Claude Haiku 5.5 cuts the small-model price by 90%, but only below 100,000 tokens

Anthropic released Claude Haiku 5.5 on October 7, completing the 5.5 generation that began with Opus 5.5 on September 22 and Sonnet 5.5 on September 28. The model ID is claude-haiku-5-5, and Anthropic says it is available now on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.

The headline is price. For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, one tenth of Haiku 4.5's $1/$5. But Haiku 5.5 is also the first current Claude model whose price depends on prompt length: above 100,000 tokens it costs $0.50/$2.50, five times the short-prompt rate and half of Haiku 4.5's.

What the pricing actually looks like

According to Anthropic's pricing documentation, the two tiers are:

• Up to 100,000 tokens: $0.10 input, $0.125 for 5-minute cache writes, $0.20 for 1-hour cache writes, $0.01 for cache hits, $0.50 output.
• Over 100,000 tokens: $0.50 input, $0.625 and $1 for cache writes, $0.05 for cache hits, $2.50 output.
• Batch API: half price, so $0.05/$0.25 below the threshold and $0.25/$1.25 above it.
• Haiku 4.5, for comparison: $1 input, $0.10 cache hits, $5 output.

Two details matter for anyone estimating costs. First, the docs describe the higher rate as applying to "prompts over 100,000 tokens", which reads as the whole request moving to the higher tier rather than only the tokens past the line. On that reading, a 99,000-token prompt costs about $0.0099 in input while a 101,000-token prompt costs about $0.0505. VentureBeat notes that Anthropic's materials do not say how a request of exactly 100,000 tokens is billed, and the docs do not say whether cached tokens count toward the threshold.

Second, Haiku 5.5 uses the newer tokenizer that Anthropic introduced with Claude Opus 4.7, while Haiku 4.5 uses the older one. Anthropic's launch post says this means "slightly more tokens per task"; its pricing docs put the newer tokenizer at roughly 30% more tokens for the same text, depending on content. That is why Anthropic's own savings estimate is around 75% on average rather than 90%: the company says the cut is 90% for requests up to 100,000 tokens and 50% above that, before the tokenizer effect. VentureBeat reports that about 90% of Haiku 4.5 requests already fall in the shorter tier.

Specs and the new effort control

Anthropic's models overview lists Haiku 5.5 with a 1M-token context window, 128K tokens of maximum synchronous output, adaptive thinking, a reliable knowledge cutoff of June 2026, and a retirement commitment of not sooner than October 7, 2027. Its default effort is medium.

Effort is the more consequential change for developers. Haiku 5.5 is the first Haiku-class model with adjustable effort, from Low to Max. That turns Haiku from a fixed "cheap and fast" option into a model you can tune per call, and it means published benchmark scores need to be read alongside the effort level that produced them.

Benchmarks, as reported by Anthropic

Anthropic compares Haiku 5.5 with Haiku 4.5, OpenAI's GPT-6 Luna and its own Sonnet 5.5. All figures below come from Anthropic's own testing; independent evaluation is not yet available.

• OSWorld 2.1 (offline subset): Haiku 5.5 72.4%, Haiku 4.5 15.7%, GPT-6 Luna 48.9%, Sonnet 5.5 83.9%.
• Terminal-Bench 4.0: Haiku 5.5 39.2%, Haiku 4.5 0.0%, GPT-6 Luna 16.4%, Sonnet 5.5 70.6%.
• FrontierCode 1.1 (Main): Haiku 5.5 46.4%, GPT-6 Luna 42.4%, Sonnet 5.5 52.1%.
• GDPval-AA v2.1 (Elo): Haiku 5.5 1620, Haiku 4.5 735, GPT-6 Luna 1437, Sonnet 5.5 1840.
• Humanity's Last Exam, with tools: Haiku 5.5 57.4%, Haiku 4.5 18.7%, Sonnet 5.5 64.5%.

VentureBeat points out that the 39.2% Terminal-Bench result was achieved at maximum effort, and that Haiku 5.5 scores about 20% at its default medium effort. That gap is a useful reminder: if you switch to Haiku 5.5 for coding-agent work and leave effort on the default, do not expect the headline number. On every evaluation Anthropic shows, Sonnet 5.5 stays ahead.

Anthropic also cites customer results, including HubSpot reporting 92.8% averaged over three runs on its CRM test suite, Box reporting an 11-point gain over Haiku 4.5 at about half the latency, and Asana reporting over 30% lower task-completion latency. These come from the customers' own internal evaluations and cannot be compared with each other.

The competitive picture

The short-prompt price is not unique. VentureBeat reports that GPT-6 Luna is also priced at $0.10/$0.50, with a higher rate above 272,000 input tokens. So both labs now sell their small models at the same base price, and the practical difference is where each one raises prices for long prompts: Anthropic at 100,000 tokens, OpenAI at 272,000. For long-document workloads, that threshold may matter more than the base rate.

Other changes in the same announcement

• Sonnet 5.5 cache reads halved from $0.20 to $0.10 per million tokens, which Anthropic estimates lowers the cost of most agentic work by about 20%. VentureBeat reports some Azure and Google Cloud customers will get the reduction over the following days.
• API credits for subscribers: $100 a month on Max 5x, $200 on Max 20x, and up to $500 pooled on Team.
• SDKs: the Python and TypeScript SDKs add beta support for computer use and browser use.
• Safeguards: Anthropic says Haiku 5.5's cybersecurity safeguards are stricter than Haiku 4.5's and looser than Sonnet 5.5's, allowing more defensive tasks while still blocking penetration testing and similar attacker-oriented work.

What changes for developers

For classification, extraction, routing, summaries and subagents with short prompts, Haiku 5.5 is now the obvious Claude candidate to test, and the price drop is large enough to change where it makes sense to call a model at all. The caveats are concrete: re-measure token counts with the new tokenizer before trusting old cost estimates, keep an eye on prompt length so routine requests do not drift over 100,000 tokens, and set effort explicitly instead of relying on the default when quality matters. Anthropic's own docs still describe Haiku as the model for high-volume, latency-sensitive tasks, and its numbers still favor Sonnet for harder agentic work.