The price war, in one paragraph
On October 7, 2026, Anthropic released Claude Haiku 5.5 — the third model in the Claude 5.5 family, after Opus 5.5 (September 22) and Sonnet 5.5 (September 28). Its headline move: pricing per 1M tokens that exactly matches OpenAI's GPT-6 Luna short-context rates — $0.10 input / $0.50 output. Small-model competition has officially become a price war.
Head-to-head: pricing
Haiku 5.5 charges $0.10 input / $0.50 output for prompts up to 100K tokens (cache reads just $0.01), rising to $0.50 / $2.50 above 100K tokens. Anthropic claims it's about 75% cheaper on average than Haiku 4.5, and its launch page says 90% lower than Haiku 4.5 for requests up to 100K tokens — vendor claims, not independently verified. Luna matches the $0.10/$0.50 short tier, but its higher tier only kicks in above 272K input tokens ($0.20/$0.75). That detail matters: for a 150K-token prompt, Luna is actually cheaper on list price.
| Metric | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Short-context price (per 1M) | $0.10 in / $0.50 out | $0.10 in / $0.50 out |
| Short tier limit | 100K input tokens | 272K input tokens |
| Over-limit price | $0.50 / $2.50 | $0.20 / $0.75 |
| Context window | 1M tokens | See official docs |
| Max output | 128K tokens | See official docs |
Head-to-head: performance
Anthropic's system-card benchmarks (vendor-run, not independently verified) put Haiku 5.5 ahead of Luna on several agentic tests: OSWorld 2.1 offline 72.4% vs Luna's 48.9% (Haiku 4.5 scored just 15.7%); Terminal-Bench 4.0 39.2% vs Luna's 16.4% (Haiku 4.5: 0%); FrontierCode 1.1 46.4% vs Luna's 42.4%. On Humanity's Last Exam it scored 45.9% without tools and 57.4% with tools. If you trust the vendor numbers, Haiku 5.5 is the stronger small agent — especially for computer-use and coding-adjacent tasks.
Head-to-head: features
Haiku 5.5 ships a 1M token context window with 128K max output, and it's the first Haiku-class model with adaptive thinking on by default — effort defaulting to medium with an adjustable effort setting. Model ID: claude-haiku-5-5. Knowledge cutoff is June 2026. It takes text and images in, text out. It's available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.
Migration gotchas (from Anthropic's own guide)
If you're moving from Haiku 4.5, three traps will throw errors: manually setting budget_tokens returns a 400 error; setting non-default temperature/top_p/top_k returns a 400; and the new tokenizer counts the same text as roughly 30% more tokens — so re-estimate your costs before assuming the price drop applies to your current spend.
- Remove manual budget_tokens settings — they now return a 400 error.
- Leave temperature/top_p/top_k at defaults — non-default values return a 400.
- Re-estimate token counts — expect ~30% more tokens per request.
- Check prompt routing if you run mixed context lengths — the 100K-token price cliff is steep.
What else came in the same announcement
Anthropic also halved Sonnet 5.5 cache reads ($0.20 → $0.10 per 1M tokens) and added monthly API credits for Max/Team subscribers: $100/month for Max 5x, $200/month for Max 20x, and up to $500 pooled for Team plans.
Which one should you pick?
For short prompts where price is all that matters, they're identical at $0.10/$0.50. For mid-size contexts (100K–272K tokens), Luna is cheaper on list price. For agentic benchmarks, Haiku 5.5 leads on Anthropic's reported numbers. The honest answer: match the model to your task — Haiku 5.5 for cheap agents and subagents in the Claude ecosystem, Luna if you want a longer cheap tier or live in OpenAI's stack.
Frequently Asked Questions
Is Claude Haiku 5.5 really 90% cheaper than Haiku 4.5?
That's Anthropic's launch-page claim for requests up to 100K tokens (about 75% cheaper on average across all lengths). It's a vendor claim, not independently verified — and remember the new tokenizer counts ~30% more tokens, which eats into the saving.
Why would I pick Haiku 5.5 over GPT-6 Luna at the same price?
On Anthropic's reported benchmarks, Haiku 5.5 leads Luna on agent-style tests like OSWorld 2.1 (72.4% vs 48.9%) and Terminal-Bench 4.0 (39.2% vs 16.4%), and it offers a 1M-token context window. Luna's advantage is a higher cheap-tier ceiling (272K vs 100K input tokens).
What breaks when migrating from Haiku 4.5?
Three things, per Anthropic's migration guide: manual budget_tokens settings return a 400 error, non-default temperature/top_p/top_k return a 400, and the new tokenizer counts the same text as ~30% more tokens, so your costs won't drop as far as the headline price suggests.
Is Haiku 5.5 good for coding?
Not for demanding agentic coding, per Anthropic's own positioning. It's aimed at high-volume, cost-sensitive work — summaries, classification, live support, voice agents — and as a subagent under Opus 5.5 or Sonnet 5.5.
Where can I access Haiku 5.5?
The Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The model ID is claude-haiku-5-5, and its knowledge cutoff is June 2026.
Related Prompts
Vector Fashion Editorial Illustration Portrait Editable AI Prompt
ChatGPT
Street Dance Fashion Illustration Editorial Editable AI Prompt
ChatGPT
Illustrated Athlete Giant Sneaker Mustard Poster Ad Editable AI Prompt
ChatGPT
