Claude Haiku 5.5: What Changed, What It Costs, Should You Switch?
Answer first
Anthropic released Claude Haiku 5.5 on October 7, 2026: its small, fast tier jumps to a 1M-token context window and drops to $0.10/$0.50 per million tokens for prompts up to 100K — 90% below Haiku 4.5 at that tier. Anthropic's own benchmarks show step-change computer-use and terminal gains; independent scores have not surfaced yet.
What changed vs Claude Haiku 4.5
- Price: $1.00/$5.00 → $0.10/$0.50 per 1M tokens for prompts up to 100K; Anthropic puts the average saving at about 75% (Reuters, Oct 7, 2026). Cache reads fall from $0.10 to $0.01.
- Two-tier pricing is new: prompts above 100K tokens cost $0.50/$2.50 instead. The headline price only applies at the short-prompt tier.
- Context window: 200K → 1M tokens. Max output: 64K → 128K tokens.
- First Haiku with an effort setting: five levels, default medium — the benchmark table's headline terminal score was run at maximum effort.
- Reads text and images; aimed at classification, summarisation, extraction, live support and subtasks for larger agents.
- Available day one via the Anthropic API (model ID
claude-haiku-5-5), AWS, Google Cloud and Microsoft Azure.
The numbers, with trust labels
| Benchmark | Haiku 5.5 | Haiku 4.5 | Source & trust |
|---|---|---|---|
| OSWorld 2.1, offline subset (computer use) | 72.4% | 15.7% | Anthropic launch table — vendor-reported (GPT-6 Luna: 48.9% in the same table) |
| Terminal-Bench 4.0 | 39.2% (max effort) | 0.0% | Anthropic launch table — vendor-reported (Luna 16.4%; Claude Sonnet 5.5 70.6%) |
| GDPval-AA v2.1 (knowledge work, Elo) | 1620 | 735 | Anthropic launch table — vendor-reported (Luna 1437) |
| Humanity's Last Exam, no tools | 45.9% | 10.2% | Anthropic launch table — vendor-reported (57.4% with tools) |
| Artificial Analysis Intelligence Index / LMArena | no score found | — | Independent: no AA or LMArena score for Haiku 5.5 had surfaced in the sources checked as of Oct 23, 2026 |
How to read this table: every performance number here is Anthropic's. The jumps are large enough to be plausible for a generational release — and they are exactly the rows Anthropic chose to publish. Two cautions from Anthropic's own material: the Terminal-Bench figure is a maximum-effort run, not the default setting; and Anthropic itself steers complex agentic coding to Sonnet 5.5 and Opus 5.5. Launch coverage summarising Anthropic's system card also flags the highest over-refusal rate among the models Anthropic tested.
What it costs (per 1M tokens, checked Oct 23, 2026)
| Tier | Input | Output | Cache read | Note |
|---|---|---|---|---|
| Prompts up to 100K tokens | $0.10 | $0.50 | $0.01 | The headline tier — most high-volume jobs land here |
| Prompts above 100K tokens | $0.50 | $2.50 | $0.05 | 5× the short-prompt rate |
| Haiku 4.5 (reference) | $1.00 | $5.00 | $0.10 | The model it replaces |
Runs on what?
- API only, closed weights. Anthropic API plus AWS, Google Cloud and Microsoft Azure, under the model ID
claude-haiku-5-5. - Text and image input, text output; 1M-token context, up to 128K output tokens.
- The effort dial is a cost control: five effort levels from low to max, default medium. Anthropic's headline agentic scores use max effort, which costs more time and tokens — your default-setting results will differ.
Should you switch? By situation
- You run Haiku 4.5 for classification, extraction or summarisation at volume → Switch after a spot-check. The price drop at the ≤100K tier is mechanical (a tenth of the input price), and the reported capability gains are in exactly these task families.
- You dropped small models from agent loops because they were the weak link → Pilot it. Computer-use and terminal scores are the story of this release; test at the effort level you will actually pay for.
- Your workload is complex agentic coding → Stay on Sonnet/Opus. Anthropic's own table puts Sonnet 5.5 at 70.6% on Terminal-Bench 4.0 against Haiku's 39.2%, and Anthropic points demanding coding agents upward.
- Your prompts routinely exceed 100K tokens → Model the second tier first. At $0.50/$2.50 the saving versus Haiku 4.5 is still about half, not 90%.
FAQ
How much cheaper is Claude Haiku 5.5, really?
At the short-prompt tier (up to 100K tokens), input and output prices are a tenth of Haiku 4.5's — $0.10/$0.50 against $1.00/$5.00 per million tokens. Anthropic's own average figure is about 75% cheaper overall, because longer prompts fall into the higher tier and a new tokenizer changes token counts.
What is the effort setting?
Haiku 5.5 is the first Haiku with an adjustable effort setting: five levels controlling how long the model thinks before answering, with medium as the default. Higher effort buys capability on hard tasks at higher latency and token cost. Anthropic's headline Terminal-Bench score of 39.2% was measured at maximum effort.
Is Claude Haiku 5.5 good at coding?
Better than any previous Haiku on Anthropic's numbers, but it is not Anthropic's coding recommendation. Anthropic's launch table shows Sonnet 5.5 far ahead on Terminal-Bench 4.0 (70.6% vs 39.2%), and Anthropic directs complex coding agents to Sonnet and Opus. Haiku's coding role is subtasks inside a larger agent.
Are there independent benchmarks for Haiku 5.5?
None that we could find as of October 23, 2026: no Artificial Analysis Intelligence Index score and no LMArena placement had surfaced in the sources checked. Every score on this card is Anthropic-reported, which is why the recommendation is to verify on your own workload before switching wholesale.
Where can I use Claude Haiku 5.5?
Through the Anthropic API (model ID claude-haiku-5-5) and, from launch day, on Amazon Web Services, Google Cloud and Microsoft Azure. It is the small-model tier of the Claude 5.5 family, alongside Sonnet 5.5 and Opus 5.5.
What we could not verify
- Independent benchmark scores: no Artificial Analysis or LMArena figures found as of Oct 23, 2026
- Whether Anthropic ran GPT-6 Luna itself for its comparison table or used OpenAI's published figures — Anthropic does not say (noted by DataNorth's launch analysis)
- Launch-partner latency claims (e.g. Asana's "over 30% latency reduction"): customer quotes published by Anthropic, not independently tested
- System-card weakness details (over-refusal, honesty under pressure) reach us via secondary summaries of Anthropic's system card, not the card itself
Sources: Anthropic announcement and system-card coverage (Reuters, Oct 7, 2026; DataNorth and AI Weekly launch analyses; VeraxonAI pricing/spec summary — checked Oct 23, 2026); Puter developer model listing. Vendor-reported figures are labelled as such wherever they appear on this card.