Anthropic launches Claude Haiku 5.5: 75% cheaper than predecessor, OSWorld accuracy jumps from 15.7% to 72.4%

On October 7, 2026, Anthropic officially released the Claude Haiku 5.5 model, with model ID claude-haiku-5-5. Positioned as the company's cheapest, fastest, and most capable small model to date, it targets high-throughput, cost-sensitive tasks including summarization, context compression, database queries, classification requests, and other rapid, repetitive workloads. The model pairs with Opus 5.5 and Sonnet 5.5 as a sub-agent in coding workflows, while its extremely fast response time makes it especially suitable for latency-sensitive use cases like real-time customer support and browser use.

On pricing, Haiku 5.5 averages about 75% cheaper than its predecessor Haiku 4.5. Specifically, requests up to 100,000 tokens are priced 90% lower, while requests over 100,000 tokens are priced 50% lower; since about 90% of Haiku 4.5 requests fall in the former category, the overall cost drop is significant. Per million tokens: cached reads $0.01-$0.05 (depending on prompt length), cache writes $0.125-$0.625, input $0.10-$0.50, output $0.50-$2.50.

In performance benchmarks, Haiku 5.5 shows a major leap over Haiku 4.5. On the GDPval-AA v2.1 measuring real-world professional work, it scores 1620 (vs Haiku 4.5 at 735, Sonnet 5.5 at 1840, GPT-6 Luna at 1437); AA-Briefcase v1.1 scores 1578 (vs Haiku 4.5 at 614); computer use OSWorld 2.1 offline subset accuracy reaches 72.4% (vs Haiku 4.5 at 15.7%); multi-disciplinary Humanity's Last Exam scores 45.9% without tools and 57.4% with tools; agentic coding Terminal-Bench 4.0 reaches 39.2%, FrontierCode 1.1 reaches 46.4%; visual reasoning Chartography reaches 46.4%. Haiku 5.5 is also the first Haiku-tier model with adjustable effort settings, offering Low, Med, High, Xhigh, and Max levels, letting users flexibly trade off cost and intelligence.

Early customer feedback further validates its performance and cost improvements. Asana reports task completion latency reduced by over 30% and per-agent-turn inference speed improved by up to 2.5x; HubSpot achieves 92.8% average on its CRM task evaluation suite (its highest ever); AlphaSense's "Ask in Document" feature is called about 8 million times weekly in production, with score on 400-query test rising from Haiku 4.5's 0.76 to 0.84; Box notes Haiku 5.5 scores 11 points higher than Haiku 4.5 with about half the latency; Rogo uses it as a sub-agent to extract segment revenue data from 10-K reports; Cognition uses it as a "sidekick" to Devin Fusion, keeping Fusion at 66.2 on FrontierCode while reducing cost and latency.

On safety, Haiku 5.5 shows major alignment improvements over Haiku 4.5, with significant reductions in misaligned behaviors and cooperative misuse. Its cyber safeguards are stricter than Haiku 4.5 but slightly below recent other Anthropic models, allowing broader defensive tasks while still blocking penetration testing and other techniques more likely to be used by attackers; its biology safeguards are aligned with Sonnet 5, Sonnet 5.5, and Opus 5, allowing research questions but limiting requests that could cause harm.

Alongside the launch, Anthropic made several other important updates. First, Sonnet 5.5 cached reads were cut in half, from $0.20 to $0.10 per million tokens, reducing Sonnet 5.5 cost on most agentic tasks by about 20%. Second, starting this week, new monthly API credits were introduced for Max and Team subscribers: Max 5x users get $100/month, Max 20x users get $200/month, Team subscribers can get up to $500 pooled. Anthropic also updated the Claude Python and TypeScript SDKs, adding computer use and browser use support in beta. Haiku 5.5 is now live on all platforms including AWS, Google Cloud, and Microsoft Azure, accessible to developers through Claude Platform with the claude-haiku-5-5 model ID.