NewsAI

Anthropic Launches Claude Haiku 5.5: 75% Cheaper on Average, First Haiku with Effort Controls, Sonnet 5.5 Cache Reads Halved

Anthropic released Claude Haiku 5.5 on October 7, positioning it as the cheapest, fastest, and most capable small model the company has ever shipped. It is built for high-volume, cost-sensitive workloads, the model costs around 75% less to run on average than Haiku 4.5, and it is the first Haiku-class model to ship with an adjustable effort setting. Anthropic also cut Sonnet 5.5 cache-read pricing in half, lowering the cost of most agentic work by roughly 20%.

Key takeaways

  • Pricing: $0.10 input / $0.50 output per million tokens for prompts up to 100K, 90% cheaper than Haiku 4.5
  • Benchmark leap: GDPval-AA 1620 (from 735), Terminal-Bench 4.0 at 39.2% (from 0%), OSWorld 2.1 at 72.4% (from 15.7%)
  • First Haiku-class model with an effort setting from Low to Max, letting developers trade cost against intelligence
  • Sonnet 5.5 cache reads drop from $0.20 to $0.10 per million tokens
  • Monthly API credits for subscribers: Max 5x gets $100, Max 20x gets $200, Team up to $500 pooled

Benchmark performance: a small-model leap

The numbers show a wide gap over its predecessor. On knowledge work, Haiku 5.5 scores 1620 on GDPval-AA v2.1, versus 735 for Haiku 4.5, 1437 for GPT-6 Luna, and 1840 for Sonnet 5.5, and 1578 on AA-Briefcase v1.1 (from 614). In computer use, it reaches 72.4% on the OSWorld 2.1 offline subset, well ahead of Haiku 4.5’s 15.7% and GPT-6 Luna’s 48.9%. Agentic coding jumps from 0.0% to 39.2% on Terminal-Bench 4.0, and the model’s 46.4% on FrontierCode 1.1 edges out GPT-6 Luna’s 42.4%. Multidisciplinary reasoning improves to 45.9% without tools and 57.4% with tools on Humanity’s Last Exam, while visual reasoning rises to 46.4% on Chartography, from 6.4%.

Effort settings: cost control built in

Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, ranging from Low to Max. Developers choose where to sit on the cost-intelligence curve for each task, and Anthropic published accuracy-versus-cost curves for three benchmarks at every setting. The company’s positioning is explicit: Sonnet 5.5 and Opus 5.5 remain the better picks for complex agentic coding of the kind Terminal-Bench 4.0 measures, while Haiku 5.5 targets narrowly scoped, high-volume work such as compaction, summarization, and subagent tasks that would often have been cost-prohibitive on previous versions.

Pricing: 90% cheaper under 100K tokens

Pricing splits at 100,000 prompt tokens. Below that line, requests cost $0.10 per million input tokens and $0.50 per million output tokens, with cache reads at $0.01, 90% below Haiku 4.5. Above it, pricing moves to $0.50 / $2.50, still 50% cheaper. About 90% of requests to the previous Haiku fell under the lower tier, which is how Anthropic arrives at the ~75% average saving. One caveat: the model ships a new tokenizer shared with Sonnet 5.5 and Opus 5.5 and uses slightly more tokens per task, a factor the company says its average already accounts for.

Early customers: fast and accurate

Early testers reported consistent gains. Asana measured over a 30% reduction in task-completion latency and up to 2.5x faster inference per agent turn. HubSpot’s CRM simulations produced the best score it has recorded on that suite, 92.8% averaged over three runs, with the fastest completion times and the lowest false-positive rate. AlphaSense, running about 8 million production calls a week through its Ask in Document feature, saw scores rise from 0.76 to 0.84 across 400 queries. Box reported an 11-point quality gain at roughly half the latency. Cognition added Haiku 5.5 to the sidekick lineup in Devin Fusion, holding a top-tier FrontierCode score of 66.2 while cutting cost and latency.

Price cuts across the range

Alongside the launch, Anthropic halved Sonnet 5.5 cache-read pricing from $0.20 to $0.10 per million tokens; because cache reads account for a large share of token consumption, most agentic tasks on Sonnet 5.5 become roughly 20% cheaper. Starting this week, subscribers also receive monthly API credits for the Claude Platform: $100 for Max 5x users, $200 for Max 20x, and up to $500 pooled across Team subscribers, usable on any model. For developers, the Claude Python and TypeScript SDKs gain beta support for computer use and browser use, tasks Anthropic says suit Haiku 5.5’s combination of speed, capability, and price.

On safety, Anthropic reports major improvements across nearly all of its alignment evaluations relative to Haiku 4.5, with fewer misaligned behaviors and a lower willingness to cooperate with misuse. Cybersecurity safeguards are more restrictive than Haiku 4.5’s: defensive tasks are permitted more broadly than under Sonnet 5.5’s safeguards, but penetration testing and similar attack-leaning techniques remain blocked. Biology safeguards match those of Sonnet 5, Sonnet 5.5, and Opus 5, with wider access available through the Life Sciences and Cyber Verification Programs.

Claude Haiku 5.5 is available today across AWS, Google Cloud, and Microsoft Azure, and on the Claude Platform as claude-haiku-5-5, with a migration guide for existing deployments.

Sources: Anthropic Blog, Haiku 5.5 System Card