Anthropic Releases Claude Haiku 5.5: Beats GPT-6 Luna in Benchmarks, Slashes Prices

Anthropic releases Claude Haiku 5.5, and the most direct impact comes from price. For requests with prompts under 100k tokens—90% of Haiku 4.5 request volume—input and output prices are cut by a full 90%; the portion exceeding 100k tokens is cut by only 50%. The previous-generation Haiku 4.5 was released about a year ago and was 10 times the price of Haiku 5.5.

This pricing makes Haiku 5.5, when not exceeding a 100k prompt, cheaper than DeepSeek-V4.1 Flash's off-peak price, except for cache-hit input. But "cheap" does not equal "money-saving": Haiku 5.5 switched to the same tokenizer as Sonnet 5.5 and Opus 5.5, so the same tasks consume more tokens, with potentially higher inflation for code, tables, and non-English content, meaning the money actually saved is less than the list-price discount. In AA evaluations, although Haiku 5.5's API price is cheap, the average cost of completing tasks has not reached the ideal range.

Official benchmarks show Haiku 5.5 beating GPT-6 Luna on every item, becoming the new small-model gatekeeper. This claim comes from vendor self-testing and lacks independent third-party verification, so it should be treated as a source_claim rather than a confirmed fact.

On the specific data: in OSWorld 2.1 (which tests an agent operating a real computer to complete multi-step tasks), Haiku 5.5's Low tier achieves 42.0% accuracy at $0.07 per run, and the Max tier achieves 72.4% at $0.61; while Haiku 4.5 only has a Max tier of 15.7% at $1.45. In other words, 5.5's Low tier is more accurate than 4.5's Max tier and costs half as much. GDPval-AA v2.1 (which tests real professional work across 44 occupations) shows a similar curve: Low tier Elo 1125 at $0.01, Max tier 1620 at $0.87; 4.5's Max tier is only 735 at $0.24.

Haiku 5.5 is also the first Haiku model to support effort adjustment, inheriting the series' five-tier configuration from low to max. The previous-generation Haiku 4.5 basically came up empty on computer operation and coding, while this generation evolves directly.

Many people expected Haiku 5.5 to replace Sonnet 5.5 on some tasks and further lower costs, but no miracle appeared this time. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2% while Sonnet 5.5 scores 70.6%, and the gap is very clear in complex multi-step coding, cross-file refactoring, and long-horizon autonomous planning.

Anthropic itself also recommends prioritizing Sonnet 5.5 or Opus 5.5 for complex agent coding. This means Haiku 5.5's positioning is not to replace Sonnet, but to fill the gap in the execution layer.

Haiku 5.5's value lies in the execution layer, that is, tasks that have already been broken down, have clear acceptance criteria, and can be run in parallel. This judgment is an inference based on its benchmark curves and price structure, not the vendor's own words.

One specific case is Cognition's Devin: using Opus 5.5 as the main model and Haiku 5.5 as the sub-agent, the FrontierCode combination reached 66.2%, higher than either model running alone. However, this is a single case, and whether it can be generalized to other agent scenarios is unknown.

At this point, the division of labor is clear: Opus 5.5 handles complex reasoning, Sonnet 5.5 handles general execution, and Haiku 5.5 handles volume and speed.

If a previous business used Haiku 4.5, this time it is not something that can go live just by changing the model name. Migrating to Haiku 5.5 involves five changes, each of which could cause requests to error out directly or return unexpected structures.

First, the budget_tokens manual thinking configuration errors out directly and must be changed to adaptive thinking plus the effort parameter. Second, temperature, top_p, and top_k are all locked to default values, so applications that rely on sampling parameters for creative control or diversified generation must change their logic. Third, assistant message prefilling is removed, so the previous approach of forcing JSON to start via prefilling must be replaced with tool calls or the structured output interface. Fourth, the computer operation tool version changes from computer_20250124 to computer_toolset_20260801, and both the interface format and return structure may differ. Fifth, adaptive thinking is enabled by default, so the first content block of a response may be a thinking block rather than the main text, and parsing logic must filter by the type field.

Anthropic provides a migration guide and recommends going through it before switching. Old businesses using Haiku 4.5 cannot go live by just changing the model name; they must rework the logic for thinking configuration, sampling parameters, prefilling, and tool versions.

Besides Haiku 5.5 itself, Anthropic also announced two adjustments. Sonnet 5.5's cache reads drop from $0.20/M to $0.10/M, and Anthropic says most agent tasks therefore see costs fall by about 20%.

Max and Team subscription users can claim API credits starting this week: Max 5x gets $100 per month, Max 20x gets $200 per month, and Team gets up to $500 per month shared within the team, usable across all models. These credits can be used to experiment with calling APIs to build tools, applications, and agents.

On the other side, OpenAI has handed out another reset card. The specific content and motivation are not explained, but this move is interpreted by outsiders as a response to Anthropic's pricing offensive.