Anthropic releases Haiku 5.5 with reported tenfold lower API prices

• Short-prompt API prices fell 90%: Anthropic's pricing thread lists $0.10 input and $0.50 output per million tokens in the cheaper tier.
• Computer-use performance reached 72.4% on OSWorld 2.1's offline subset, the result cited in aakashgupta's post, versus 15.7% for Haiku 4.5.
• Website design is an early creative use case: viktoroddy posted a demo attributed to Haiku 5.5.

Anthropic's October 7 launch footnote discloses a new tokenizer that uses more tokens per piece of work. An independent SVG test stretched past five minutes at the highest effort setting and still cost just over three cents.

The tenfold discount promoted in the launch post applies to prompts up to 100,000 tokens. Anthropic's pricing table gives these rates per million tokens:

• Up to 100K prompt tokens: $0.10 input, $0.50 output, $0.01 cache reads.
• Over 100K prompt tokens: $0.50 input, $2.50 output, $0.05 cache reads.

Haiku 4.5 charged $1 input, $5 output, and $0.10 for cache reads. The new tiers therefore cut token prices by 90% and 50%, respectively.

Anthropic says roughly 90% of Haiku 4.5 requests fit the cheaper tier, but estimates an average cost reduction of around 75% per completed piece of work, accounting for changed token consumption.

Simon Willison, who runs the Claude Token Counter tool, counted about 25% more tokens for the same long prompt with Haiku 5.5 than with Haiku 4.5.

Willison used the prompt “Generate an SVG of a pelican riding a bicycle” in his hands-on tests, covering five effort settings:

• low: seven seconds, costing 0.0936 cents.
• medium: the default setting.
• high.
• xhigh.
• max: five minutes, nine seconds, costing 3.3826 cents.

Reasoning could not be disabled. Willison reported a good bicycle frame at every setting above low.

Viktor Oddy posted a website-design demo that he attributed to Haiku 5.5, following a request to design a website.

Anthropic's launch evaluations separate computer operation from command-line coding:

• OSWorld 2.1, offline subset: Haiku 4.5's 15.7% → Haiku 5.5's 72.4%, +56.7 points. GPT-6 Luna scored 48.9%.
• Terminal-Bench 4.0: Haiku 4.5's 0.0% → Haiku 5.5's 39.2%, +39.2 points. Sonnet 5.5 scored 70.6%.

OSWorld measures completing long, multi-step tasks on a real computer. Terminal-Bench measures professional tasks within a command-line interface.

Haiku 5.5 is available in Claude Platform and Claude Code, according to Anthropic's developer thread. Anthropic positions it as a helper underneath Opus 5.5 or Sonnet 5.5 for:

• Summaries.
• Compactions.
• Database queries.

One example in the launch announcement has a larger model building a presentation while a Haiku subagent extracts the segment-revenue line from a company's 10-K filing. Cheap research helpers are the creative prize here.

The Gray Swan test pictured below measures attackers' success across repeated attempts, with lower scores indicating stronger resistance. All models used extended thinking, and Claude models were tested without protections specific to prompt injection.

The system card also reports a regression: Haiku 5.5 over-refused more than any other model tested in Anthropic's automated behavioral audit.

Anthropic is rolling out monthly Claude Platform credits alongside the release:

• Max 5x: $100.
• Max 20x: $200.
• Team: $20 per Standard seat and $100 per Premium seat, pooled up to $500.

Under the Help Center terms, new subscribers qualify after seven days, credits go to one linked Console organization, and unused balances expire each billing cycle. Free, Pro, and Enterprise plans are excluded.

Interactive Claude Code sessions and Claude usage through third-party cloud providers are excluded. Self-run claude -p calls qualify when billed as Agent SDK usage with an API key from the linked organization; runs started by the Claude Code GitHub Action remain ineligible.