Haiku 5.5 is 75% cheaper and now actually good at browser tasks. Small-model economics just changed.

Anthropic released Haiku 5.5 on October 7. The headline is 75% cheaper on average versus Haiku 4.5, which is already notable. But the number I keep coming back to is the OSWorld score.

Haiku 4.5 scored 15.7% on OSWorld 2.1's offline subset. That's a score that says: you can try to use this model in a browser-automation loop, but most of the time it will fail. Haiku 5.5 scores 72.4%. That is not an incremental improvement. That is a different product category.

What the benchmarks actually say

All numbers from Anthropic's launch post:

• OSWorld 2.1 (offline subset): 72.4% vs 15.7% for Haiku 4.5
• Terminal-Bench 4.0: 39.2% vs 0.0% for Haiku 4.5 — the model couldn't do terminal tasks at all before
• Humanity's Last Exam (no tools): 45.9% vs 10.2%
• GDPval-AA v2.1: 1620 vs 735 (Anthropic's internal agentic eval suite)

Pricing structure worth understanding carefully. Two tiers based on request length:

• Under 100K tokens: $0.10 per million input, $0.50 per million output — 90% cheaper than Haiku 4.5
• Over 100K tokens: $0.50 per million input, $2.50 per million output — 50% cheaper than Haiku 4.5

The 75% average reduction Anthropic headlines is a blended figure across both tiers. If most of your calls are under 100K tokens, the actual savings are closer to 90%.

The effort level thing

First Haiku-class model with adjustable effort. That is the feature I will actually use day-to-day.

If you've been building multi-agent workflows, you've probably been doing one of two things: either using Haiku everywhere because it's cheap (and accepting a high failure rate on harder sub-tasks), or using Sonnet everywhere and absorbing the cost. Haiku 5.5 gives you a third option: a Haiku you can turn up.

Anthropic's positioning is explicit — Haiku 5.5 as a subagent alongside Opus 5.5 and Sonnet 5.5 orchestrators. The model handles repetitive sub-tasks (summarization, classification, form-filling, short reasoning) at $0.10/M while the orchestrator runs on a more expensive model for planning.

One genuine gotcha: the tokenizer is updated. Haiku 5.5 uses slightly more tokens per equivalent input than Haiku 4.5 did. Anthropic doesn't publish a specific overhead percentage in the release post. Worth running your representative inputs through the tokenizer before assuming the 90% cost cut is fully realized. For most workloads the overhead is small. For very tight latency/cost budgets, measure it.

What I think the OSWorld number actually means

72.4% on OSWorld 2.1's offline subset. Let me be precise about what that is and isn't.

OSWorld is a well-established computer-use benchmark — the standard in the field for testing whether a model can navigate a GUI, fill forms, open applications, respond to screen state. The "offline subset" removes network-dependent tasks (tasks that require a live internet connection to complete). Full OSWorld numbers are harder, and Anthropic doesn't publish those.

So 72.4% means: on the tasks OSWorld can evaluate without a live internet connection, Haiku 5.5 succeeds almost 3 out of 4 times. That's actually usable. Haiku 4.5's 15.7% was not.

Or maybe I'm being too optimistic. The offline subset is easier than the real-world online version. Tasks that require dynamic web content, login flows, or navigation across pages might perform differently. Before I redesign a system around this benchmark, I'd run it against my actual browser-automation use cases.

The Terminal-Bench number is the one I find more decisive. Terminal-Bench 4.0: 39.2% vs 0.0%. Zero. Haiku 4.5 literally could not do terminal tasks. This isn't a regression from 30% to 39% — this is a model that gained a capability class it didn't have before. For CLI automation, scripting, and server-side agent work, that matters.

Source spread

• Anthropic — Claude Haiku 5.5 — hype (own announcement). All benchmark numbers and pricing sourced here. Benchmarks are Anthropic-reported, not independently reproduced.
• vals.ai — Claude Haiku 5.5 model card — builder. Third-party listing confirming the October 7 release date and basic specs.

Pros & cons

What's genuinely useful:

• 90% cost reduction for sub-100K requests is a primary-source number. For high-volume classification, summarization, or short-context pipelines, this changes the economics.
• OSWorld 72.4% makes browser-use workflows viable at Haiku pricing — previously you needed Sonnet or higher for reliable computer-use.
• First effort levels on Haiku means per-request cost control without model-swapping.
• Available everywhere on day one: AWS, GCP, Azure, Claude Platform. No access gotcha.
• Cache reads at $0.01/M under 100K. Repeated system-prompt patterns are nearly free to cache.

What to watch:

• Updated tokenizer adds overhead. Measure before migrating.
• GDPval-AA and AA-Briefcase are Anthropic-internal benchmarks. No independent reproduction yet.
• OSWorld offline subset ≠ full OSWorld. Network-dependent browser tasks may perform differently.
• The 50% cost cut for over-100K tokens is real but smaller than the headline number. Long-context workloads need to do the math separately.
• Cybersecurity safeguards are tighter on Haiku 5.5 than Haiku 4.5. If you were doing security research or offensive tooling with the prior model, expect more restrictions.

Further reading

• Anthropic — Claude Haiku 5.5 launch post — all specs, pricing, and benchmarks
• OSWorld benchmark — the computer-use benchmark methodology and leaderboard
• Anthropic model overview — full family comparison