Claude Haiku 5.5 Drops Price by 90%: Specs, Cost and Tests
Anthropic launched Claude Haiku 5.5 with a 90% price cut to $0.10/$0.50 per million tokens. Here is a look at benchmark scores, costs, and hands-on test results.
Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.
👁 115 reads
Anthropic released Claude Haiku 5.5 on October 7, 2026, cutting API pricing by 90% down to $0.10 per million input tokens and $0.50 per million output tokens-matching GPT-6 Luna. I immediately put the model through 3 practical prompt tests: arithmetic and proofreading passed cleanly, with just one small omission in a creative text task. A comprehensive guide with a 13-task benchmark is coming together in our Claude Haiku 5.5 hub.
Status as of October 8, 2026
- Model: Claude Haiku 5.5, API identifier
claude-haiku-5-5, on OpenRouter asanthropic/claude-haiku-5.5. - API Pricing: $0.10 / $0.50 per 1M tokens (prompts up to 100k tokens), $0.50 / $2.50 above 100k tokens.
- Context Window: 1M tokens, up to 128k output tokens, training cutoff June 2026.
- Multilingual Evaluation: 11 out of 12 points across my initial quick hands-on test.
- Availability: Claude web and mobile apps (including Free tier), API, Claude Code, and OpenRouter.
Claude Haiku 5.5 Features and Benchmarks
This is the third release in Anthropic’s 5.5 series, following Opus (September 22) and Sonnet (September 28). Anthropic positions it as the cheapest, fastest, and most capable compact model in its lineup, targeting high-volume workflows: summarization, classification, sub-agents, and browser automation. For the first time, Haiku features adaptive reasoning with an effort parameter (ranging from low to max, defaulting to medium).
Official performance figures reported by Anthropic:
- OSWorld 2.1 (computer use): 72.4% compared to 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna.
- Terminal-Bench 4.0: 39.2% compared to 0% for Haiku 4.5 and 16.4% for Luna; Sonnet 5.5 stands at 70.6%.
- Humanity’s Last Exam (without tools): 45.9% (Haiku 4.5: 10.2%, Sonnet 5.5: 56.9%).
Independent evaluation from Artificial Analysis scored Haiku 5.5 at 43 points on its quality index (Haiku 4.5 scored 17, Luna 38, Gemini 3.8 Flash 41, Sonnet 5.5 56). However, there is a caveat: at maximum effort, Haiku 5.5 generates roughly 162k output tokens per task-about three times as many as Luna-bringing its average task cost to $0.21 versus $0.07 for Luna. A lower token price does not automatically guarantee a cheaper overall bill.
Claude Haiku 5.5 Pricing
| Per 1M Tokens | Haiku 5.5 (up to 100k) | Haiku 5.5 (above 100k) | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|---|
| Input | $0.10 | $0.50 | $1 | $2 |
| Output | $0.50 | $2.50 | $5 | $10 |
| Prompt Cache Read | $0.01 | $0.05 | $0.10 | $0.10 |
| Prompt Cache Write | $0.125 | $0.625 | $1.25 | $2.50 |
Batch mode provides an additional 50% discount. Anthropic estimates real-world savings will average around 75% compared to Haiku 4.5 because the updated tokenizer (shared with Sonnet and Opus 5.5) consumes slightly more tokens for identical inputs. Alongside this release, prompt cache read pricing on Sonnet 5.5 dropped from $0.20 to $0.10. My initial evaluation run cost just $0.001 in API credits.
Hands-On Quick Test
I tested the model across three standardized prompt tasks using the low effort setting via OpenRouter. Note that we tested it in Polish to evaluate cross-lingual fidelity; the prompts and model outputs below are faithful English translations of the original Polish test material.
| Task | Score | Response Time | API Cost |
|---|---|---|---|
| Text with idiom and inflections | 2/3 - Idiom natural, city name included only once | 4.2 s | $0.00022 |
| Tax and invoice calculation | 3/3 - Accurate breakdown and formatting | 3.9 s | $0.00017 |
| 6-error proofreading | 6/6 - Plus one unnecessary correction | 8.4 s | $0.00061 |
Idiomatic Writing
[Translated from the Polish response] In the end, energy bills surged, and the company came out like Zablocki on soap, losing both capital and the trust of its partners.
The idiom was applied correctly and naturally. However, the instructions required mentioning the city name Bielsko-Biala at least twice across different inflected cases, but the model included it only once. Automated grading awarded 3/3, but I deducted a point for the missed constraint.
Tax Math
The calculation was completely accurate, formatted exactly as requested, and delivered within 4 seconds with no issues.
Proofreading
[Translated from the Polish response] REVISED: Yesterday I went to the store to buy bread and butter. I do not know if I will make it before closing, because the store closes at six. The cashier said that one must always keep the receipt.
All 6 intentional spelling and grammar errors were resolved. However, the model also altered an acceptable colloquial phrasing (“they close the store” to “the store closes”), treating it as a subject-verb agreement issue. That reflects over-correction rather than a severe error. It generated 755 reasoning tokens during this task-the highest compute consumption across the three trials.
Global Availability
- Claude Apps (web, iOS, Android) - Available across Free, Pro, Max, Team, and Enterprise plans, with model selection enabled.
- Claude Code - Supported (requires an eligible paid subscription or API billing).
- Anthropic API - Identifier
claude-haiku-5-5; scheduled for retirement no earlier than October 7, 2027. - Cloud Platforms - Amazon Bedrock, Google Cloud, and Microsoft Foundry.
- OpenRouter - Available globally, verified on October 8, 2026.
- Max and Team Plans - Includes API usage allowances: $100 for Max 5x, $200 for Max 20x, and up to $500 monthly pool for Team plans.
Should You Switch?
- Recommended: Workflows already built on the Anthropic API-migrating from Haiku 4.5 yields up to a 90% reduction in base token rates for short context queries.
- Recommended: Sub-agents, classification, document summaries, and computer-use automation, where benchmark gains over GPT-6 Luna are significant.
- Caution: High effort reasoning modes significantly increase output length; evaluate actual token volume against your specific task pipelines rather than relying solely on the per-token price card.
Further Reading
- Claude Haiku 5.5 - Comprehensive Guide and 13-Task Benchmark
- Claude Sonnet 5.5 and Claude Pricing Breakdown
- AI Cost Comparison Guide
Sources
- Anthropic: Introducing Claude Haiku 5.5 - Pricing, benchmarks, plan allowances
- Anthropic: Models Overview - Model IDs, context limits, deprecation schedule
- Anthropic: Claude Haiku - Plan availability
- Artificial Analysis: Claude Haiku 5.5 - Index score, token consumption analysis
- OpenRouter: Claude Haiku 5.5
Verified on October 8, 2026.
