What Is Claude Haiku 5.5? Anthropic's Fastest Small Model Explained

Claude Haiku 5.5 is the model for tasks that are simple but happen thousands of times a day. Anthropic released it on October 7, 2026, at a tenth of its predecessor's per-token rate for most requests.

If you pay for a large model to sort emails, summarize calls, or answer routine support questions, that gap matters. Most of those jobs don't need the most powerful model, and paying for one adds up fast.

Here's what Claude Haiku 5.5 is, what changed from Haiku 4.5, what it costs, and how it scores in Anthropic's tests and in independent ones. All figures are as of October 2026.

What is Claude Haiku 5.5?

Claude Haiku 5.5 is Anthropic's fastest and lowest-cost Claude model, built for high-volume, time-sensitive work such as classification, extraction, routing, and subagent tasks. Released on October 7, 2026, Haiku 5.5 costs $0.10 per million input tokens for prompts up to 100,000 tokens and reads up to 1 million tokens at once.

In its launch announcement, Anthropic names the jobs it was built for:

• Quick, repetitive work: summaries, compressing long conversations, database lookups, and sorting requests into categories.
• Subagent tasks: a larger model such as Opus 5.5 or Sonnet 5.5 plans the work and hands Haiku the small, well-defined pieces.
• Speed-sensitive jobs: live customer support and browser automation, where every second of waiting counts.

A subagent is simply a helper model. Think of a lead model as the manager who decides what needs doing, and Haiku as the assistant who fetches a figure from a report or files a ticket. Developers call it with the model ID claude-haiku-5-5.

Where Haiku 5.5 fits in the Claude 5.5 family

Haiku 5.5 is the fastest and cheapest of Anthropic's four current model tiers. It costs one-twentieth of Sonnet 5.5 per token on short prompts, and it is the tier Anthropic recommends for volume rather than difficulty.

The Claude model lineup, pricing as of October 2026. All four models share a 1-million-token context window, 128,000-token output, and a June 2026 knowledge cutoff. Source: Claude Platform docs.

The practical takeaway is to treat these tiers as a team, not a ranking. A bigger model is worth its price on hard problems, while Haiku handles the many small jobs around them.

What's new in Claude Haiku 5.5

Haiku 5.5 brings the main features of Anthropic's larger 5.5 models down to the lowest price tier. Five changes matter most.

1. A much lower price, with a 100K-token line

Haiku 5.5 charges 90% less per token than Haiku 4.5 for prompts up to 100,000 tokens, and 50% less above that. Anthropic says about 90% of requests to Haiku 4.5 fell under the line, so most workloads land in the cheap tier.

The average saving is lower, at about 75%. Haiku 5.5 uses a newer tokenizer that counts the same text as roughly 30% more tokens than Haiku 4.5 did, which eats into the per-token discount. Text that cost $1 to read on Haiku 4.5 costs about $0.13 on Haiku 5.5 in the low tier, not $0.10. Add the higher rate on prompts over 100,000 tokens, and Anthropic puts the average saving at about 75%.

2. Adjustable effort, for the first time on a Haiku model

Effort controls how long the model thinks before answering. Haiku 5.5 offers settings from low to max and starts at medium. Low suits quick labels and routing, while higher settings help on tricky, multi-step checks.

Haiku 4.5 had no effort dial, only a fixed thinking budget that developers set by hand. More effort means better answers on hard tasks, but also more thinking tokens on your bill.

3. A 1-million-token context window and 128K-token output

The context window is how much text a model can read in one request. Haiku 5.5 takes up to 1 million tokens, five times Haiku 4.5's 200,000, and can write up to 128,000 tokens in a single reply. Through Anthropic's Batch API, a beta option raises that output limit to 300,000 tokens.

That lets Haiku read the same large inputs as the bigger model directing it, such as a full contract set or a long support history. Reading that much, though, pushes a request past the 100,000-token line and into the higher price tier.

4. Much stronger computer and browser use

Haiku 5.5 can now operate a computer reliably enough to be useful. On OSWorld 2.1, a test of completing tasks on a real desktop, Anthropic reports 72.4% against 15.7% for Haiku 4.5.

Anthropic points to repetitive jobs like form filling, data entry, and moving information between apps. It also added computer use and browser use tools to its Python and TypeScript SDKs in beta on launch day.

5. Built for subagent work

Anthropic designed Haiku 5.5 to work under a larger model. Fable or Opus plans the work, and Haiku runs the small subtasks, which makes it practical to run many agents in parallel.

Customers describe the same split. Financial AI company Rogo says a bigger model builds a presentation while Haiku 5.5 pulls the revenue figure it needs from an annual report.

Claude Haiku 5.5 vs Haiku 4.5: what changed

Haiku 5.5 is a generational jump over Haiku 4.5: five times the context, a tenth of the short-prompt price, and usable scores on tasks Haiku 4.5 could barely attempt.

Claude Haiku 4.5 vs Claude Haiku 5.5 specs, pricing, and scores, as of October 2026. Sources: Anthropic and Artificial Analysis.

Haiku 4.5 is still available, but Anthropic now lists it as a legacy model and recommends moving to Haiku 5.5.

How much does Claude Haiku 5.5 cost?

Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Once a prompt goes over 100,000 tokens, every rate rises five times.

Claude Haiku 5.5 API pricing as of October 2026, from Anthropic's Claude Platform docs. The Batch API takes 50% off input and output.

A token is a small chunk of text, often a whole short word or part of a longer one. Input is what you send the model, and output is what it writes back.

In everyday terms, the low tier is very cheap. A support reply that reads 3,000 tokens of context and writes 500 tokens costs about $0.00055, so 1,000 replies come to roughly $0.55. A 150,000-token contract summary is a different story: at the higher tier, 1,000 of them with 2,000-token outputs cost about $80.

Two more factors move the real bill. Hidden thinking tokens count as output, so higher effort settings cost more per request. On the other side, Anthropic says prompt caching can save up to 90% when you reuse the same instructions or documents, and batch processing saves 50% on work that can wait.

Claude Haiku 5.5 benchmarks: Anthropic's numbers vs independent tests

Anthropic's tests show Haiku 5.5 far ahead of Haiku 4.5 and ahead of OpenAI's GPT-6 Luna, but still behind Sonnet 5.5 everywhere. Independent results from Artificial Analysis confirm a big jump and add two caveats Anthropic's table leaves out.

1. What Anthropic reports

Vendor-reported results from Anthropic's Claude Haiku 5.5 launch announcement, October 2026. Elo scores are relative ratings, not percentages.

The gap to Sonnet 5.5 is widest on long coding jobs. Haiku 5.5 completes about four in ten Terminal-Bench tasks, against seven in ten for Sonnet 5.5, which is why Anthropic still recommends its larger models for complex coding.

2. What Artificial Analysis measured

Artificial Analysis, an independent benchmarking firm, scores Haiku 5.5 at 43 on its Intelligence Index at max effort. That is 26 points above the previous Haiku and a little ahead of other small models such as GPT-6 Luna at 38, according to its launch analysis. Sonnet 5.5 still leads at 56. Its own Terminal-Bench 4.0 run put Haiku 5.5 at 33%, below the 39.2% in Anthropic's table, because the two test setups differ.

The same analysis flags two caveats:

• Heavy token use: at max effort, Haiku 5.5 used about 162,000 output tokens per test task, roughly three times GPT-6 Luna. More tokens mean a higher bill per finished job, even at the same per-token price.
• Less stored knowledge, fewer made-up answers: Haiku 5.5 answered 36% of factual questions correctly, below GPT-6 Luna's 44%. Its hallucination rate was lower, though, at 40% against 77% for Luna.

Artificial Analysis also expects one weak score to rise. On AutomationBench, which tests multi-step business workflows, Haiku 5.5 scored 35%. A safety issue during pre-release testing made the model refuse too often, and Artificial Analysis plans to re-run the test once Anthropic's fix is in place.

What Claude Haiku 5.5 is good at, and where it falls short

Haiku 5.5 is strongest on short, repeated, well-defined jobs and weakest on long, open-ended ones. The table below sums up where it fits.

Where Claude Haiku 5.5 fits, based on Anthropic's guidance and benchmark results as of October 2026.

Early customers report results that match the strong rows. AlphaSense, which runs about 8 million document questions a week, saw a statistically significant accuracy gain over Haiku 4.5 on 400 test queries. Box reported scores 11 points higher than Haiku 4.5 at about half the latency. Both are testimonials Anthropic published, not independent tests.

The weak rows have simple workarounds. Hand complex builds to a larger model, and give Haiku the source material rather than asking it to recall facts unaided.

Where you can use Claude Haiku 5.5

Claude Haiku 5.5 is available in the Claude apps for every plan, through Anthropic's API, and on the three major cloud platforms.

Claude Haiku 5.5 availability as of October 2026, from Anthropic's Haiku page and the GitHub changelog.

Anthropic has committed to keeping Haiku 5.5 available until at least October 7, 2027. That gives teams that build on it a full year before any retirement.

Should you switch from Haiku 4.5?

For most high-volume tasks, yes. Haiku 5.5 is cheaper per token on short prompts and far more capable, but swapping the model name is not the whole job.

Check three things before you move a live product or automation:

• Settings that now fail: Haiku 5.5 rejects any non-default value for temperature, top_p, or top_k. A setup that pinned one of these, such as a classifier set to temperature 0, will return an error until the setting is removed.
• Thinking is on by default: Haiku 5.5 decides when to think, and that thinking shows up in its responses and on your bill. Pick an effort level on purpose rather than relying on the default.
• Token counts go up: the newer tokenizer counts the same text as about 30% more tokens. Recount your prompts, both for cost and to see which ones now cross the 100,000-token line.

Plan the move soon. Anthropic's docs list Haiku 4.5's retirement as no sooner than October 15, 2026, just eight days after Haiku 5.5 launched, so start testing Haiku 5.5 on a slice of real traffic now.

What Claude Haiku 5.5 means if you're building AI-powered apps

Haiku 5.5 is a reminder that the right AI model depends on the job, not the leaderboard. Most features inside an app, like tagging support tickets or summarizing a customer's history, don't need the most powerful model available.

A simple way to think about it: match each feature to the smallest model that does it well. Use a fast, low-cost model for frequent, simple jobs, and save a larger one for the work where mistakes are expensive. Splitting features this way can cut your running costs sharply as usage grows.

Use Haiku 5.5 for volume and a larger model for depth

Claude Haiku 5.5 is Anthropic's answer to the jobs that happen all day: sorting, summarizing, looking things up, and helping a bigger model finish its work. It costs a tenth of Haiku 4.5's rate on short prompts, reads up to 1 million tokens, and finally handles computer use well enough to be practical.

It is not a replacement for Sonnet 5.5 or Opus 5.5. Independent tests show a real jump in capability, but also heavy token use at max effort and weaker recall of facts. Long prompts over 100,000 tokens also cost five times as much.

The smart next step is to list the AI tasks you run most often and try the simplest ones on Haiku 5.5 at medium effort. If you're turning those tasks into an app, choose from Claude, GPT, and Gemini models for each feature and start building on Emergent.