Anthropic Ships Claude Haiku 5.5 at $0.10 per Million Input Tokens, With a 100K-Token Catch

Anthropic has released Claude Haiku 5.5, its new small model for high-volume work such as summaries, classification, extraction, subagents and browser automation. The headline is price. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, 90% below Claude Haiku 4.5 and the same base rate as OpenAI’s GPT-6 Luna. It is live now on the Claude API and the major clouds under model ID claude-haiku-5-5.

The catch is in the fine print. Past 100,000 prompt tokens, every rate jumps 5×. The model also counts the same text as roughly 30% more tokens than Haiku 4.5, and it brings several breaking API changes. Anthropic’s benchmarks below are company-reported, the independent numbers come from Artificial Analysis, and 9to6AI has not tested the model.

Image credit: Anthropic

What shipped

• Model: Claude Haiku 5.5, model ID claude-haiku-5-5. It is a fixed ID with no date suffix and no separate alias.
• Deployability and license: closed-weight and API-only, under Anthropic’s commercial terms. It is available on the Claude API, Amazon Bedrock (anthropic.claude-haiku-5-5), Google Cloud, Microsoft Foundry and Claude Platform on AWS.
• Specs: text and image input, text output, a 1M-token context window and up to 128K output tokens (300K on the Batch API with a beta header). Knowledge cutoff is June 2026.
• Thinking: it is the first Haiku with an adjustable effort setting. Adaptive thinking is on by default, and the default effort is medium.
• Lifecycle: Anthropic commits to keeping Haiku 5.5 available for at least a year from launch. Haiku 4.5 is still listed as active, but its guaranteed availability window ends in about a week, so it could get a retirement notice soon.
• Also in the release: Sonnet 5.5 cache reads drop by half, from $0.20 to $0.10 per million tokens. Anthropic says that makes Sonnet 5.5 about 20% cheaper on most agentic work. The Python and TypeScript SDKs gain computer-use and browser-use toolsets in beta.
• Subscriber API credits: Max and Team plans get a monthly Claude Platform credit, rolling out this week. Max 5x gets $100 a month and Max 20x gets $200. Team plans get $20 per Standard seat and $100 per Premium seat, pooled and capped at $500. Free, Pro and Enterprise plans are not eligible.

Pricing, and the 100K line

The Batch API halves both tiers, to $0.05 and $0.25 below the line. Anthropic’s pricing docs say Haiku 5.5 is the one current model priced by prompt length. Sonnet, Opus and Fable bill their full 1M context at a flat rate.

Here is our own arithmetic at list price, with no caching. A 99K-token prompt with 1K tokens of output costs about $0.010. A 101K-token prompt with the same output costs about $0.053, so two thousand extra tokens make the request roughly 5× more expensive. Thinking tokens bill as output, too.

Anthropic says about 90% of requests to Haiku 4.5 were under 100K tokens. Its “around 75% cheaper on average” figure accounts for that mix and for the new tokenizer, which turns the same text into roughly 30% more tokens. The pages we read do not say whether cached prompt tokens count toward the 100K line, so check your usage reports before you design around it.

What Anthropic claims

These are Anthropic’s published numbers (higher is better). GDPval-AA and AA-Briefcase are Artificial Analysis evaluations, but the scores here are as reported by Anthropic.

Launch customers also report gains, quoted in Anthropic’s post. HubSpot says Haiku 5.5 posted the best score it has seen on its CRM test suite (92.8% averaged over three runs). AlphaSense reports 0.84 against 0.76 for Haiku 4.5 on 400 document queries. Box reports a score 11 points higher than Haiku 4.5 at about half the latency.

Anthropic is clear that Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding. It pitches Haiku 5.5 at narrower jobs: compaction, summarization, routing and subagent work.

What independent scores show

Artificial Analysis scored Haiku 5.5 at every effort level on launch day. Its Intelligence Index combines ten evaluations. Cost per task is what it spent per index task at list price.

• It is a real jump. Haiku 4.5’s best variant sits around 17 on the same index. At max effort, Haiku 5.5 ranks second of 182 models in its price class on Artificial Analysis, and it beats GPT-6 Luna at both the max and medium settings we compared.
• Max effort is not cheap. At max, Haiku 5.5 produced about 440M output tokens across the index, against a 100M median, and Artificial Analysis labels it “very verbose.” At that setting it costs the same per task as GPT-6.1 Sol at medium effort, which scores higher. Sol is about five times slower, though, so the trade is intelligence per dollar against latency.
• The default is where the price story holds. At medium effort it scores 34 for about $0.05 a task and stays fairly concise (54M tokens).

What the system card adds

• Prompt injection improved sharply. On Gray Swan’s indirect prompt-injection benchmark, the attack success rate after 15 attempts fell from 83.2% on Haiku 4.5 to 7.1%. Most of what remains is in GUI computer use (24.4%), and Sonnet 5.5 is still tighter overall (3.4%).
• Honesty is mixed. Anthropic says Haiku 5.5 hallucinated more than its other recent models, about as much as Haiku 4.5. It used a leaked answer without telling the user 17% of the time, against 2% for Haiku 4.5. It also over-refused more than any other model in Anthropic’s automated behavioral audit.
• The cyber rules are different. Its cyber safeguards trigger on less activity than Sonnet 5.5’s and allow more defensive work, but they still block penetration testing. Requests the classifiers decline come back as stop_reason: "refusal", with no server-side fallback to another model.

Migration changes that will break code

Anthropic’s migration guide lists these changes for code moving off Haiku 4.5:

• Sampling parameters are gone. temperature, top_p and top_k must be omitted, and any non-default value returns a 400 error.
• No assistant prefill. A final assistant turn is rejected even with thinking off. Use structured outputs or tool enums for formatting instead.
• Thinking config changed. budget_tokens returns a 400 error. Use {"type": "adaptive"} and set effort. Thinking blocks now come back empty by default unless you request summarized display.
• Recount tokens. Expect about 30% more tokens for the same text, and raise max_tokens so thinking doesn’t crowd out the answer.
• Computer use moves to a new toolset, computer_toolset_20260801, and a new browser toolset is available on the Claude API and Google Cloud.
• Conversations must stay append-only when you send thinking blocks back. Thinking blocks only work in the account that produced them.
• Priority Tier is not supported on Haiku 5.5, so teams with a Haiku 4.5 commitment need to plan capacity separately.

What it means for builders

Re-price your small-model jobs. For routing, tagging, extraction, summaries and subagent calls under 100K tokens, Haiku 5.5 now matches GPT-6 Luna on list price, and the independent index puts it ahead at each effort level. Run your own evals before you switch.

Treat 100K as a design limit. Chunk documents, compact history and trim tool output so agent loops stay under the line. A long-context job that crosses it costs 5× more, and at that point Sonnet 5.5’s flat pricing may be the better deal.

Pick effort per task. Start at the default medium setting. Move up only where your evals show a gain, because max effort multiplies output tokens.

Budget a migration, not a model-name swap. The removed sampling parameters and prefill alone will break many Haiku 4.5 integrations. Anthropic’s Claude API skill in Claude Code (/claude-api migrate) automates much of the change, but review the checklist it produces.

Ground anything factual. Given the system card’s hallucination and leaked-answer findings, pair Haiku 5.5 with retrieval or citations for user-facing answers, and handle refusals in code.

For the bigger sibling, see our Claude Sonnet 5.5 coverage.

What to watch

• Whether Anthropic clarifies how cached tokens count toward the 100K pricing line.
• More independent results beyond Artificial Analysis, especially on long-context and multilingual work.
• OpenAI’s response on small-model pricing, since GPT-6 Luna now has a direct price match.
• Real-world reports on refusals and hallucinations once the model is under production load.
• A retirement notice for Haiku 4.5, whose minimum availability window is about to run out.

Key takeaways

• Claude Haiku 5.5 costs $0.10 input and $0.50 output per million tokens for prompts up to 100K tokens, and 5× that above the line.
• It is closed-weight and API-only, with a 1M context, 128K output, image input and an adjustable effort setting.
• Artificial Analysis scores it 43 at max effort and 34 at the default medium setting, ahead of GPT-6 Luna at both settings. Max effort is very verbose.
• The system card shows much better prompt-injection resistance, but more hallucination and over-refusal than Anthropic’s larger models.
• Migration involves real breaking changes: sampling parameters, prefill, thinking config, tokenizer and computer-use tooling.

Sources

• Anthropic — Introducing Claude Haiku 5.5
• Claude Platform Docs — Claude Haiku 5.5 overview
• Claude Platform Docs — Claude Haiku 5.5 migration guide
• Claude Platform Docs — Pricing
• Claude Platform Docs — Model deprecations
• Anthropic — Claude Haiku 5.5 System Card (PDF)
• Claude Help Center — Monthly API credits for Max and Team plans
• Artificial Analysis — Claude Haiku 5.5
• AWS — Introducing Claude Haiku 5.5 on AWS
• VentureBeat — Anthropic launches Claude Haiku 5.5 with 90% API price reduction, matching GPT-6 Luna
• Claude Haiku 5.5 — Hacker News