Claude Haiku 5.5 Cuts Average Costs About 75 Percent

Anthropic is pushing its small-model tier into agent workflows that used to be too expensive to run at volume. In an October 7 announcement, the company said Claude Haiku 5.5 is its cheapest, fastest and most capable Haiku yet, with average costs about 75% lower than Haiku 4.5.

What happened

For prompts up to 100,000 tokens, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, compared with $1 and $5 for Haiku 4.5, Anthropic said. Above that threshold, prices rise to $0.50 and $2.50. Roughly 90% of prior Haiku requests fell in the cheaper band. Crypto Briefing relayed the same schedule and noted Haiku 5.5 is the first Haiku with an adjustable effort setting.

Benchmark tables in the Anthropic post show large jumps versus Haiku 4.5: 72.4% versus 15.7% on the offline subset of OSWorld 2.1, and 39.2% on Terminal-Bench 4.0 where Sonnet 5.5 still leads at 70.6%. Customer quotes from Asana, HubSpot, AlphaSense, Box and Cognition describe lower latency and higher hit rates on triage, CRM audits, document question-answering and coding sidekick roles. Alongside the launch, Anthropic halved Sonnet 5.5 cache-read pricing to $0.10 per million tokens, estimating about 20% lower cost on many agentic jobs, and said Max and Team subscribers would receive monthly Claude Platform API credits.

Why it matters

Price and speed at the small-model layer change which tasks get automated. Anthropic positions Claude Haiku 5.5 for summaries, classification, database queries, live support and browser use, and as a subagent beside Opus 5.5 and Sonnet 5.5 on coding. That packaging matters for companies already managing token bills at millions of calls per week, as AlphaSense's quote implied. It also arrives while peer labs jockey on evaluation leaderboards we tracked in the Arena valuation story and while Microsoft talks about emergency brakes on model autonomy in our Nadella emergency brake report.

Pricing tables tell the adoption story faster than benchmark charts. At $0.10/$0.50 per million tokens under 100k, Claude Haiku 5.5 undercuts prior Haiku enough that workloads once reserved for rules engines or smaller open models become economical on Anthropic's stack. Crypto Briefing underlined that Sonnet remains the better coding lead while Haiku takes narrowly scoped jobs—exactly the division of labour enterprises already use with human teams. Cache-read cuts on Sonnet 5.5 compound the saving on agent loops that reread long contexts.

Safety notes in the Anthropic post matter for procurement. Haiku 5.5 carries tighter cyber safeguards than Haiku 4.5 in some dimensions while remaining less restrictive than Sonnet on defensive tasks, and it shares biology safeguards with recent larger models. That mix will be read closely by firms in regulated verticals. The monthly API credits for Max and Team seats are a go-to-market shove: lower the cost of experimentation and more internal tools get built. Whether those tools stay on Claude Haiku 5.5 after credits lapse is the retention test.

Competition from other small models will police Anthropic's pricing. The launch post compared Haiku 5.5 with GPT-6 Luna on several benchmarks, claiming leads on OSWorld and knowledge-work suites while still trailing Sonnet on hard agentic coding. Buyers will run their own evals. The strategic bet behind Claude Haiku 5.5 is that a fast, cheap, effort-tunable model becomes the default subagent layer in coding products such as Devin Fusion, where Cognition reported a FrontierCode score of 66.2 with Haiku as sidekick. If those designs spread, token mix shifts toward Haiku even when the "lead" model stays large.

Enterprises that adopt Claude Haiku 5.5 for high-volume paths still need Sonnet or Opus for hard coding and long-horizon plans; the win is a cheaper mix, not a single-model monopoly.

Documentation updates in the Claude Platform SDKs for computer use and browser use are aimed squarely at Haiku's speed profile. A model that is cheaper and faster is the natural driver for UI-level agents that click, type and summarise in real time. Claude Haiku 5.5 succeeds if those beta paths move into production without blowing token budgets.

What's next

the new Haiku model is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure, Anthropic said, with beta computer-use and browser-use support landing in Python and TypeScript SDKs. Watch whether enterprise eval suites reproduce the early customer gains and how the new Max/Team API credits change experimentation. This article is for information only and is not investment advice.