A Claude API prompt that crosses 100,000 tokens now lands in a different price tier. Claude Haiku 5.5 input costs $0.10 per million tokens up to that boundary and $0.50 above it. Output rises from $0.50 to $2.50 per million tokens on the same split. That cliff matters more to an agent developer than Anthropic's claim that this is its fastest small model: a long conversation, a large retrieval bundle, or an accumulated tool trace can make the rate jump fivefold between two requests.
The release drew immediate developer attention. In MrKeyoor's October 7 snapshot, the Hacker News discussion had reached 669 points and 336 comments. The thread is a demand signal, not independent evidence for Anthropic's performance claims. Its recurring question is still useful: how cheap is Haiku 5.5 once a real workload reaches the 100,000-token line?
Anthropic released Claude Haiku 5.5 on October 7 across its own API, Amazon Web Services, Google Cloud, and Microsoft Azure. The model ID on the Claude API is claude-haiku-5-5. Anthropic says it costs about 75% less per task on average than Haiku 4.5 after accounting for both the new rates and a tokenizer change. For prompts no longer than 100,000 tokens, the posted token rates are 90% lower. For longer prompts, they are 50% lower.
The cheapest rate covers a defined shape of work
The full table makes Anthropic's target clearer. Up to 100,000 prompt tokens, cache reads cost $0.01 per million tokens and cache writes cost $0.125. Past the cutoff, those rates become $0.05 and $0.625. Haiku 4.5 charged $0.10 for cache reads and $1.25 for writes without this new split. Anthropic says roughly 90% of requests to the previous Haiku fell below 100,000 tokens, which explains why the company can describe the change as a broad reduction even with the higher long-prompt tier.
The target workloads are short, frequent jobs: classification, database queries, summaries, context compaction, and narrow subagent assignments. One early customer example in Anthropic's announcement is AlphaSense's document-question feature, which handles about 8 million calls a week. AlphaSense reported a score of 0.84 for Haiku 5.5 against 0.76 for Haiku 4.5 in a 400-query test. That is a company-supplied result on one production task, rather than a general quality verdict. It does show the volume at which a few tenths of a dollar per million tokens can change a bill.
Long-running agents create a less tidy pattern. Their prompts can grow as the system appends messages, tool outputs, retrieved files, and prior reasoning blocks. A job may begin in the lower tier and later cross the boundary without any change to the model ID. Teams that estimate cost from the first turn of a session will miss that change. The safest unit for comparison is the completed task, including retries and every tool round trip, because the advertised token rate covers only one part of the run.
There is another wrinkle. Anthropic's migration guide says the same input text produces about 30% more tokens on Haiku 5.5 than on Haiku 4.5, with the exact increase depending on the content. Existing token counts, max_tokens settings, and budget alerts therefore need fresh measurements. A prompt that previously counted near 80,000 tokens could approach or pass the new price boundary after migration. Anthropic says its 75% average cost reduction already accounts for the tokenizer, but an individual workload can land elsewhere.
The small model has moved into agent work
Anthropic positions Haiku 5.5 as a worker beneath Sonnet 5.5 or Opus 5.5, handling bounded steps while a larger model plans or reviews. That is a different role from a simple chat model. It puts price, latency, tool behavior, and failure handling into the same engineering decision. A cheap subagent that needs repeated retries can cost more than a slower model that finishes in one pass.
The company's benchmark table reports 39.2% on Terminal-Bench 4.0 for Haiku 5.5, compared with 0.0% for Haiku 4.5. On the offline subset of OSWorld 2.1, which tests computer use, Anthropic reports 72.4% against 15.7%. Those are large jumps in Anthropic's own evaluations, though the bigger Sonnet 5.5 remains ahead at 70.6% on Terminal-Bench and 83.9% on OSWorld. The practical reading is narrower than the launch copy: Haiku may now be useful for more agent steps, while the vendor's data still puts Sonnet ahead on complicated command-line and computer tasks.
Haiku 5.5 is also the first Haiku model with an adjustable effort setting. Adaptive thinking is enabled by default, and developers can choose an effort level to trade response cost against reasoning. This changes output handling. A response may begin with a thinking block even when the request does not explicitly enable thinking, so code that assumes the first content block contains the answer can break. Anthropic tells developers to select blocks by their type field instead.
Browser and computer use arrive through beta SDK toolsets. The SDK documentation makes an important boundary explicit: Anthropic supplies classes and routing, while the developer supplies the browser or desktop, implements the actions, and writes the URL policy. The sample driver is labeled as non-production code. Anthropic also advises blocking private network ranges and the cloud metadata address, then running the browser in a profile that contains no account data the model should not access. A lower inference bill does not remove those deployment duties.
A migration can fail before the first answer
Changing claude-haiku-4-5 to claude-haiku-5-5 is only the first item in Anthropic's ten-point migration checklist. Requests that use a fixed thinking budget must move to adaptive thinking. Haiku 5.5 rejects assistant-prefill turns, and most custom temperature, top_p, or top_k settings return a 400 error. Computer-use integrations on the Claude API and Google Cloud must move from the older computer_20250124 tool to the computer_toolset_20260801 toolset.
Anthropic provides a Claude Code migration command for projects that use its API:
/claude-api migrate this project to claude-haiku-5-5
The command can update model IDs and parameter changes, then produce a manual checklist. Anthropic says it asks for the migration scope before editing files. That automation still leaves application-specific work: recounting representative prompts, checking output parsers, and exercising agent loops that save or replay thinking blocks.
Stored conversations deserve their own test. A Haiku 5.5 thinking block works only in the account that produced it or an account linked to that one. Changing the system prompt, tools, or earlier messages and then sending the block back can also produce a 400 error. Services that replay one conversation store through several customer accounts may silently lose a cross-account thinking block, according to the migration guide. Within one account, altered earlier context can fail outright.
Refusals have a new integration consequence too. Haiku 5.5 can return stop_reason: "refusal", and Anthropic says there is no server-side fallback. The company describes its cyber safeguards as more restrictive than Haiku 4.5's while allowing a wider set of defensive tasks than Sonnet 5.5. An application that currently treats every stopped response as completed text needs an explicit refusal path before switching traffic.
One enterprise limitation sits outside ordinary request code. Anthropic says Priority Tier commitments for Haiku 4.5 do not carry over because Haiku 5.5 does not support Priority Tier. A team using reserved capacity has to plan that migration separately even though the model is available through the major cloud platforms.
What the price cut changes
Haiku 5.5 makes narrow model calls cheap enough to reconsider work that previously ran through rules, queues, or a larger model. The strongest candidates have measurable answers: a classifier with a labeled test set, a retrieval helper with known citations, or a subagent whose patch can be checked. Anthropic's customer examples follow that pattern. HubSpot tested the model in simulated CRM portals, while Asana ran it against tasks such as bug triage and project setup. Neither result establishes how the model will behave on a different company's data.
The first production check should happen at the 100,000-token boundary. Record total input, output, cache traffic, latency, retries, and task success on both sides of it, using Haiku 5.5's own tokenizer. Then run the same cases through the model currently in service. If Haiku wins, the reason will appear in cost per accepted result rather than the launch benchmark table. What to watch next is whether real agent sessions stay inside the cheaper tier, because that single boundary decides how much of the headline price cut reaches the bill.