Claude Haiku 5.5 explained: a tenth of the price, if you watch the tokens

Anthropic released Claude Haiku 5.5 on 7 October 2026 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, 90 per cent below Claude Haiku 4.5. It is also far stronger and faster: Artificial Analysis scores it 43 on its Intelligence Index v4.3.2, against 17 for Haiku 4.5, and measured it at about 240 output tokens a second. It is the small, quick member of the 5.5 family, below Claude Sonnet 5.5 and Claude Opus 5.5.

For a UK marketing or operations team, that makes it the obvious candidate for high-volume work such as tagging, routing and summarising. The catch is that the per-token price is not the per-job price. A new tokenizer counts more tokens for the same text, Haiku 5.5 thinks before it answers by default, prompts over 100,000 tokens cost five times as much per token, and none of the documented routes keeps processing inside the UK.

Prices, model IDs and migration rules were checked against Anthropic’s launch page, what’s new page, migration guide, effort documentation and pricing documentation on 9 October 2026. Independent figures from Artificial Analysis, Vals.ai and CursorBench were checked the same day. All prices are US list prices.

The answer in brief

Claude Haiku 5.5 is the cheapest and fastest model Anthropic sells, and the first Haiku that is good enough to do real agent work. The saving is real for short, repetitive jobs and much smaller for long prompts or high effort settings.

• Released 7 October 2026 as claude-haiku-5-5, with a 1M-token context window and up to 128K output tokens.
• Costs $0.10 input and $0.50 output per million tokens up to 100,000 tokens of prompt, then $0.50 and $2.50. Haiku 4.5 costs $1 and $5.
• Scores 43 on the Artificial Analysis Intelligence Index v4.3.2, ahead of Gemini 3.8 Flash (41) and GPT-6 Luna (38) and behind Sonnet 5.5 (56), at roughly twice their output speed.
• The same text produces about 30 per cent more tokens, and at maximum effort it writes a great deal, so Artificial Analysis measured its cost per task at only a quarter below Haiku 4.5’s.
• Effort defaults to medium; Anthropic recommends low for chat and simple high-volume requests.
• Moving from Haiku 4.5 needs code changes: temperature settings, prefilled replies and manual thinking budgets now return errors.
• There is no UK-only processing. EU-only processing is available through Amazon Bedrock or Google Cloud Vertex AI, at a 10 per cent premium on Anthropic’s pricing and without batch discounts.

What is Claude Haiku 5.5, and where can you use it?

Claude Haiku 5.5 is Anthropic’s small, fast model, described in its models overview as built for “high-volume, latency-sensitive tasks such as classification, extraction, and routing”. Anthropic’s Haiku page lists uses from live chat, voice agents and support to form filling, and the launch page pitches it as a helper that Opus 5.5 or Sonnet 5.5 can hand simple steps to, sometimes called a sub-agent. It accepts text and images, returns text, and its reliable knowledge runs to June 2026.

It is a big step up in specification. The context window grows from Haiku 4.5’s 200K tokens to 1M, and maximum output from 64K to 128K. It is the first Haiku with an effort setting, a dial that controls how much the model reasons before it answers, with five levels from low to max. It supports computer use through a new tool version and adds browser use on the Claude API and Google Cloud (Anthropic, what’s new in Haiku 5.5). Anthropic lists its retirement as not sooner than 7 October 2027.

Anthropic’s Haiku page says Free, Pro, Max, Team and Enterprise users can select Haiku 5.5 on Claude.ai, on web, iOS and Android, and it is in Claude Code. Developers can use it through the Claude API, Amazon Bedrock (anthropic.claude-haiku-5-5), Claude Platform on AWS, Google Cloud Vertex AI and Microsoft Foundry, and it became generally available in GitHub Copilot on launch day.

Anthropic is clear about its limits. Its models overview places Opus 5.5 at “long-running agentic coding and knowledge work”, Fable 5.1 at “demanding reasoning and long-horizon agentic work”, and Sonnet 5.5 as the “best combination of speed and intelligence”. The Haiku launch page adds that Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding, and its own benchmark table shows why.

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, the same list price as OpenAI’s GPT-6 Luna. Above 100,000 tokens of prompt, every token in the request costs five times as much: $0.50 and $2.50. It is the only Anthropic model priced by prompt length (Anthropic pricing documentation).

Source: Anthropic pricing documentation, checked 9 October 2026.

Three details matter when you budget. The 100,000-token threshold counts every input token in the request, cache reads and writes included, so a long cached prompt still pays the higher rate. Batch processing, for jobs that can wait, halves every price. And location costs extra: US-only processing on Anthropic’s own API is billed at 1.1 times the standard rate, and Anthropic’s pricing page says regional and multi-region endpoints on Bedrock and Google Cloud “include a 10% premium over global endpoints”.

Teams on Claude subscriptions now get some API use included. Anthropic’s launch page introduces a monthly API credit for Max and Team subscribers: $100 on Max 5x, $200 on Max 20x, and $20 per Standard seat or $100 per Premium seat on Team, pooled and capped at $500. Credits do not roll over, and Free, Pro and Enterprise plans are not eligible (Anthropic, API credits for subscribers).

Is Claude Haiku 5.5 really 90 per cent cheaper than Haiku 4.5?

Per token, yes, for prompts up to 100,000 tokens. Per job, the saving is smaller and depends on how you use it. Anthropic’s own estimate is about 75 per cent on average: its launch page footnote says Haiku 5.5 is 90 per cent cheaper below the threshold and 50 per cent cheaper above it, that “On Haiku 4.5, 90% of requests fell into the former category”, and that the figure allows for the change in how many tokens are used.

Two things eat into the saving. First, the tokenizer, the part of the model that splits text into billable pieces. Anthropic says the newer tokenizer “produces approximately 30% more tokens for the same text”, and Simon Willison measured about 1.25 times as many on his own prompt, calling it a hidden price increase. It also moves the threshold: a prompt of about 77,000 tokens on Haiku 4.5 now crosses 100,000 on Haiku 5.5.

Second, output. Thinking is billed as output tokens, including thinking you never see: by default Haiku 5.5 returns its thinking blocks empty, but you pay for the full reasoning. The usage.output_tokens_details.thinking_tokens field shows how much. Artificial Analysis calls Haiku 5.5 “very verbose”: at max effort it produced 440M output tokens across the Intelligence Index, against 78M for Haiku 4.5. Its measured cost per index task was $0.21 against Haiku 4.5’s $0.28, a quarter lower rather than nine tenths (Artificial Analysis, v4.3.2). That figure is provisional, does not yet apply the over-100,000 price tier, and compares Haiku 5.5 at its highest effort setting with Haiku 4.5’s reasoning mode.

To see what this means for a real budget, we priced two typical jobs at list prices. The short job is 10,000 support tickets a month, each with a 2,000-token prompt and a 300-token answer as Haiku 4.5 counts them. The long job is 1,000 brand-compliance checks a month, each sending a 90,000-token brief and getting 1,000 tokens back. We added 30 per cent to the token counts for the two 5.5 models and left out thinking tokens.

Source: AIWIZ calculation from Anthropic list prices, checked 9 October 2026. The 30 per cent token uplift is Anthropic’s approximation; the exact increase depends on the content.

On short tickets, Haiku 5.5 comes out 87 per cent cheaper than Haiku 4.5, and batch processing would halve its $4.55 again. On the long briefs, the prompt crosses the threshold and the saving falls to 35 per cent. Thinking is the variable we left out: every 1,000 thinking tokens per ticket adds $5.00 a month on Haiku 5.5 and $100.00 on Sonnet 5.5. Run your own prompts through the model for a week before you trust any estimate, ours included.

Which effort setting should you use for Claude Haiku 5.5?

Start with low for simple, high-volume work and medium for most other jobs. Haiku 5.5 defaults to medium, and Anthropic’s effort guidance says to use low “for chat, short tool tasks, and simple, high-volume requests”, high “for knowledge work, longer agent tasks, and strict instruction following”, and the two highest levels only where your own tests show a quality gain. Thinking is on by default, but the same page says you can switch it off at high effort or below.

Cursor’s coding benchmark shows how steeply cost climbs with effort. Each step up buys a few points at roughly double the price.

Source: Cursor, CursorBench 4.0, checked 9 October 2026. Cost per task applies each model’s published prices to the tokens it used.

The useful finding is in the middle of the table. Haiku 5.5 at high effort scored 42.3 per cent for $0.32 a task, beating Sonnet 5.5 at medium, which scored 39.2 per cent for $0.52. At max effort Haiku costs fourteen times as much as at low effort for a gain of 17.5 points. Artificial Analysis found the same pattern on its index: in its launch analysis Haiku 5.5 scored 38 at high effort using about a third of the output tokens it needed to reach 43 at max.

How does Claude Haiku 5.5 compare with Haiku 4.5, Sonnet 5.5 and its rivals?

Claude Haiku 5.5 is far ahead of Haiku 4.5 and clearly behind Sonnet 5.5 on every measure published so far. Against other small models, it leads on Artificial Analysis’s overall index and is much faster, but it is weaker on factual knowledge.

Anthropic’s launch table uses new test suites and gives few settings for Haiku 5.5, so treat it as the vendor’s best case.

Source: Anthropic, Introducing Claude Haiku 5.5, published 7 October 2026, checked 9 October 2026. Scores from earlier Terminal-Bench versions are not comparable.

On every row Haiku 5.5 is far above Haiku 4.5 and below Sonnet 5.5, and where Anthropic shows GPT-6 Luna, Haiku 5.5 beats it. The widest gap to Sonnet is Terminal-Bench 4.0, which tests multi-step work in a command line: 39.2 per cent against 70.6 per cent.

Independent testing puts Haiku 5.5 slightly lower but in the same place.

Source: Artificial Analysis model pages, Intelligence Index v4.3.2, checked 9 October 2026. Speeds are live measurements and change daily. The Haiku 5.5 cost figure is provisional.

Where the same test was run independently, the scores came in below Anthropic’s: Artificial Analysis measured 33 per cent on Terminal-Bench 4.0 and Vals.ai 35.35 per cent, against Anthropic’s 39.2 per cent. On the Vals Index, Haiku 5.5 at max effort scored 54.31 per cent, 16th of 45 models, at $2.99 per test; its best result there was 90.44 per cent on Vibe Code Bench, third of 110. It was not yet listed on the LMArena text leaderboard on 9 October 2026.

Against its two closest rivals the trade-offs are clear. GPT-6 Luna has the same list price below 100,000 tokens and scores lower (38), but it cost a third as much per index task ($0.07), because Haiku 5.5 writes so much more; Willison judged Luna “a much better deal” for longer prompts. Gemini 3.8 Flash costs $0.75 and $3.75, scores 41 and is stronger on knowledge: in its launch analysis, Artificial Analysis put Haiku 5.5’s accuracy on its AA-Omniscience knowledge test at 36 per cent, against 55 per cent for Gemini 3.8 Flash and 44 per cent for GPT-6 Luna. Our GPT-6 Sol post covers Luna’s pricing in more detail.

Early customer results, chosen by Anthropic, point the same way. HubSpot’s Ze’ev Klapow said Haiku 5.5 “got the best score we’ve seen on this suite yet, at 92.8% averaged over three runs” on HubSpot’s CRM tests, and Box reported it “scored 11 points higher than Haiku 4.5 at about half the latency”.

Security is a real improvement. In Anthropic’s system card, the share of prompt-injection attacks that succeeded within 15 attempts fell to 7.1 per cent, from 83.2 per cent for Haiku 4.5. Prompt injection is when instructions hidden in a web page, email or document hijack the model, so this matters for any team letting it read content it did not write. Attacks through computer use, where the model operates a screen, still succeeded 24.4 per cent of the time.

Can UK businesses keep Claude Haiku 5.5 data in the UK?

No. As of 9 October 2026, none of the documented routes processes Claude Haiku 5.5 requests only in the UK. The closest option is EU-only processing on Amazon Bedrock or Google Cloud Vertex AI.

Anthropic’s own API lets you choose US-only or global processing, and nothing in between (Anthropic data residency documentation). On Amazon Bedrock, the London region (eu-west-2) can send requests through the EU cross-Region profile, eu.anthropic.claude-haiku-5-5, but in-Region inference is not supported, so a request sent from London may be processed in another EU region such as Ireland or Frankfurt. On Vertex AI, Anthropic’s documentation says newer models use global or multi-region endpoints, with eu as the European option and no London endpoint. Anthropic’s Foundry documentation lists only global and US deployments for Microsoft Foundry.

The EU route has costs beyond the 10 per cent premium. Anthropic’s Bedrock and Vertex documentation both list the Message Batches API as unsupported, so you lose the 50 per cent batch discount. Bedrock also lacks structured outputs, and the new computer use and browser use toolsets are not yet available there. In our short-ticket example, the EU route would cost about $5.01 a month against $2.28 for global batch processing on Anthropic’s API.

On the legal side, the ICO’s guidance says “All the countries in the European Economic Area (EEA) have full adequacy” for UK GDPR transfers (ICO). Anthropic’s data processing addendum includes a UK addendum that applies “to any processing of Customer Personal Data that is subject to the UK GDPR”. That is not legal advice, and your data protection officer should confirm the route before personal data goes in. Our Claude Opus 5.5 post walks through the same European options for the larger model.

What breaks if you migrate from Claude Haiku 4.5?

Several request settings that worked on Haiku 4.5 now return an error, so changing the model name is not enough. Anthropic’s migration guide lists these changes:

• Manual thinking budgets (budget_tokens) return an error; switch to adaptive thinking.
• Temperature, top_p and top_k must be removed or left at their defaults. Teams that used temperature to vary marketing copy will need to get variety from the prompt instead.
• Prefilled assistant replies, a common trick for forcing a format, are rejected even with thinking off.
• Computer use on the Claude API and Google Cloud needs the new computer_toolset_20260801 tool.
• Thinking tokens count towards max_tokens, so limits tuned for Haiku 4.5 can cut answers short.
• Safety classifiers can decline a request with stop_reason: "refusal", and there is no automatic fallback, so your code needs one.
• Priority Tier commitments on Haiku 4.5 do not carry over, because Haiku 5.5 does not support it.

Claude Code can do much of the work: the migration guide gives the command /claude-api migrate this project to claude-haiku-5-5. Watch the model alias, though. In Claude Code v2.1.293 and later, haiku points to Haiku 5.5 only on Anthropic’s API; on Bedrock, Google Cloud, Foundry and Claude Platform on AWS it still points to Haiku 4.5, so name claude-haiku-5-5 explicitly there (Claude Code model configuration).

Anthropic’s deprecations page lists Haiku 4.5 as active, with retirement not sooner than 15 October 2026 and no deprecation notice posted as of 9 October 2026. That date is the earliest it could go, not a switch-off date, but it is close enough to start testing now.

What are Claude Haiku 5.5’s weaknesses?

Its biggest weakness is factual reliability. Artificial Analysis recorded a 40 per cent hallucination rate alongside its 36 per cent knowledge-test score, and Anthropic’s system card says “it hallucinated more than other recent models, and about as much as Claude Haiku 4.5”. Do not let it publish figures, prices or product claims without a check.

It is also cautious. The system card says it “over-refused more than any other model we tested” in Anthropic’s automated behavioural audit, and the launch page says its safeguards “still block penetration testing and other techniques more likely to be used by attackers”. Security teams can apply to Anthropic’s Cyber Verification Program for wider access.

Its coding ceiling is well below Sonnet 5.5 on long, multi-step tasks, and its verbosity at high effort is a cost risk. Both are manageable if you route the hard work to Opus 5.5 or Sonnet 5.5 and set effort per job.

What should UK marketing teams test first?

Start with the short, repetitive work that already costs you time: tagging search queries, product feeds or support tickets, routing enquiries, summarising calls and threads, and pulling fields out of forms and PDFs. These jobs stay under the 100,000-token threshold, usually run well at low effort, and are easy to check against a sample of human decisions.

Source: AIWIZ, based on Anthropic’s documentation, Artificial Analysis and CursorBench, checked 9 October 2026.

One policy change is worth knowing before you automate content. Anthropic’s usage policy update, published 8 October and taking effect on 12 November, adds a section on deceptive campaigns that “applies to deceptive activity of any kind (whether political or commercial)”, including amplifying content through fake accounts or posts. Networks of fake social accounts used to amplify a brand fall squarely within it.

Should your business switch to Claude Haiku 5.5?

Yes for high-volume, short-prompt work, after a measured trial. Claude Haiku 5.5 does jobs that Haiku 4.5 could not, at a fraction of the price and at speeds that suit live chat and support. Set effort to low for simple tasks, keep prompts under 100,000 tokens, and compare cost per finished job, not cost per token, against Haiku 4.5 and GPT-6 Luna.

Above Haiku, follow Anthropic’s own split. Use Sonnet 5.5 when you want speed and quality together on everyday work, and Opus 5.5 for complex, long-running agentic coding and knowledge work; Anthropic’s overview suggests starting with Opus 5.5 if you are unsure which model to use. Move up to Claude Fable 5.1 for the most demanding reasoning, or when Opus 5.5 at higher effort still falls short. If you need processing to stay in Europe, plan for Bedrock or Vertex AI and the loss of batch pricing.

Treat what comes back as a draft from a quick, capable new starter: useful at volume, but checked before it reaches a customer. If you want help choosing which workflows to move, see our AI adoption services or talk to AIWIZ.