最新模型
Anthropic Launches Claude Haiku 5.5, Adding Reasoning Effort Controls and Tiered Long-Prompt Pricing to Haiku for the First Time
Anthropic released Claude Haiku 5.5 on October 7, with adjustable reasoning effort and lower per-token prices for short prompts. AWS is offering the model as well, but long-prompt pricing, cross-region routing, and the cost of real-world tasks still need separate evaluation.
Anthropic launched Claude Haiku 5.5 on October 7, positioning it for classification, summarization, context compression, and narrowly scoped agent subtasks. It is the first Haiku model to offer effort controls, letting developers adjust reasoning investment to the task. Its model ID on Claude Platform is claude-haiku-5-5. Official announcement
Pricing is tiered by prompt length. For prompts of up to 100,000 tokens, input and output cost $0.10 and $0.50 per million tokens, respectively. Above that threshold, the rates are $0.50 and $2.50. The per-token price for short prompts is 90% lower than Haiku 4.5, but Anthropic estimates that average task costs will fall by about 75%. It also notes that the new tokenizer may use slightly more tokens for the same work. Migration assessments should therefore compare the full task bill, not just per-token rates. Pricing and computation details
The AWS model card lists a one-million-token context window and a maximum output of 128,000 tokens. The model accepts text and image inputs and generates text. Adaptive thinking is enabled by default, with effort set to medium; five levels are available, from low to max. When thinking is turned off, the maximum effort level is high. These settings let the same model serve different latency budgets, but teams should measure whether higher effort is worthwhile using their own data. AWS model card
Capability figures are still primarily based on vendor evaluations. Anthropic reports a Terminal-Bench 4.0 score of 39.2%, below Sonnet 5.5 at 70.6%. Its OSWorld 2.1 score is 72.4%, but that result is limited to an offline subset. These results support using Haiku for discrete tasks with verifiable outcomes, but do not show that it can take over every long-running workflow. Engineering teams can start by testing document-field extraction, small code changes, and summaries of tool results, while tracking retries, error rates, and completion times. Benchmark table
Deployment options also differ. AWS now offers the model through Bedrock and Claude Platform on AWS. Bedrock supports Messages, Invoke, and Converse calls, with access through geographic or global cross-region inference configurations. Teams should confirm the routing scope and location requirements, then compare success rates, tail latency, and cost per successful task at different effort levels to determine whether lower prices translate into real savings. AWS deployment guide