Claude Haiku 5.5 for Computer Use: Scores, Prices, Limits
Vorec Team · 2026-10-23 · About 8 min read
Claude Haiku 5.5, released by Anthropic on 7 October 2026, is a small, fast model that Anthropic positions for high-volume work and for speed-sensitive tasks including browser use. Anthropic reports 72.4% on the offline subset of OSWorld 2.1, a benchmark of operating a real computer, against 15.7% for Haiku 4.5 and 83.9% for Sonnet 5.5. Input tokens cost $0.10 per million for prompts up to 100,000 tokens, against $1.00 for Haiku 4.5.
That combination, a much higher computer-use score than the previous small model at a fraction of its token price, is why the release matters beyond chat. Agents that click through software are expensive to run repeatedly, and a cheaper model that can do part of that work changes which jobs are worth automating. This article lays out what Anthropic published, what the numbers do and don't tell you, and what it means for one specific use: agents that operate an app so it can be recorded and documented.
Everything below comes from Anthropic's Haiku 5.5 announcement and the Claude Platform docs, checked 23 October 2026. All benchmark figures are Anthropic's own. Model line-ups change quickly, so treat the comparisons as a snapshot of that date.
Anthropic-reported, 7 October 2026. OSWorld 2.1 (offline subset): Haiku 5.5 72.4%, Haiku 4.5 15.7%, GPT-6 Luna 48.9%, Sonnet 5.5 83.9%. Input price per million tokens (prompts up to 100k): Haiku 5.5 $0.10, Haiku 4.5 $1.00, Sonnet 5.5 $2.00. Model ID: `claude-haiku-5-5`.
What Anthropic released
The announcement describes Haiku 5.5 as designed for high-volume, cost-sensitive tasks such as summaries, compaction, database queries and classification, and as a subagent alongside Opus 5.5 and Sonnet 5.5 on coding work. Because Anthropic calls it its fastest model to date, it says the model works especially well for speed-sensitive tasks like live customer support and browser use. A footnote qualifies that: it is the fastest at each model's standard speed, but runs less quickly than Opus models in Fast Mode.
Three other details matter for anyone building agents:
- Adjustable effort. Haiku 5.5 is the first Haiku-class model with an effort setting, so you can trade cost against capability per request.
- SDK support for computer and browser use. Alongside the launch, Anthropic updated its Python and TypeScript SDKs to support computer use and browser use in beta, and says Haiku 5.5 is especially well suited to those tasks. We covered those SDK toolsets separately.
- Availability. Anthropic lists it on all platforms, including AWS, Google Cloud and Microsoft Azure, and on the Claude Platform as `claude-haiku-5-5`.
The computer-use number, and how to read it
OSWorld 2.1 measures how well agents can operate a real computer to finish long, multi-step tasks. Here is the full row from Anthropic's table:
| Model | OSWorld 2.1 (offline subset) |
|---|---|
| Claude Haiku 4.5 | 15.7% |
| GPT-6 Luna | 48.9% |
| Claude Haiku 5.5 | 72.4% |
| Claude Sonnet 5.5 | 83.9% |
Some care is needed before drawing conclusions:
- These are vendor-run results. Anthropic ran or reported every number in the table, and points to the Haiku 5.5 system card for its evaluation method. Independent results may differ.
- It's a subset. The figures are labelled "offline subset", not the full benchmark.
- Your app is not the benchmark. A benchmark of general desktop tasks says little about how a model handles one specific product with its own menus, custom controls and sign-in flow. The only reliable test is running your own tasks.
- It's partial credit. Anthropic's OSWorld chart plots a "partial-credit score", so 72.4% is not the share of tasks completed end to end.
- Effort changes the score. Anthropic's accuracy-versus-cost charts show results at each effort setting, so a quoted score depends on how hard the model was allowed to work.
What the table does support is a relative statement: on Anthropic's own measurements, the jump from Haiku 4.5 to Haiku 5.5 on computer use is large, and Sonnet 5.5 is still ahead.
What it costs: the price table, with the arithmetic
Anthropic's prices per million tokens. Haiku 5.5 has two rates, depending on whether the prompt is up to or over 100,000 tokens:
| Per 1M tokens | Haiku 5.5 (up to / over 100k) | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| Input | $0.10 / $0.50 | $1.00 | $2.00 |
| Output | $0.50 / $2.50 | $5.00 | $10.00 |
| Cache writes | $0.125 / $0.625 | $1.25 | $2.50 |
| Cache reads | $0.01 / $0.05 | $0.10 | $0.10 |
Derived from that table, so you can check it:
| Comparison (prompts up to 100k) | Calculation | Result |
|---|---|---|
| Haiku 5.5 input vs Haiku 4.5 | $0.10 ÷ $1.00 | 10% of the price, i.e. 90% lower |
| Haiku 5.5 output vs Haiku 4.5 | $0.50 ÷ $5.00 | 90% lower |
| Haiku 5.5 input vs Sonnet 5.5 | $0.10 ÷ $2.00 | 1/20 of the price |
| Haiku 5.5 over-100k input vs Haiku 4.5 | $0.50 ÷ $1.00 | 50% lower |
Anthropic's headline figure is that Haiku 5.5 costs around 75% less to run on average than Haiku 4.5. Its footnote explains why that is less than the 90% list-price cut: 90% applies to requests up to 100,000 tokens (which Anthropic says were 90% of Haiku 4.5 requests), 50% applies above that, and Haiku 5.5's updated tokenizer uses slightly more tokens per task.
For computer-use agents, the over-100k rate deserves attention. Every step of a computer-use loop typically sends screenshots and history back to the model, so long sessions can push prompts past 100,000 tokens. That is our reading of how such loops tend to grow, not something Anthropic states; measure your own prompt sizes before budgeting on the lower rate.
Anthropic also halved Sonnet 5.5's cache-read price to $0.10 per million tokens on the same day, which it says makes Sonnet 5.5 around 20% cheaper on most agentic work.
Where Anthropic says Haiku 5.5 is not the right choice
The announcement is direct about this. On Terminal-Bench 4.0, a test of complex multi-step tasks in a command line, Haiku 5.5 scores 39.2% against Sonnet 5.5's 70.6%, and Anthropic writes that Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks. Haiku 5.5, it says, is best suited to more narrowly scoped tasks that would have been cost-prohibitive with earlier Claude models, such as compaction, summarization or subagent work.
A practical way to apply that, in our reading: give Haiku 5.5 well-defined, repeatable steps, and keep a larger model for planning, for unfamiliar interfaces, and for deciding when something has gone wrong.
What a cheaper computer-use model means for recorded demos and tutorials
Agents that operate software are increasingly used to produce material about that software: a demo of a new feature, a bug reproduction for a ticket, or a walkthrough that becomes a tutorial. Those jobs share a pattern: the same task is run again and again, every time the interface changes.
That is the kind of high-volume, narrowly scoped work Anthropic describes Haiku 5.5 as suited for. Our reading of what follows from that:
- Re-runs get cheaper. If a release changes three screens, re-running the walkthrough to capture them again matters more when each run costs less.
- Split the work. A larger model can plan the walkthrough and judge whether the result looks right, while a smaller one carries out the clicks it has been given.
- A recording is not a tutorial. An agent's run produces actions and screenshots. Turning that into something a person can learn from still needs narration, pacing and a human check that the steps make sense.
- Review stays essential. Anthropic's chart labels the OSWorld result a partial-credit score, so even a strong score includes tasks done only partly or wrongly. Anything published from an agent run should be watched by a person first.
Where Vorec fits
Vorec's `record-tutorial` skill lets Claude Code (or Codex, through the same CLI) research an app, plan a walkthrough and drive the Vorec macOS recorder to capture it locally. You review the take before anything is uploaded. Vorec then drafts narration matched to the captured workflow and generates the voiceover, so nothing is spoken into a microphone, and the same capture can also produce a written step-by-step guide. In the editor you can add smooth cursor motion, cursor-follow zoom and click-based auto-zoom, and freeze-sync can hold a frame so an explanation has time to finish. You can also record yourself with the Vorec app, or upload an existing recording. Plans start at $9/month (Starter); a free trial is available (current allowances on the pricing page).
FAQ
When was Claude Haiku 5.5 released?
7 October 2026, per Anthropic's announcement.
Can Claude Haiku 5.5 use a computer?
Anthropic reports 72.4% on the offline subset of OSWorld 2.1 and says the model works especially well for browser use. Its Python and TypeScript SDKs added computer use and browser use support in beta at the same time.
How much does Claude Haiku 5.5 cost?
$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens; $0.50 and $2.50 above that. Cache reads are $0.01 and $0.05.
Is Haiku 5.5 better than Sonnet 5.5 for computer use?
Not on Anthropic's numbers: Sonnet 5.5 scores 83.9% on the OSWorld 2.1 offline subset against Haiku 5.5's 72.4%. Haiku 5.5 is far cheaper per token.
What is the model ID?
`claude-haiku-5-5`.
Why does Anthropic say it is 75% cheaper when the list price is 90% lower?
Because 90% applies to prompts up to 100,000 tokens, 50% applies above that, and Haiku 5.5's tokenizer uses slightly more tokens per task. The 75% is Anthropic's average across those effects.
Related reading: Claude browser and computer use SDK toolsets, Computer use agents explained, OpenAI Agents API computer use and AI agent screen recording.
Sources: Anthropic, Introducing Claude Haiku 5.5 (7 October 2026) and the Haiku 5.5 system card; Claude Platform docs, Browser and computer use with the SDK toolsets. All checked 23 October 2026.
Want the recording to narrate itself? Record with Vorec, or upload one you already have, and get a narrated tutorial plus a written guide. Start free. Trial includes up to 3 projects; exports carry a watermark.