Claude Haiku 5.5: Anthropic Cuts Its Small Model's Price by 90%
On October 7, 2026, Anthropic shipped Claude Haiku 5.5 and led with a price cut, not a benchmark crown: $0.10 per million input tokens and $0.50 per million output for prompts under 100,000 tokens — a tenth of what Haiku 4.5 charged. The interesting part is what sits under the headline number: a 5x price cliff at 100K tokens, a tokenizer that counts roughly 30% more tokens for the same text, and benchmark gains that are all vendor-reported.
- 90% cheaper per token, 75% per job — Under 100K-token prompts, Haiku 5.5 costs $0.10/$0.50 per million tokens versus Haiku 4.5's $1/$5. Anthropic's own average-savings figure is around 75%, because a newer tokenizer counts roughly 30% more tokens for the same text.
- The benchmarks are vendor-run — Anthropic reports OSWorld 2.1 (offline subset) jumping from 15.7% to 72.4% and Terminal-Bench 4.0 from 0% to 39.2%. The independent Vals AI run measured 35.35% on Terminal-Bench 4.0 — lower, but in the same neighborhood.
- Mind the 100K cliff and the effort dial — Prompts above 100,000 tokens pay 5x rates ($0.50/$2.50), and the headline benchmark scores are max-effort numbers — at the default medium effort, published scores drop.
This article is based on Anthropic's launch page and model documentation, read on October 11, 2026, cross-checked against independent coverage (Simon Willison, The Decoder, VKTR, Quartz, Cryptobriefing) and an independent benchmark run by Vals AI. All benchmark figures are vendor-reported unless stated otherwise. We have not run a billed long-term workload on Haiku 5.5 ourselves; treat the cost math as arithmetic, not our own measurement.
Anthropic released Claude Haiku 5.5 on October 7, 2026, and pitched it as its fastest, most affordable small model to date — built for high-volume, latency-sensitive work like classification, routing, extraction, summarising and subagent tasks, not for frontier reasoning. The previous Haiku was nearly a year old and had aged badly on price: $1 per million input tokens and $5 per million output was expensive for a small model even in 2025, and a full 10x the price of OpenAI's GPT-6 Luna.
The new price sheet
Haiku 5.5 exactly matches Luna's $0.10/$0.50 base rate, but with a wrinkle Luna doesn't have: a hard pricing step at 100,000 prompt tokens. Here is the full structure from Anthropic's launch materials:
| Charge (per million tokens) | Haiku 5.5, up to 100K | Haiku 5.5, over 100K | Haiku 4.5 |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache reads | $0.01 | $0.05 | $0.10 |
| Cache writes | $0.125 | $0.625 | $1.25 |
Source: Anthropic's launch pricing table. Alongside the launch, Anthropic also halved cache-read pricing on Claude Sonnet 5.5 to $0.10 per million tokens, which it estimates cuts the cost of most agentic workloads by around 20%.
The step matters in practice, not just on paper. Anthropic says roughly 90% of requests to the previous Haiku model fell under 100,000 tokens, so most existing users land in the cheap tier. But the boundary is a cliff, not a ramp: a 99,000-token prompt bills every token at $0.10 input, while a 100,001-token prompt bills every token at $0.50 — five times the rate on the whole request. Long-document summarisation and big-codebase analysis sit exactly where that cliff bites.
Why "90% cheaper" and "75% cheaper" are both true
The per-token cut is genuinely 90% on input and output below 100K tokens. The average saving is not, because the migration guide documents that Haiku 5.5's tokenizer turns the same text into roughly 30% more tokens than its predecessor did. Anthropic's own claim is careful: "on average, it now costs around 75% less to run," with footnotes noting the model is "priced 90% lower than Claude Haiku 4.5 for requests up to 100,000 tokens, and 50% lower for requests over 100,000 tokens." Simon Willison flags the tokenizer change as a hidden price increase; Anthropic's position is that the 75% average figure already accounts for it. Both things can be true — and for budgeting, 75% is the number that survives contact with a real invoice.
The effort dial: free capability, priced headroom
Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, set via a model parameter with medium as the default. The published benchmark scores are max-effort numbers, and effort is not free: at the default medium effort, the GDPval-AA score Anthropic shares drops from 1620 to 1277. In other words, the price sheet is only half the cost model — pushing a task to high or max effort consumes more headroom to reach the headline scores. For routine extraction work, medium effort at $0.10/$0.50 is the point; the max-effort numbers describe what the model can do, not what most requests should pay for.
What the benchmarks actually show
The jumps versus Haiku 4.5 are large — with two honest caveats: every number below is Anthropic's own evaluation, and the OSWorld figure is the offline subset of the benchmark (82 tasks run from screenshots, not the live web environment).
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 (Elo) | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1 (Elo) | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1, offline subset | 72.4% | 15.7% | 48.9% | 83.9% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| Humanity's Last Exam (no tools) | 45.9% | 10.2% | not reported | 56.9% |
Anthropic's launch table, as reproduced by The Decoder, DataCamp and others. On every row where both are reported, Haiku 5.5 finishes ahead of GPT-6 Luna — but Sonnet 5.5 still leads on all of them.
Independent verification is thin but directionally consistent: Vals AI's Terminal-Bench 4.0 leaderboard, updated October 7, 2026, puts Haiku 5.5 at 35.35% on its own Mini-SWE-agent harness (12th of 45 models, about $3.05 per test) — below Anthropic's 39.2% claim, but nowhere near Haiku 4.5's zero. The system card, walked through by kingy.ai, also reports the OSWorld strict pass rate at 37.1% versus the 72.4% partial score — a reminder that "score" depends heavily on how the harness grades.
Where it fits, and where it doesn't
- High-volume queues — support-ticket triage, document classification, structured extraction, routing. This is the workload the pricing was designed around, and where the 75% average saving applies most cleanly.
- Subagents — Anthropic explicitly positions it to pair with Opus 5.5 and Sonnet 5.5 as a cheap subagent on coding work. A planner on Sonnet delegating file-churn tasks to Haiku 5.5 is now a defensible cost structure.
- Computer-use loops — OSWorld measures an agent operating a computer from screenshots, a token-hungry workload where a 10x price drop changes what's economical to run at all.
- Not a Sonnet replacement — Sonnet 5.5 beats it on every published row. If your task is currently failing at the small tier, a 90% discount doesn't fix capability gaps.
Availability, guarantees, and the watermark
Haiku 5.5 is available from launch day on the Claude API and platform, Amazon Bedrock, Google Cloud and Microsoft Foundry/Azure, under the model ID claude-haiku-5-5. The 1M-token context window and 128K max output match the docs. Anthropic's model overview commits to keeping it available until not sooner than October 7, 2027 — a one-year floor that matters for anyone locking a pipeline to a cheap tier.
One compliance detail worth knowing: because it launched after August 2, 2026, Haiku 5.5's output carries Anthropic's invisible text watermark — a statistical pattern in token choices that is undetectable to a reader, implemented for EU AI Act provenance rules. Anthropic is explicit that nothing is added to the text and there are no hidden characters; it is a detector-readable property of the generation process, not a metadata layer.
How to decide in ten minutes
If you run a high-volume workload on any model above $0.50 per million output tokens, the test is cheap: move one production queue to Haiku 5.5 at default effort for a day, and compare three numbers — quality sample, tokens billed (not tokens expected), and cost per task. Watch the tokenizer inflation and the 100K cliff specifically, because both inflate the bill in ways a per-token price comparison hides. Anthropic is also rolling out monthly API credits for subscribers ($100 per month for Max 5x, $200 for Max 20x, up to $500 pooled across Team plans), which can offset migration-month bills. The verdict we'd publish: for bulk work under 100K tokens, this is now the default starting point in the Claude lineup — and for everything else, it's the cheapest way to find out your task needed a bigger model.
- Anthropic — Claude Haiku 5.5 model page (pricing, effort setting, benchmarks)
- Claude Platform docs — model overview (1M context, 128K output, retirement date)
- Claude Platform docs — Haiku 5.5 migration guide (tokenizer note, platform IDs)
- Anthropic — How Claude's text watermark works
- AWS Machine Learning Blog — Introducing Claude Haiku 5.5 on AWS
- Simon Willison — Introducing Claude Haiku 5.5 (October 7, 2026)
- The Decoder — Haiku 5.5 arrives with massive price cuts
- VKTR — Anthropic cuts small-model pricing 90% with Claude Haiku 5.5
- Quartz — Anthropic is cutting prices on its smallest AI model by up to 90%
- Vals AI — Terminal-Bench 4.0 leaderboard (independent: 35.35%)
- kingy.ai — Haiku 5.5 system-card walkthrough (OSWorld offline subset, strict pass rates)
- Cryptobriefing — Anthropic launches Claude Haiku 5.5 with 75% lower average costs
Prices and benchmark figures are standard published rates as of October 11, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.