Claude Haiku 5.5: Ten Cents per Million Tokens, and the Fine Print Behind It

Two weeks ago I wrote that the comparison that mattered most for high-volume AI work, OpenAI’s GPT-6 Luna against Anthropic’s small model, was still to come. It arrived on Tuesday, 7 October 2026. Anthropic released Claude Haiku 5.5, the last model in its 5.5 generation after Opus 5.5 (22 September) and Sonnet 5.5 (28 September). The headline is price: $0.10 per million input tokens and $0.50 per million output tokens, the same as Luna and a tenth of what Haiku 4.5 cost. For anyone running classification, extraction, routing or sub-agent pipelines, the small tier is where most of the token bill sits, so this launch matters more in practice than a new flagship would.

What was released

Model ID: claude-haiku-5-5 on the Claude API, with the same ID on Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-haiku-5-5 on Amazon Bedrock. It is a fixed snapshot with no date suffix.
Context and output: a 1M-token context window and 128K tokens of output, according to Anthropic’s models overview. That is the same window as Sonnet and Opus, and five times the 200K of Haiku 4.5.
Thinking: adaptive thinking is on by default at medium effort, with levels from low to max.
Retirement commitment: no sooner than 7 October 2027.

Pricing

Per million tokens	Prompt ≤ 100K tokens	Prompt > 100K tokens
Input	$0.10	$0.50
Output	$0.50	$2.50
Cache read	$0.01	$0.05
Cache write	$0.125	$0.625

Haiku 4.5 cost $1 input and $5 output. So list prices are 90% lower below the 100K threshold and 50% lower above it. Anthropic says about 90% of Haiku 4.5 requests fall below 100K tokens, and estimates that real workloads cost about 75% less once token-count changes are included. On the same day Anthropic halved Sonnet 5.5’s cache-read price from $0.20 to $0.10 per million tokens.

The benchmarks (vendor-reported)

Benchmark	Haiku 5.5	Haiku 4.5	GPT-6 Luna	Sonnet 5.5
GDPval-AA v2.1 (Elo, knowledge work)	1620	735	1437	1840
OSWorld 2.1 (computer use)	72.4%	15.7%	48.9%	83.9%
Terminal-Bench 4.0 (agentic coding)	39.2%	0.0%	16.4%	70.6%
Humanity’s Last Exam (no tools / tools)	45.9% / 57.4%	10.2% / 18.7%	—	—

The jump over Haiku 4.5 is very large: from almost useless to competent on computer use and on terminal-based coding agents. Against Luna, Haiku leads on every row Anthropic published.

Three caveats apply. These are Anthropic’s numbers, not independent results. The Terminal-Bench figure was measured at maximum effort; VentureBeat reports that at the default medium effort the score is roughly 20%, which still beats Luna but by much less, and max effort burns far more thinking tokens. And Sonnet 5.5 stays clearly ahead on everything, especially coding (70.6% against 39.2%). Anthropic itself positions Haiku for “classification, extraction, and routing”, not for writing your codebase.

The fine print developers should read

The 100K cliff. Above 100,000 prompt tokens, every price multiplies by five. Luna’s long-context surcharge starts at 272K. A 1M-token window with a 100K price threshold means long-document and big-repo jobs should be costed separately; they will not see the 90% cut.

The tokenizer. Haiku 5.5 uses the newer tokenizer introduced with Claude Opus 4.7. Anthropic’s migration guide says the same text produces about 30% more tokens than on Haiku 4.5. A simple calculation: 1.3 × $0.10 is about $0.13 for what used to cost $1 of input below 100K, a real 87% saving. Above 100K it is 1.3 × $0.50 = $0.65 against $1, only about 35%. Thinking tokens, now on by default, add to output on top of that.

Breaking API changes. Migrating is not just a model-ID swap. According to the official migration guide:

temperature, top_p and top_k must be removed; non-default values return HTTP 400.
Assistant prefill is rejected even with thinking off. Use structured outputs or tools for formatting instead.
thinking: {type: "enabled", budget_tokens: N} returns 400; use adaptive thinking plus the effort parameter.
Responses may start with thinking blocks, so code that reads content[0] as the answer must select blocks by type.
Thinking blocks are bound to the account that produced them and to an unchanged conversation prefix, so pipelines that edit history or replay conversations across accounts need changes.
Computer use moves to the new computer_toolset_20260801; the old tool returns 400.
Requests can end with stop_reason: "refusal" from safety classifiers, and Priority Tier is not supported on Haiku 5.5.

Code written for “cheap, deterministic, temperature 0” pipelines on Haiku 4.5 is exactly the kind that breaks.

Why it matters

The small tier is what runs at scale: the sub-agents an orchestrating model delegates to, the classifier in front of every support ticket, the extractor reading every invoice. In two weeks both leading Western labs have pushed that tier to the same $0.10 / $0.50 price point, while claiming capabilities on computer use and agentic tasks that their mid-tier models struggled with a year ago. Combined with the Sonnet 5.5 cache price cut, Anthropic is clearly targeting agent workloads, where cached context is re-read on every step.

The practical consequence is architectural. If a small model can reliably operate a browser or a terminal for a tenth of a cent per call, it becomes reasonable to fan out many cheap agents and send only the hard cases to Sonnet or Opus. Routing, not model choice, becomes the main cost lever.

Open questions

Independent numbers. None of the benchmarks have yet been reproduced by third parties. The useful test is cost per completed task on your own workload, including retries and thinking tokens, not price per token.
Determinism. Removing temperature control is a step back for teams that relied on low-temperature runs for reproducible extraction. Anthropic’s answer is structured outputs; whether that is enough for every pipeline is not yet clear.
Safety trade-offs. Anthropic says cyber safeguards are tighter than on Haiku 4.5, and coverage notes that penetration-testing requests are blocked under standard safeguards. Security teams using Haiku for automated scanning may hit refusals. Unite.AI, citing Anthropic’s system card, also notes occasional over-refusal of benign requests.
The price war’s floor. Matching Luna to the cent suggests both companies are pricing against each other rather than against cost. How long that is sustainable, and what happens to the open-weight models competing in this tier, is the story to watch next.