Claude Haiku 5.5 lands at #27 in Agent Arena on a budget

Anthropic’s newest small model just took a seat on one of AI’s more demanding leaderboards. Claude Haiku 5.5 (High) ranks #27 overall in arena.ai’s Agent Arena, and it gets there while costing less per task than the models that came before it.

What Anthropic shipped

Claude Haiku 5.5 was released on October 7, 2026. Anthropic pitches it as the cheapest, fastest and most capable small model in the Claude 5.5 family.

That family has a clear pecking order. Opus 5.5 and Sonnet 5.5 are the heavier models. Haiku 5.5 is the lighter sibling aimed at high-volume, cost-sensitive jobs.

Anthropic targets the model at agentic work. That covers database queries, coding workflows, automation and customer support.

The biggest structural upgrade is memory. Haiku 5.5 carries a 1M-token context window, up from the 200k window on Haiku 4.5.

The model also supports browser tools and subagent delegation. Subagent delegation lets one model hand pieces of a task to other model instances.

The leaderboard and the price tag

Agent Arena evaluates models on real-world, long-horizon tasks. Those are multi-step jobs where a model has to stay on track over many actions, not just ace a single prompt.

Haiku 5.5 (High) placed #27 there as of mid-October 2026. It also ranks #30 in arena.ai’s Code Arena in the WebDev category.

API costs start as low as $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100k tokens. The pricing is tiered, so those figures represent the entry level for shorter prompts. Larger prompts land in different tiers.

Average running costs are approximately 75% lower than Haiku 4.5.

The model’s cost-performance profile is described as Pareto-competitive. You can usually buy more capability, but you’ll pay for it.

Benchmark gains over the last generation

On GDPval-AA v2.1, Haiku 5.5 scored 1620 points, compared with 735. That benchmark is geared toward economically useful work.

On OSWorld 2.1 offline, the model hit 72.4%, versus 15.7%. OSWorld tests whether a model can actually operate a computer environment, clicking, navigating and completing tasks the way a human user would.

Where you can use it

Distribution is broad. Haiku 5.5 is available on Claude.ai across all plans, through the Anthropic API, and via Amazon Bedrock, Google Cloud and Microsoft Foundry.

Why a #27 ranking is still a big deal

A model that cracks the top 30 on long-horizon agent tasks while cutting average costs by approximately 75% versus its predecessor fits a clear need. It gives companies a credible option for the high-volume layer of their AI stack.

Teams can route difficult reasoning to Opus 5.5 or Sonnet 5.5, then hand repetitive, high-volume work to Haiku 5.5. Subagent delegation makes that kind of division of labor more practical inside a single workflow.

The tiered pricing deserves a close read. The headline rates apply to prompts under 100k tokens, and workloads that lean on the full 1M-token window will see different numbers.