarisation, subagent work. The benchmark table and the positioning agree, which is not always the case in a launch.

The other rows tell the same story at different altitudes. OSWorld 2.1 on the offline subset: 72.4% against Haiku 4.5's 15.7% and Sonnet 5.5's 83.9%. Humanity's Last Exam with no tools: 45.9% against 10.2% and 56.9%. GDPval-AA v2.1, a knowledge-work rating: 1620 against 735 and 1840. In each case the new Haiku lands closer to Sonnet than to its predecessor. Whether that holds on your own tasks is the question only your own evaluation answers, and it is the one a vendor table cannot.

The cost claim deserves the same label. Anthropic says Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average. That is a calculation, not a list-price comparison. Anthropic's footnote blends the 90% cut for prompts up to 100,000 tokens with the 50% cut above it, weights them by Haiku 4.5's request mix (90% short), and adjusts for the new tokenizer using more tokens per task. Your blend will differ. The one customer figure in the post comes from Asana, quoted on Anthropic's page:

A testimonial chosen by the vendor is weaker evidence than a benchmark the vendor ran, and both are weaker than your own trace. It is quoted here because speed is the claim Anthropic leads with and the only one a named customer attests to on Anthropic's launch page.

Haiku 5.5 is available on launch day across every platform Anthropic sells through: the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. The model ID is claude-haiku-5-5 everywhere except Bedrock, which prefixes it as anthropic.claude-haiku-5-5. Anthropic describes no staged rollout and no waitlist.

Three other things shipped in the same announcement, and two of them cost nothing to claim.

The decision has three inputs: what you run on Haiku 4.5 today, how long your prompts are after re-tokenization, and whether you run anything on Sonnet that only needed Sonnet because Haiku 4.5 was weak. The router below is our reading of Anthropic's published numbers. It is not a quality judgement about Haiku 5.5 on your tasks; run your own evaluation before any move that touches customers.

The one move we would not make is the one the headline invites: moving everything onto Haiku 5.5 because it is cheap. Anthropic's own table puts it thirty points behind Sonnet 5.5 on agentic coding, and a small model that fails a task three times costs more than a larger one that passes once. The price list makes the small tasks almost free. It does not make the large ones small. If you want help deciding which of your workloads fall on which side, that is work our AI transformation team does with clients every month.