Claude Haiku 5.5: a spectacular price drop for Anthropic’s small models

A launch focused on volume and latency

On October 7 2026, Anthropic announced Claude Haiku 5.5 as the direct successor to Haiku 4.5, positioning it as the cheapest and quickest model in the small‑model category. Designed for high‑volume, latency‑sensitive tasks—such as summarisation, dataset compression, database queries, classification, customer‑support assistance, or browser automation—Haiku 5.5 aims to replace heavier solutions in contexts where execution speed outweighs deep reasoning capabilities. The company highlighted that the model’s reduced inference time makes it especially suitable for real‑time applications where every millisecond counts.

A dramatically larger context window

Compared with Haiku 4.5, which offered a 200 000‑token input window and a 64 000‑token output window, Haiku 5.5 expands those limits to one million input tokens and 128 000 output tokens. This enlarged capacity enables the processing of much longer documents without the need for chunking, a feature that is particularly valuable for summarising massive reports or executing complex queries over large databases. The model also incorporates a configurable parameter called “effort”, whose default setting is “Medium”. Anthropic’s published benchmark scores were obtained in “maximal‑effort” mode, indicating that performance can vary depending on the chosen setting.

Two‑tier pricing: how the savings break down

Anthropic proposes a rate of $0.10 per million input tokens and $0.50 per million output tokens for requests under 100 000 tokens. Beyond that threshold, the prices are multiplied by five, yet they remain on average at half the cost of Haiku 4.5 according to the company’s statements. Anthropic claims that these new rates make Haiku 5.5 roughly 75 % cheaper to run than its predecessor. By contrast, 01net reports a 90 % reduction for short queries and a 50 % reduction for longer ones, creating a slight discrepancy with Anthropic’s official pricing description.

Measured performance on benchmarks

Benchmark results released by Anthropic show substantial gains. On Terminal‑Bench 4.0, Haiku 5.5 achieves 39.2 % versus 0 % for Haiku 4.5. The GDPval‑AA v2.1 score rises from 735 to 1 620, OSWorld 2.1 climbs from 15.7 % to 72.4 %, and Chartography improves from 6.4 % to 46.4 %. All these scores were obtained in maximal‑effort mode; in medium mode, the Terminal‑Bench 4.0 score drops to roughly 20.3 %. Compared with Sonnet 5.5, Haiku 5.5 still lags on FrontierCode 1.1 (46.4 % versus 52.1 %) and on Terminal‑Bench 4.0 (39.2 % versus 70.6 %).

Safeguards and usage restrictions

Anthropic indicates that Haiku 5.5’s safeguards are stricter than those of Haiku 4.5 but less restrictive than those of Sonnet 5.5. The model permits defensive‑type tasks while continuing to block intrusion attempts or penetration‑testing activities. This intermediate stance is intended to provide greater freedom for commercial applications while maintaining a security level compatible with compliance requirements.

Concrete implications for developers

For teams integrating Anthropic’s API, the price drop translates into an immediate reduction in operational costs, especially if the majority of requests stay below the 100 000‑token threshold—which Anthropic says accounts for 90 % of Haiku 4.5 traffic. Additionally, the company announced that reading from the cache for Sonnet 5.5 will cost half as much, benefiting agents that heavily rely on cached data. New monthly API credits ranging from $100 to $500 are also offered to Max and Team subscription tiers, providing extra flexibility for high‑usage projects.

Open questions that remain

Several uncertainties persist. The exact per‑token price for Haiku 4.5 has not been disclosed, complicating precise cost comparisons. The real impact of the “effort” parameter on expenses has not been independently verified, and the benchmarks supplied come solely from Anthropic. No third‑party tests have yet confirmed the advertised performance, particularly against OpenAI’s GPT‑6 Luna, whose price multiplier only kicks in after 272 000 tokens, compared with Haiku 5.5’s 100 000‑token breakpoint.

What changes right now

In summary, Claude Haiku 5.5 introduces a substantial cost reduction for high‑volume workloads while offering a dramatically larger context window and markedly better performance than Haiku 4.5. Developers can expect significant savings on short queries, provided they stay under the 100 000‑token limit, and productivity gains from the new browser‑and‑computer control capabilities delivered via the Python and TypeScript SDKs. However, the two‑tier pricing model, the configurable effort setting, and the lack of independent benchmark verification all call for heightened caution before making a large‑scale migration to this new model.