Anthropic announced Claude Haiku 5.5, its cheapest, fastest, and most capable small model yet, costing around 75% less to run than Haiku 4.5. Designed for high-volume, cost-sensitive tasks, it includes an adjustable effort setting and is available on AWS, Google Cloud, and Microsoft Azure. Anthropic also halved the price of cache reads on Claude Sonnet 5.5 to $0.10 per million tokens.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Anthropic announced the launch of its small model Claude Haiku 5.5 via its official X account, calling it its "cheapest, fastest, strongest" small model, with an average running cost about 75% lower than Haiku 4.5, and introducing an adjustable effort setting to the Haiku series for the first time; it also halved the cache read price of Sonnet 5.5 to $0.10 per million tokens. All key figures come from the vendor's own statements, with no third-party verification.
Anthropic's official Claude X account published a post announcing the launch of Claude Haiku 5.5 and simultaneously lowering the cache read price of Claude Sonnet 5.5. This post is the sole source of all information in this article, and it does not include benchmark names, specific scores, or evaluation methods.
The post describes Haiku 5.5 as the "cheapest, fastest, strongest" small model Anthropic has ever released, and claims its average running cost is about 75% lower than Claude Haiku 4.5. The same post also announced halving the cache read price of Sonnet 5.5 to $0.10 per million tokens, and claims this reduces Sonnet 5.5's running cost on most long-running work by about 20%.
It must be made clear that the above cost, capability, and alignment figures are all Anthropic's own statements, constituting vendor claims rather than independently confirmed facts. The post itself provides no verifiable comparison basis, nor any third-party source for cross-validation.
The cost claim for Haiku 5.5 rests on two pillars. The first is the reduction relative to the previous generation: Anthropic claims its average running cost is about 75% lower than Haiku 4.5. The second is the task distribution: Anthropic claims tasks under 100K tokens account for about 90% of its previous-generation Haiku model's request volume, and Haiku 5.5 offers particularly outstanding value on such tasks.
Together, these two figures form a narrative that "most requests fall in the most cost-effective range." But the source does not explain what task mix and billing basis the 75% reduction is based on, nor does it explain the time range and statistical basis for the 90% request volume figure. Therefore, these two figures can only be understood as vendor claims, not as reproducible cost conclusions.
The post also claims Haiku 5.5 is designed for high-concurrency, cost-sensitive tasks and can reliably handle repetitive work such as summarization and classification. This positioning is consistent with its identity as a small model, but it likewise provides no specific data on throughput, latency, or concurrency scale.
Anthropic claims Haiku 5.5 is its first Haiku model with an adjustable effort setting, letting users trade off between cost and intelligence based on the task. This is one of the few points in this release that represents a product-form change rather than a pure performance claim.
However, the source does not explain the specific levels, value range, or default behavior of the adjustable effort. This means "trading off based on the task" is currently only a directional description, and users cannot judge from this post how much practical room for operation there actually is.
Inferring from product logic (this is an inference), the adjustable effort works in tandem with the "high-concurrency, cost-sensitive" positioning: the same model can handle repetitive work such as summarization and classification at low effort, and take on more complex tasks at high effort. But this inference is not supported by level or pricing information in the source, and can only be regarded as an interpretation of the launch intent.
Anthropic claims Haiku 5.5 shows significant improvement over Haiku 4.5 in coding, computer use, and knowledge work, and claims its speed is sufficient to support real-time customer support and browser-use scenarios. The post also claims it is suitable for use as a subagent alongside Claude Opus 5.5 and Sonnet 5.5 in coding work.
On alignment, Anthropic claims Haiku 5.5 shows substantial improvement over Haiku 4.5 on almost all alignment evaluations, with a marked reduction in instances of misaligned behavior.
These statements all point in the same direction—better than the previous generation—but the source provides no benchmark names, specific scores, or evaluation methods, nor does it disclose the specific evaluation sets, metrics, and magnitudes for the alignment evaluations. Therefore, "significant improvement" and "substantial improvement" are currently unverifiable vendor claims, and readers should not equate them with confirmed capability levels.
Anthropic claims Haiku 5.5 is now available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. But the source does not specify particular regions or service forms, nor does it provide specifications such as Haiku 5.5's pricing table, context window, or latency.
The Sonnet 5.5 cache read price cut also has a scope problem: the source does not explain whether this price is valid long-term or whether it applies to all platforms. The post claims that after the price cut, Sonnet 5.5's running cost on most long-running work is reduced by about 20%, but the very term "most long-running work" is itself not further defined.
For teams that rely on cache read costs for budgeting, these gaps mean the actual impact of the price cut will require more complete pricing documentation before it can be assessed.
What can be confirmed: Anthropic's official account released Haiku 5.5 and simultaneously announced a cache read price cut for Sonnet 5.5. This is the act of release itself, which is a confirmed fact.
What cannot be confirmed: the 75% cost reduction, the significant improvement in coding and knowledge work, the substantial improvement in alignment evaluations, the claim that tasks under 100K tokens account for 90% of request volume, and the roughly 20% cost reduction for Sonnet 5.5. These are all Anthropic's own statements, with no third-party independent verification and no supporting benchmark scores or evaluation methods.
In addition, the actual effectiveness of Haiku 5.5 collaborating with Opus 5.5 and Sonnet 5.5 as subagents likewise has no independent evidence. The post shows about 95.5K views, 343 reposts, and 3K likes; these are X platform counts, reflecting dissemination heat rather than technical verification.
All information in this article comes from one launch post on Anthropic's official X account, with no third-party or independent source for cross-validation. Key figures on cost, capability, and alignment are all vendor claims, with no benchmark names, specific scores, or evaluation methods given; "cheapest, fastest, strongest" is a marketing expression with no verifiable comparison basis provided. The levels and default behavior of the adjustable effort, Haiku 5.5's pricing table and context window, and the validity period and applicable scope of the Sonnet 5.5 price cut are all not explained in the source.
The launch of Haiku 5.5 conveys a clear product direction: use a small model plus adjustable effort to cover high-concurrency, cost-sensitive long-tail tasks, and use the Sonnet 5.5 cache price cut to lower the cost of long-running work. But this official post only gives direction, not verifiable figures. Until third-party evaluations or complete pricing documentation appear, the 75% reduction, capability improvements, and alignment improvements should be regarded as vendor claims, not confirmed facts.
Primary report
Primary source