Claude Haiku 5.5: Pricing, Free Access, and Benchmark Tests

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.

👁 127 reads

Continuously updated. Status as of October 8, 2026, 11:00 AM.

Claude Haiku 5.5 is the most affordable and fastest model in the Claude 5.5 family, succeeding Haiku 4.5 from October 15, 2025. Anthropic launched it on October 7, 2026, priced at $0.10 per million input tokens and $0.50 per million output tokens - a 90% discount compared to Haiku 4.5 and matching GPT-6 Luna. In our Polish-language writing evaluation, it scored 76.8 out of 100, compared to 59.2 for Haiku 4.5 and 82.8 for GPT-6 Luna.

What is Claude Haiku 5.5?

Haiku represents Anthropic’s most lightweight line of models, engineered for massive volume, ultra-fast responses, and cost-sensitive workloads such as classification, summarization, customer support routing, and sub-agents within Claude Code. Anthropic pre-announced Haiku 5.5 twice: on September 22 alongside Opus 5.5 and on September 28 with Sonnet 5.5. The official October 7 release designates it as the cheapest, fastest, and most capable small model Anthropic has ever shipped.

Key updates over Haiku 4.5 include:

• 90% lower pricing for requests under 100k tokens, and 50% lower above that threshold. Anthropic notes that sub-100k prompts account for roughly 90% of legacy Haiku traffic.
• Adaptive thinking with an effort parameter (low, medium, high, xhigh, max; default is medium) - appearing in the Haiku family for the first time.
• 1M token context window with up to 128k output tokens, trained on data through June 2026.
• New tokenizer, shared with Sonnet and Opus 5.5. In our measurements, the same Polish test text took 341 tokens on Haiku 5.5 versus 254 on Haiku 4.5 (a 34% increase).
• Stricter cybersecurity safeguards - VentureBeat reported that standard safety filters now strictly block activities like penetration testing.

Haiku took the longest to see an architectural refresh. Haiku 5 was skipped entirely, despite speculative claims on X pointing to late August and early September launches. Rumors in September even suggested the Haiku line would be phased out - which Anthropic conclusively dismissed with the 5.5 generation announcements.

How to access Claude Haiku 5.5

1. In the Claude app (claude.ai, iOS, Android) - choose Haiku 5.5 from the model picker above the chat bar. Anthropic provides access on all plans, including the Free tier.
2. In Claude Code - run /model and pick Haiku 5.5 (requires Claude Code version 2.1.293 or higher, plus an active paid plan or configured API billing). Settings and rate limits are covered in our guide on accessing Haiku 5.5 for free and in Claude Code.
3. Via the Anthropic API - grab an API key from platform.claude.com (guide in our walkthrough on getting an API key) and call claude-haiku-5-5. Note that non-default values for temperature, top_p, and top_k trigger a 400 error, continuing the policy introduced in models since 4.7.
4. Through major cloud providers - Amazon Bedrock (anthropic.claude-haiku-5-5), Google Cloud, and Microsoft Foundry (claude-haiku-5-5), live from launch day.
5. Via OpenRouter - under anthropic/claude-haiku-5.5, routing through Anthropic, Claude Platform on AWS, Amazon Bedrock, Google, or Azure. We conducted our benchmark tests this way - see our overview of OpenRouter.

If you have legacy pipelines running Haiku 4.5, Anthropic’s deprecation board marked it on October 8 as “active” with support guaranteed through “not earlier than October 15, 2026.” Because Anthropic provides at least 60 days of advance notice before shutting down endpoints, Haiku 4.5 will remain operational through at least December even if retired shortly.

Claude Haiku 5.5 vs Haiku 4.5, Sonnet 5.5, and Opus 5.5

Vendor benchmarks cited in Anthropic’s release documentation:

The most dramatic leap is in autonomous agent environments: on computer operation workflows, Haiku 5.5 sits closer to Sonnet 5.5 than to GPT-6 Luna. However, for terminal coding, it still lags noticeably behind Sonnet (39.2% versus 70.6%).

Claude Haiku 5.5 API pricing

All prices reflect direct vendor list rates without tax. Cache writes on Haiku 5.5 cost $0.125 ($0.625 over 100k tokens). Concurrently, Anthropic dropped cache reads on Sonnet 5.5 from $0.20 to $0.10 and introduced monthly API credits for subscribers: $100 for Max 5x, $200 for Max 20x, and up to $500 per team in the Team tier. Gemini 3.8 Flash rates are locked through December 31, 2026.

Two key dynamics may make your real-world bill higher than headline rates suggest:

• Tokenizer differences on non-English text. In our AI pricing index, running 1,000 standard non-English test queries cost approximately $0.81 on Haiku 5.5, $0.50 on GPT-6 Luna, and $5.98 on Haiku 4.5. Even with matching per-token rates, Haiku can end up pricier than Luna depending on language token density.
• High effort settings produce long reasoning traces. Artificial Analysis measured that at “max” effort, Haiku 5.5 outputs around 162k reasoning tokens per run, roughly three times more than Luna, driving the average run cost to $0.21 versus $0.07 for Luna.

Anthropic itself estimates that blended real-world costs for Haiku 5.5 will average roughly 75% less than Haiku 4.5, rather than the pure 90% headline cut.

Claude subscription tiers

Subscribers to Claude web and mobile access get Haiku 5.5 included across all tiers:

Full subscription features and plan options are detailed in our guide to Claude pricing and plans.

Claude Haiku 5.5 performance in multilingual tests

Budget models historically struggle with complex non-English phrasing. In the source author’s previous Polish-language benchmark tests, Haiku 4.5 scored a modest 59.2 out of 100. Haiku 5.5 improved markedly by 17.6 points to 76.8, although it did not overtake GPT-6 Luna (82.8), Sonnet 5.5 (80.4), or Gemini 3.8 Flash (79.4). However, it comfortably surpassed DeepSeek V4.1 Flash (72.9) and Mistral Large 4 (69.2).

Key qualitative observations from our test outputs (we tested these prompts in Polish; translated here for reference):

• Strong performance: Client emails, professional LinkedIn posts, and short creative narratives earned 8.2 to 8.3 out of 10 from automated evaluation judges. Phrasing flowed naturally without unnatural translation artifacts; em-dashes appeared at a rate of 5.7 per 1,000 words, close to Opus 5.5 (5.1) and lower than Luna (7.7).
• Weaker areas: 5-point concise summarization scored just 5.8 points. The model exceeded an 80-word ceiling and produced an awkward grammatical construction (roughly translated as “about what reported the warehouse manager”), flagged by both independent evaluation judges.
• Complex proofreading: 26 out of 30 errors corrected, up from 23/30 on Haiku 4.5, while Gemini 3.8 Flash scored a perfect 30/30.
• Complex regional knowledge quiz: 43 out of 48 correct (Haiku 4.5: 40/48; Luna: 48/48). It successfully caught all trick questions containing false premises (such as hypothetical awards never granted).
• Negative constraint adherence: Satisfied 23 out of 26 strict instructions, displaying a tendency toward verbosity when tight length caps are imposed.

Polish-language benchmark: Haiku 5.5 tested

The source author evaluated the model against 13 Polish-language benchmark prompts via OpenRouter, with reasoning effort set to “low” and uniform token constraints applied. Automated scoring was handled by external models (GPT-6 Sol and Gemini 3.1 Pro) calibrated against four fixed anchor models (Opus 5.5, Luna, Qwen3.8 Max, GLM-5.3). The test was completed on October 8, 2026. Side-by-side prompt responses are cataloged in our AI Model Arena.

Costs reflect direct OpenRouter billing. Regarding throughput: across all 13 prompts, Haiku 5.5 generated an average of 164 tokens per second (including reasoning tokens), compared to 101 for Haiku 4.5 and 98 for Sonnet 5.5. (We restrict direct speed comparisons to Anthropic models due to differing tokenizer designs across labs). At the “low” effort setting, Haiku 5.5 thought more thoroughly than its predecessor, generating 12.4k reasoning tokens across the suite versus 7.2k on Haiku 4.5 - yet delivered a 6x lower overall run cost thanks to the 10x base price reduction.

Verdict: for standalone copy generation in non-English languages, GPT-6 Luna still offers stronger quality per dollar spent. Haiku 5.5 shines when you are already committed to Anthropic’s ecosystem, deploying autonomous agents, or needing stronger grounding (scoring a 40% hallucination rate on AA-Omniscience versus 77% for Luna).

Comparison with competing budget models

Pricing data sourced from OpenRouter as of October 8, 2026; Artificial Analysis Index figures are taken from their launch coverage. As of October 8, neither GPT-6.1 Luna nor “Sol mini” had been released - OpenRouter lists the September 22 version of GPT-6 Luna. See our full results in the multilingual AI model leaderboard, or check our hands-on breakdown in Haiku 5.5 vs GPT-6 Luna vs Gemini 3.8 Flash.

Pre-launch rumors versus reality

• September 22, 2026 (Opus 5.5 launch) & Mike Krieger on X: Sonnet 5.5 and Haiku 5.5 were promised “in the coming weeks.” This proved accurate: Haiku arrived 15 days later.
• September 23, 2026, @Mr_Salio on X: Claimed Haiku 5.5 would offer “Opus 5.5 parity.” In reality, Haiku hits 43 on the AA Index versus Opus’s 58.
• September 2026 pricing leaks: Rumored output pricing around $1.25 per million tokens. The official price came in much lower at $0.50.
• September 29, 2026, Polymarket: 94% implied probability of a new Haiku by October 15. Correct.
• Timeline extrapolation: Historical patterns suggested October 14 (16 days post-Sonnet, mirroring the 4.5 cycle). Anthropic shipped a week ahead of that schedule.

Who is Claude Haiku 5.5 for?

• Recommended: Engineering teams already using Anthropic’s API - dropping in Haiku 5.5 for Haiku 4.5 slashes costs by 75-90% with zero architectural migration.
• Recommended: Claude Code sub-agents, browser workflows, computer use, data extraction, and routing - where benchmarks significantly beat Luna.
• Recommended: Free-tier Claude users who gain access to a fast model with a 1M context window.
• Worth testing: High-volume localized multilingual copy - GPT-6 Luna remains cheaper and more accurate in our synthetic testing.
• Not recommended: Complex software engineering and tasks with rigid output format constraints - Sonnet 5.5 remains the superior choice here.

Frequently asked questions

When was Claude Haiku 5.5 released?

October 7, 2026. Anthropic had teased its arrival since September 22. Initial launch impressions are documented in our report on the Claude Haiku 5.5 90% price drop.

What does Claude Haiku 5.5 cost?

Direct API rates are $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100k tokens. Prompts exceeding 100k tokens cost $0.50 and $2.50 respectively. Batch processing receives a 50% discount.

Is Claude Haiku 5.5 free to use?

Yes, through the Claude app. Anthropic has enabled Haiku 5.5 across Free, Pro, Max, Team, and Enterprise accounts on the web, iOS, and Android. Free-tier message limits vary dynamically based on system load. Direct API usage has no free tier.

Should you pick Haiku 5.5 or Sonnet 5.5?

Sonnet 5.5 is 20 times more expensive ($2 / $10 per million tokens) but considerably more powerful: scoring 80.4 in our testing and 70.6% on Terminal-Bench versus Haiku’s 39.2%. Choose Sonnet for mission-critical code and demanding reasoning; use Haiku for high-volume operational tasks.

How does Haiku 5.5 compare to GPT-6 Luna?

Token list prices are identical. In non-English language testing, Luna scored higher (82.8 vs 76.8) and consumed fewer tokens. However, Haiku holds clear leads on agentic tasks (OSWorld 72.4% vs 48.9%) and overall AA Index ranking (43 vs 38).

Where is Haiku 5.5 available?

Globally. The Claude web app, mobile apps, and developer API are operational without novel regional blocks. Direct access through OpenRouter and major cloud hyperscalers worked worldwide on day one.

What is happening to Claude Haiku 4.5?

As of October 8, 2026, Haiku 4.5 remains active, with Anthropic guaranteeing availability until “not earlier than October 15, 2026.” Anthropic provides a minimum 60-day deprecation window once retirement is officially declared.

Why was there no Claude Haiku 5?

Anthropic never provided an official explanation. Following Haiku 4.5, the company chose to align version numbering straight to 5.5 to match Opus and Sonnet 5.5.

Kinetyra