# Claude Haiku 5.5 Cuts Small-Model Prices to One Tenth of Haiku 4.5 and Jumps From 15.7% to 72.4% on OSWorld

At $0.10 per million input tokens and $0.50 per million output tokens with a one-million-token context, Anthropic's new small model costs less than Meta's Muse Spark 1.1, which was priced to undercut the closed labs.

- Published: 2026-10-11T06:16:05.908Z
- Canonical: https://polylog.news/ai/2026-10-11/claude-haiku-5-5-cuts-small-model-prices-to-one-tenth-of-hai
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic News](https://www.anthropic.com/claude-haiku-5-5), [Anthropic News](https://www.anthropic.com/claude-sonnet-5-5), [Meta AI](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/)

Anthropic released [Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) on October 7, nine days after [Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5). The release completes the 5.5 family and replaces Haiku 4.5, the model at the center of this week's Philadelphia incident. The price is [$0.10 per million input tokens and $0.50 per million output tokens](https://openrouter.ai/anthropic/claude-haiku-5.5) for prompts up to 100,000 tokens. Above that length the price rises to $0.50 and $2.50. Cache reads start at $0.01 per million tokens, and batch requests cost half. That is one tenth of Haiku 4.5's list price, and Anthropic says real workloads come out about 75% cheaper on average. Context grows from 200,000 to [one million tokens](https://thenewstack.io/anthropic-claude-haiku-5-5/), with output up to 128,000 tokens and a June 2026 knowledge cutoff.

Anthropic's own benchmark numbers show the largest gains on agentic tasks. On the offline subset of OSWorld 2.1, the score rises from 15.7% to 72.4%. On Terminal-Bench 4.0, Haiku 5.5 scores [39.2%, compared with 0% for Haiku 4.5, 16.4% for OpenAI's GPT-6 Luna and 70.6% for Sonnet 5.5](https://computingforgeeks.com/claude-haiku-5-5-released-features-benchmarks/). It scores 64.8% on SWE-Bench Pro. On SWE-bench Multilingual it scores 83.7%, against Haiku 4.5's 67.4%, a smaller gain than the agentic tests show. These numbers come from the vendor. The one independent figure so far is Artificial Analysis's [Intelligence Index score of 43 at maximum effort](https://openrouter.ai/anthropic/claude-haiku-5.5), second of 182 models in its price class.

The price cut puts downward pressure on the rest of the market. Sonnet 5.5 already lists at [$2 input and $10 output](https://anotherwrapper.com/tools/llm-pricing/claude-sonnet-5-5) and runs about 30% faster than Sonnet 5. Meta entered paid APIs with [Muse Spark 1.1 at $1.25 and $4.25](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/), a price [reported as roughly a quarter](https://the-decoder.com/metas-muse-spark-1-1-api-pricing-squeezes-openai-and-anthropic-as-the-ai-price-war-heats-up/) of what OpenAI and Anthropic charged at the time. Haiku 5.5 now costs less than Muse Spark on both input and output for prompts under 100,000 tokens. Practitioners should note that the price rises fivefold above that length, so the advertised long context carries a separate price curve.

## What this means

The exposure runs through subagent economics. Orchestrators that send summarization, browsing and tool calls to a small model now get computer-use scores in the 70% range at about a tenth of the old cost. The losers are mid-tier closed models and hosted open-weight providers whose main argument was price. A closed lab cutting its own small-model price tenfold suggests that the efficiency gains are real, or that Anthropic is buying agent workload share at a loss, and its public numbers cannot tell those two apart. Many observers will credit the OSWorld score, but scores from an offline subset of a vendor-run benchmark often fall once an independent party reruns them. The likelier outcome is a matching small-model price cut from OpenAI or Google within weeks, followed by vendors switching default subagent models to Haiku 5.5. The forward read fails if independent OSWorld and Terminal-Bench reruns land far below Anthropic's figures, because that would make the cut a pricing move with no capability behind it.

## What to watch

- Independent reruns of OSWorld 2.1 and Terminal-Bench 4.0 on Haiku 5.5. These show whether the agentic gains hold outside Anthropic's own harness.
- The next small-model price from OpenAI and Google. A match below $0.50 per million output tokens would confirm that the small-model tier now competes on price.
- Whether coding-agent products such as Cursor and Claude Code make Haiku 5.5 their default subagent. Such a switch would turn the pricing into real usage.