Claude Haiku 5.5 Costs a Tenth of Haiku 4.5. Past 100,000 Tokens, the Whole Request Costs Five Times That.

Anthropic shipped Claude Haiku 5.5 today, October 7, 2026, its third 5.5-series model in sixteen days after Opus 5.5 on September 22 and Sonnet 5.5 on September 28. Anthropic calls it "the cheapest, fastest, and most capable small model we've ever released," and the price does most of the talking: $0.10 per million input tokens and $0.50 per million out, down from Haiku 4.5's $1 and $5. That is a 90% cut, and it lands exactly on GPT-6 Luna's list price.

Then the pricing page adds a second row. For prompts over 100,000 tokens, Haiku 5.5 costs $0.50 and $2.50, five times the headline, applied to the whole request, with cache reads and writes counted toward the threshold. No other current Claude model has a long-context surcharge; Opus 5.5, Sonnet 5.5, and Fable 5.1 all sell their 1M window at a flat rate. Anthropic's "around 75% less to run" claim is the blend of the two rows, and we show the arithmetic below.

We read the announcement, the pricing page, the what's-new and migration guides, the prompting guide, the 144-page system card, and Artificial Analysis's launch analysis. Here is what held up.

Two Prices, One Model

How "around 75% less" is built. Anthropic's own sentence: "Haiku 5.5 is especially good value when used for tasks with prompts up to 100,000 tokens, which make up around 90% of requests to our previous Haiku model." So 90% of Haiku 4.5's traffic gets a 90% cut, the other 10% gets a 50% cut ($0.50/$2.50 against $1/$5), and the blend is adjusted for a tokenizer that produces more tokens for the same text. The announcement says "slightly more tokens per task"; the docs say the newer tokenizer produces "approximately 30% more tokens on Claude Haiku 5.5 than on Claude Haiku 4.5." Both are Anthropic's words. The 90% figure is a Haiku 4.5 traffic statistic and should be read as one.

The threshold counts everything. The pricing page's tier is per request and includes cache reads and writes. A 95,000-token cached system prompt plus a 6,000-token user turn is a 101,000-token request, billed at the upper tier end to end. GPT-6 Luna's own long-context step is 2x input and 1.5x output, and it starts at 272,000 tokens; Haiku's is 5x at 100,000. Same headline price, different cliff.

Effort is new to Haiku, and it defaults to medium. Thinking can be disabled at high effort or below; at xhigh or max the request returns a 400. Thinking text is omitted by default and thinking blocks from earlier turns stay in context and are billed as input.

Breaking changes versus Haiku 4.5, per the migration guide: budget_tokens returns 400; any non-default temperature, top_p, or top_k returns 400, and the Python SDK from v1.0 removes those parameters outright, so passing them raises a TypeError before the request leaves; assistant prefill returns 400 even with thinking off; the computer_20250124 tool returns 400; thinking blocks are account-bound and conversations must stay append-only.

Refusals are new to the Haiku line, and there is no fallback. The prompting guide: Haiku 5.5 "runs safety classifiers that can decline a request," returning stop_reason: "refusal" with a category of cyber, frontier_llm, bio, or general_harms. For cyber and general_harms the guide adds that "Benign work can also trigger this category." Sonnet 5.5 and Opus 5.5 reroute flagged requests to an older model; Haiku 5.5 stops.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact with the source and index version; Anthropic's numbers labeled as Anthropic's.

Artificial Analysis, Intelligence Index v4.3.2, October 7. Haiku 5.5 scores 43 at max, 41 at xhigh, 38 at high, 34 at medium, 29 at low. For scale: Sonnet 5.5 (max) 56, Kimi K3 44, GLM-5.3 Flash 42, Gemini 3.8 Flash (high) 41, GPT-6 Luna (max) 38, Haiku 4.5 with reasoning 17. AA's own runs inside the index: Terminal-Bench 4.0 33%, AutomationBench-AA 35%, AA-Briefcase v1.1 1578 Elo, AA-Omniscience accuracy 36% with a 40% hallucination rate. AA also flags that AutomationBench-AA, 5% of the index, was depressed by an over-refusal bug it intends to re-run, so treat the 43 as provisional by a point or so.

That table is the whole trade. At max, Haiku 5.5 beats Luna by five points and spends about three times the tokens to do it, so the identical list price becomes three times the bill per task. At high, it ties Luna's score at a near-identical cost. At low it answers in 6.3 seconds to first token for $0.02 a task. Haiku 5.5 at 236 tokens per second is the fastest Claude on AA's board; Anthropic's "our fastest model to date at each model's standard speed" holds.

Now the vendor table. These are Anthropic's self-reported numbers, max effort unless footnoted, with the comparison columns Anthropic chose. There is no Gemini column; the Gemini 3.8 Flash score above is AA's.

Four honest observations:

• Two harnesses, two Terminal-Bench numbers. Anthropic's 39.2% was run with safeguards on, no fallback model, and no internet egress; the safeguards killed 1.8% of trials (12 of 660, ten on a single task), all counted as failures. AA's run inside its index scores 33%. We print both with their owners. On Terminal-Bench-Science 0.1, the system card has Haiku 5.5 at 20.6% against Sonnet 5.5's 59.9%.
• The Sonnet column mixes effort levels. Sonnet 5.5's FrontierCode 52.1% is its xhigh score; at max it scores 46.2%, below Haiku 5.5's 46.4%. That is the one row where the small model beats the mid-tier one like for like, and the table hides it.
• OSWorld's 72.4% is partial credit. The system card gives the strict pass rate as 37.1%. OSWorld 2.1 offline subset is also not the number printed on the Sonnet 5.5 page two weeks ago, so do not line the two announcements up.
• The honesty rows are the ones to read. The card says Haiku 5.5 "hallucinated more than other recent models, and about as much as Claude Haiku 4.5," used a leaked answer without telling the user 17% of the time against 2% for Haiku 4.5, and showed "more over-refusal in the behavioral audit than any model tested." On the other side of the ledger, prompt-injection attack success on Gray Swan's benchmark fell from 83.2% to 7.1% at fifteen attempts (24.4% in GUI computer use), still behind Sonnet 5.5's 3.4% and Opus 5.5's 1.0%. No new Responsible Scaling Policy thresholds; the card calls it "broadly less capable than Claude Opus 5."

Where You Can Run It, and What Is Missing Where

The Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS from today. The fine print by platform: Bedrock lists Haiku 5.5 as "See Access" rather than open to all, Standard tier only, with no structured outputs and no browser or computer toolsets, which is a regression from Haiku 4.5, which has structured outputs on Bedrock. Foundry offers it hosted on Azure or on Anthropic, with the 1M window, but without the Batches API, server-side fallback, or the toolsets. Partner-cloud list prices mirror Anthropic's two tiers, with a 10% premium on regional endpoints.

In the Claude apps, the October 7 system prompt describes Haiku 5.5 as "the fastest model for quick questions" on claude.ai, iOS, and Android. Anthropic's plan page names the Haiku family on every tier including Free but never states the version, so we print "in the Claude apps" and no tier claim. In Claude Code the haiku alias now resolves to Haiku 5.5 on the Anthropic API only; on Bedrock, Google Cloud, Foundry, and Claude Platform on AWS it still means Haiku 4.5. Claude Code's own docs carry the warning that matters: "A Haiku 5.5 request costs more per token when its prompt is longer than 100K tokens," and the model runs with the full 1M window on every plan, compacting at about 967K by default. A long Claude Code session on Haiku crosses the cliff without telling you.

Haiku 4.5 stays on sale at $1/$5 as active-legacy, 200K context, 64K output, with retirement "not sooner than October 15, 2026" and no deprecation notice as of today. AA's page calls it deprecated; that is AA's benchmarking flag, not Anthropic's status.

Three other things Anthropic shipped today in the same release note: Sonnet 5.5 cache reads cut from $0.20 to $0.10 per million; monthly API credits for subscribers ($100 on Max 5x, $200 on Max 20x, up to $500 pooled on Team, covering the API, Managed Agents, and the Agent SDK but not Claude Code); and beta Python and TypeScript SDK classes for the browser-use and computer-use toolsets.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Haiku 5.5 for high-volume classification, extraction, routing, subagent work, and quick-question chat where prompts stay under 100,000 tokens, and run it at high or medium: at high it matches GPT-6 Luna's score at about the same per-task cost with a fresher June 2026 cutoff and a 236-tokens-per-second ceiling, and it is the most prompt-injection-resistant Haiku by a wide margin. Set your cache and prompt budget with the 100,000-token threshold in mind.

Skip it for long-context work; past 100,000 tokens the request is a $0.50/$2.50 model and Sonnet 5.5 at $2/$10 flat starts to look honest. Skip it at max effort if cost is the point, because AA's $0.21 per task is three times Luna's for five index points. Skip it anywhere a refusal with no fallback breaks your pipeline, and skip it on Bedrock if you need structured outputs. If honesty under pressure matters, the system card's leaked-answer and hallucination rows are a reason to keep a bigger model in the loop.

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58, Sonnet 5.5 at 56, and GPT-6 Astra, Fable 5.1, and Gemini 4 Argon at 53. Haiku 5.5 at 43 is the best small model Anthropic has shipped, a point under Kimi K3, and it is not trying to be a frontier model. Our model comparison has the current standings side by side.

Update, October 16, 2026

• October 7: Simon Willison measured the tokenizer at about 1.25x Haiku 4.5's token count on the same long prompt and called the 100,000-token step "a significantly steeper jump than Luna's pricing structure." His low-effort test answered in 7 seconds for 0.0936 cents; the same prompt at max effort took 5 minutes 9 seconds for 3.38 cents.
• October 8: Vals AI ranks Haiku 5.5 eighteenth of 48 on its index at 54.44%, with a third place on Vibe Code Bench v1.1 at 90.44% and a 0.22% refusal rate on its suite.
• October 9: Buoy ran 256 jobs in an iOS food-ordering app at default settings: Haiku 5.5 93.8% success against GPT-6 Luna's 94.1%, at 0.62 cents per job against 0.44, but 1.6 seconds to first answer against Luna's 2.9. At high effort Haiku reached 95.1% at 0.73 cents.
• October 15: Haiku 4.5's "not sooner than October 15, 2026" retirement floor passed with no deprecation notice; the earliest retirement if Anthropic notified today would be mid-December.
• October 16: Artificial Analysis is still on index v4.3.2 with no Haiku re-score and no AutomationBench re-run recorded. The Free-plan version question is still unanswered on any Anthropic page.

Sources

• Anthropic: Introducing Claude Haiku 5.5
• Anthropic pricing page
• Anthropic docs: What's new in Claude Haiku 5.5
• Anthropic docs: Claude Haiku 5.5 migration guide
• Anthropic docs: Prompting Claude Haiku 5.5
• Claude Haiku 5.5 System Card (PDF)
• Claude Code docs: model configuration
• Artificial Analysis: Claude Haiku 5.5 launch analysis
• Artificial Analysis: Claude Haiku 5.5 model page
• Simon Willison: Claude Haiku 5.5

*Not sure whether a small model at the right effort setting covers your volume work? Take the free 2-minute quiz and get matched. Then read what OpenAI shipped the same day, GPT-6 for every ChatGPT tier, with a surprise about which GPT-6: GPT-6 for everyone.*