Claude Haiku 5.5: 90% cheaper under 100K tokens

The cheap model just got a lot cheaper. Anthropic released Claude Haiku 5.5 today at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. That is 90% below Haiku 4.5, and it matches OpenAI's GPT-6 Luna to the cent. If your business runs ticket triage, product tagging, or email summaries on an API, your bill for that work could drop by most of a zero. Read the fine print first.

What actually happened

Per Anthropic's announcement, Haiku 5.5 has two price tiers. Prompts up to 100K tokens cost $0.10 input / $0.50 output per million tokens. Prompts over 100K cost $0.50 / $2.50. Haiku 4.5 was $1.00 / $5.00 flat. Cache reads drop to $0.01 per million in the lower tier.

Anthropic says the average cost to run is about 75% lower, not 90%. Two reasons: roughly 90% of Haiku 4.5 requests fell under 100K tokens, and the new tokenizer "uses slightly more tokens per task." VentureBeat lays the price against the field: GPT-6 Luna at $0.10 / $0.50, Gemini 3.5 Flash-Lite at $0.30 / $2.50.

Other details that matter:

• It is the first Haiku with an adjustable effort setting. The default is medium.
• Model ID claude-haiku-5-5, live on Anthropic's API, AWS, Google Cloud, and Azure.
• Anthropic recommends it for summaries, classification, database queries, live support, and as a subagent under Sonnet 5.5 or Opus 5.5. For complex agentic coding, it points you to the bigger models.
• Sonnet 5.5 cache reads were cut from $0.20 to $0.10 per million tokens.
• Max and Team subscribers get monthly API credits: $100 (Max 5x), $200 (Max 20x), up to $500 pooled for Team.

Why it matters for your business

Do the math on your own traffic. Take a classification job with 2,000 input tokens and 200 output tokens per request, 100,000 requests a month. On Haiku 4.5 that is about $300. On Haiku 5.5 it is about $30, plus whatever the new tokenizer adds. Small numbers, but the same ratio holds at ten times the volume.

Watch the 100K line. If you stuff whole catalogs, contracts, or long chat histories into each prompt, you pay the upper tier. That is still 50% cheaper than Haiku 4.5, not 90%. Trimming context is now worth real money.

Test before you swap. Vendor benchmarks are vendor benchmarks. Run your last 500 real requests through both models, compare outputs, and check the token counts. A price cut that changes your outputs is not a saving.

Jobs you priced out are back in. Tagging every product review, summarizing every support call, scoring every inbound lead. At $0.10 per million input tokens, the model cost on most small-business volumes stops being the line item that kills the idea.

Sources: Anthropic, VentureBeat.