divergent

Haiku 5.5 makes a computer for every agent practical

Haiku 5.5 uses a computer for less than half the cost of Sonnet. Your lead agent can now give each of its subagents a computer of its own.

Nick Vernij · · 4 min read

Anthropic released Claude Haiku 5.5. It approaches Sonnet-class computer-use performance at a fraction of the cost.

What caught our eye is Haiku 5.5 scoring 72.4% on OSWorld 2.1, a benchmark for agents operating real computers, outperforming other models in cost-per-task by a big margin.

Computer-use is finally affordable

Computer use is moving from an expensive frontier-model demo toward something you can run repeatedly and at scale. On OSWorld 2.1, Haiku completes tasks while spending 50% less per attempt:

Cost per task at the same OSWorld 2.1 scoreHaiku 5.5Comparable model
42%
$0.07
$0.14 GPT-6 Luna
67%
$0.28
$0.93 Sonnet 5.5
73%
$0.61
$1.38 Sonnet 5.5
Cost per attempt for each model to reach the same OSWorld 2.1 score. Lower is better.

Browsing faster than any other model

OpenRouter’s current median measurements show how quickly each model generates a response:

OpenRouter median output speed, in tokens per second. Higher is better.OpenRouter median time to first token (p50). Lower is better.

That makes Haiku about 1.4× faster than Sonnet, 1.5× faster than Opus, and 1.8× faster than Luna. These are live provider measurements and will move over time.

See OpenRouter’s live measurements

Haiku’s natural role is the subagent

Use Opus or Sonnet to plan and make judgment calls. Delegate bounded execution to Haiku: research a record, test a flow, update a dashboard, or complete a browser task.

Ideally you would give each Haiku subagent its own computer, files, and narrowly scoped credentials, so you can run them in parallel and then return the results to the lead agent for review.

At Haiku’s price and speed, this is much cheaper and more reliable than sending one expensive model through a long workflow.

The model can click. It still needs a computer.

Anthropic developed a standard for computer-use with their defined SDKs and tool calls. An in-house SDK is likely what makes this model perform well for so few tokens; the Anthropic team had the chance to move this standard into their training data and optimize for the best use of it.

However, Haiku does not come with a computer, and the environment matters. A useful computer-use agent needs:

  • An isolated machine, so one agent cannot leak into another task.
  • A real browser that the model can see and control, and where the user can interrupt at any time.
  • Files and state that survive longer than one API call.

Coincidentally, the Divergent team launched embedded sandboxes today.

Any agent can control an isolated computer from your server, while your user watches the screen live inside your product.

Haiku 5.5trip
Cheapest nonstop from Amsterdam to SF, Oct 20 to 27? Mornings if possible.
computertype "San Francisco"✓ typed
computerleft_click [668, 330]✓ Cheapest
computerleft_click [480, 376]✓ KLM 10:25
KLM at 10:25, €512 round trip. Prices are low for these dates, so now is a good time to book.
Book it. Window seat.
Haiku 5.5 finds a flight in a Divergent sandbox with Anthropic’s computer tool. Run the example app from GitHub.

When a site requires authentication, the user can take control, sign in, and hand the browser back without exposing their password to Claude or your server.

The integration uses Anthropic’s standard computer toolset, so it works with Haiku out of the box.

See the implementation guide or start from the example app on GitHub.

What we would build now

The most compelling Haiku 5.5 applications will not look like chatbots with a cursor. They will look like small teams:

  • A support agent reproducing each bug in a clean machine.
  • A QA coordinator forking the same application state across test cases.
  • An operations agent reconciling records across tools that have no shared API.
  • A coding agent giving every implementation, test, and review subagent an isolated workspace.

Computer use has always been possible with the latest frontier models; however, Haiku makes it much more economically viable, opening up many more use cases.

We’d love to give your Haiku agent a computer. Divergent sandboxes start in under a second and run a full browser with multiple tabs. Your users can watch the agent work live with our Screen SDK.

Sources