Anthropic has released Claude Haiku 5.5, targeting high-concurrency and cost-sensitive scenarios, and it is also the company's fastest-responding lightweight model to date.
The official positioning is very clear: it mainly handles data filtering, content summarization, and classification retrieval, or serves as a sub-agent for Opus 5.5 and Sonnet 5.5 in complex tasks. Notably, the new model is the first to add an adjustable thinking-effort setting, allowing a trade-off between cost and intelligence based on task complexity.
However, I think that with so many model tiers, and each tier adding a different degree of thinking-capability adjustment, the mental cost for users in actual use is very high.
This round of iteration is all about price cuts, and Haiku is no exception.
For short prompt requests under 100K tokens, Haiku 5.5 is 90% cheaper than the previous generation, at $0.1 per million input tokens and $0.5 per million output tokens, with cache hits at half price—just $0.01—and overall operating costs down about 75% on average.
Haiku 5.5's pricing is clearly aimed at GPT-6 Luna, and its cache-hit price is even lower. As a result, among models of comparable capability, Haiku 5.5 has become the new king of cost-effectiveness. DeepSeek's cost-effectiveness is being gradually eroded, and Haiku 5.5's price is already far lower than 4.1 Flash.
In benchmarks, Haiku 5.5 not only comprehensively widens the gap with the previous generation, but also beats GPT-6 Luna on some metrics:
- Computer operation test OSWorld 2.1 offline subset score: 72.4%, versus only 15.7% for the previous generation
- Professional knowledge evaluation GDPval-AA: 1620 points, notably higher than Luna's 1437
- Agent coding evaluation Terminal-Bench 4.0: 39.2%, a huge improvement over 4.5's 0%
On the supporting side, the company has added beta browser-operation support to its Python and TypeScript SDKs, and is offering Max and Team subscribers $100 to $500 per month in API credits.
As of now, Anthropic is still considerably more generous than OpenAI, and OpenAI's recent wave of moves has really damaged its reputation.