Claude Haiku 5.5 Costs $0.10 Per Million Input Tokens, Then Writes 440 Million Tokens to Finish One Benchmark

Anthropic's Claude Haiku 5.5 costs $0.10 per million input tokens, but Artificial Analysis found it wrote 440 million tokens on its Intelligence Index, the most of any of the 60 models in that comparison.

Anthropic released Haiku 5.5 on October 7 and announced it with the API model ID claude-haiku-5-5. It is available on Amazon Web Services, Google Cloud and Microsoft Azure. The announcement does not state a context window; Artificial Analysis lists 1 million tokens.

Haiku 5.5 price list against Haiku 4.5

Anthropic prices prompts up to 100,000 tokens at $0.10 input and $0.50 output per million tokens. Longer prompts cost $0.50 input and $2.50 output. Anthropic says Haiku 5.5 costs about 75% less to run on average than Haiku 4.5.

Model	Input per 1M tokens	Output per 1M tokens
Haiku 5.5 (prompts up to 100K)	$0.10	$0.50
Haiku 4.5	$1.00	$5.00
Sonnet 5.5	$2.00	$10.00

The per-token cut from Haiku 4.5 is 90%, not 75%. Anthropic's post does not say what the 75% measures.

What Artificial Analysis found

Artificial Analysis scores Haiku 5.5 (Max) at 43 on its Intelligence Index, second of 176 models, against a median of 13. It lists output speed at 237.3 tokens per second, rank 10 of 148.

Two numbers cut the other way. Time to first token is 336.77 seconds against a median of 1.84. And a full run of the index costs about $0.21 per task, rank 43 of 58, even though the per-token price sits well under the class medians of $0.25 input and $0.87 output.

The cause is volume: 440 million output tokens on the index against a median of 95 million. The page shows the reasoning variant and says a non-reasoning variant may also exist, so the 336-second delay may not apply to every setting.

Anthropic's own benchmarks

Anthropic published the figures below and points to its system card for method. The post credits Artificial Analysis only for GDPval-AA v2.1. These are vendor-supplied scores, and the post names no evaluator for the rest.

1. Sonnet 5.583.9 %
2. Haiku 5.572.4 %
3. GPT-6 Luna48.9 %
4. Haiku 4.515.7 %

Source: Anthropic Claude Haiku 5.5 announcement, accessed 2026-10-14

On Terminal-Bench 4.0, an agentic coding test, Haiku 5.5 scores 39.2% against 16.4% for OpenAI's GPT-6 Luna and 70.6% for Sonnet 5.5. On FrontierCode 1.1 it scores 46.4%, to 42.4% for GPT-6 Luna and 52.1% for Sonnet 5.5. On Humanity's Last Exam it scores 45.9% without tools and 57.4% with them.

Anthropic also cites customer-run tests, including a HubSpot CRM suite that Haiku 5.5 passed at an average of 92.8% over three runs. The post carries testimonials from named engineers, among them Aaron Vinh, Staff Software Engineer at Asana, and Ze'ev Klapow, Distinguished Software Engineer at HubSpot.

Haiku 5.5 safety limits

Anthropic says Haiku 5.5's cybersecurity safeguards are stricter than Haiku 4.5's but looser than Sonnet 5.5's, and still block penetration testing and similar attacker-oriented techniques. Its biology safeguards match Sonnet 5, Sonnet 5.5 and Opus 5. The post gives no formal safety level.

Token volume is a cost for Anthropic too: it moved its own monorepo off Git when agent-driven merges outgrew the old tool. For how a rival priced a launch with a time-limited discount, see Gemini 4 Argon's $2 and $10 pricing.

Artificial Analysis lists the reasoning variant only, so no independent token count exists yet for a non-reasoning run, as of Wednesday. Anthropic says its system card holds the evaluation methods for the benchmarks above.