Skip to content
CrawlfolioSourced research on web data for AI
Menu

Explainer

Gemini vs Claude computer use, after both left beta

Claude's computer toolset left beta and new models reject the old tool; Gemini Computer Use went GA Oct 30. What changed, what it costs, which to pick.

A curving white marble staircase with a wooden handrail on thin brass posts, sweeping up under a pale oval opening in a dim hall with a few armchairs by curtained windows
Photo: Serhat Beyazkaya / Unsplash · Unsplash License

Both computer use tools are out of beta now: Claude’s since August 19 and Gemini’s since October 30. If you run an agent that drives a screen, you have to pick the model that drives it and decide what to change before three dates: November 30, December 1 and January 1.

The short version: Anthropic’s newest models reject the old Claude tool, and three Gemini models double in price. On list prices Claude Haiku 5.5 is the cheapest per step, only Gemini documents a phone environment, and Claude keeps screenshots in your own environment. Below are the dated changes, a 20-step cost illustration and my pick for each job.

Short on time?

Short on time? Here's my take.

Cheapest per step
Claude Haiku 5.5
About $0.04 of input in my 20-step illustration, against $0.14 on Gemini 3.5 Flash-Lite. My estimate from list prices, not a measured run.
Android or phone automation
Gemini (3.5 Flash or later)
Gemini is the only one of the two with a documented mobile environment.
Regulated data, no vendor retention
Claude computer toolset
Client-side, screenshots stay in your environment, and it's ZDR and HIPAA eligible.
Browser-only jobs on Claude
Claude browser toolset
It reads the accessibility tree with element refs and tabs, falling back to screenshots and coordinates.
Already on Gemini 4 Carbon or Argon
Price out Gemini 4 Hydrogen
Carbon and Argon go from $2/$10 to $4/$20 per million tokens on December 1. Hydrogen lists $1.50/$7.50.

Prices verified . Sponsors cannot buy a verdict or a position.

What changed this fall, side by side? #

Claude’s change was a series of steps from August 19 to November 9, not one event. Gemini’s tool became built in during June and went GA on October 30, and its prices step up on December 1 and January 1.

Date (2026) Claude Gemini
Jun 24 Computer use becomes a built-in tool in Gemini 3.5 Flash, in public preview (TNW)
Aug 19 computer_toolset_20260801 out of beta, no beta header; browser_toolset_20260801 launches (Google Cloud from Aug 20)
Sep 22 Opus 5.5: old computer_20251124 returns a 400, except on Bedrock
Sep 28 Sonnet 5.5: the earlier tool “isn’t accepted”
Sep 30 Sonnet 4.5 deprecated, retiring Nov 30 on the Claude API
Oct 7 Haiku 5.5: computer_20250124 also returns a 400
Oct 8 gemini-3.5-flash now routes to gemini-3.6-flash
Oct 26 New paid-tier projects store Generate Content requests by default (store=true)
Oct 30 Computer Use GA on the Interactions and GenerateContent APIs
Nov 9 Fable 5.2 and Mythos 5.2: old tool returns a 400 everywhere but Bedrock
Nov 30 Sonnet 4.5 retires Last day of $2/$10 for 4 Carbon and 4 Argon
Dec 1 Carbon and Argon go to $4/$20; gemini-2.5-pro and -flash reroute
Jan 1, 2027 3.8 Flash goes from $0.75/$3.75 to $1.50/$7.50

Sources: Claude release notes, Gemini changelog, Gemini pricing. Prices are USD per million input/output tokens, verified 2026-11-16.

Will your Claude computer use agent break? #

Only when you move to a new model while still sending the old tool. From Opus 5.5 on, Anthropic says “the earlier computer_20251124 tool returns a 400 error; on Amazon Bedrock, computer_20251124 keeps working”. By November 9 that held for Fable 5.2 and Mythos 5.2 “on every platform except Amazon Bedrock”. Release notes.

Older models keep their versions. computer_20251124 still runs on Fable and Mythos 5.1 and 5, Opus 5, Sonnet 5, Opus 4.8 to 4.5 and Sonnet 4.6; computer_20250124 on Sonnet 4.5 and Haiku 4.5, plus Opus 4.1, Sonnet 4 and Opus 4 on the platforms where they aren’t retired. Computer use docs.

The docs give a nine-step migration. These are the parts that change behaviour:

  1. New shape. The toolset has 17 members (screenshot, zoom, the click types, scroll, type, key, wait and more) and no action field. It rejects name, display_width_px, display_height_px, display_number and enable_zoom.
  2. Zoom flips on. “Zoom is enabled by default on the toolset, whereas enable_zoom defaults to false.” Turning it off “removes about 410” tokens per request.
  3. Several actions per turn. Handle every tool_use block, and after a failure return “Not executed: an earlier computer action in this turn failed.” for the rest.
  4. Smaller changes. key gets a repeat count from 1 to 100, and you resize screenshots yourself.

That last one matters: oversized computer use screenshots are rejected, not downscaled. “Claude 4.7 and later models” accept up to 2,576 px; others 1,568 px. Vision docs. On October 7, Python SDK 1.12.0 and TypeScript SDK 0.132.0 added classes for both toolsets.

Dark speckled stone stairs climbing beside a grey wall, with a zigzag metal handrail lit from beneath by a warm strip of light
The rail changes direction at every landing, and a move to a newer model asks the same of your code: follow each change to the tool definition in turn rather than assuming the old shape still fits. Photo: Alessia Cocconi / Unsplash · Unsplash License

What changed on the Gemini side? #

Computer use stopped being a separate model. On June 24 it became “now a native tool inside Flash” in public preview, replacing the 2.5 model, and Google’s changelog marked it GA on October 30. TNW, changelog.

The tool works in three environments. AI Studio’s docs list a browser environment with 20 actions, a mobile one with 10 (including open_app and long_press) and a desktop one with 17 (including hotkey). Coordinates are integers from 0 to 999, so you don’t declare a screen size.

Browser is the default, and Google Cloud’s docs cover mobile and desktop under “Gemini 3.5 Flash or later”, so I’d plan on 3.5 Flash or newer for them.

Watch the model IDs. AI Studio lists six models for the tool and Google Cloud lists ten, so use the list for your platform. gemini-3.5-flash already routes to gemini-3.6-flash, and from December 1, gemini-2.5-pro and gemini-2.5-flash route to gemini-3.8-flash and gemini-3.5-flash-lite. Changelog.

One more oddity: the AI Studio page shows code only for the Interactions API, and neither docs page carries a GA label. I go by the changelog’s date.

So what does a 20-step run cost? #

On list prices, Claude Haiku 5.5 is the cheapest per step, then Gemini 3.5 Flash-Lite. Neither vendor charges a separate tool fee: Claude computer use “follows the standard tool use pricing”, and Google says it is “priced at the same rates as Gemini and uses the same SKUs”. Claude pricing, Google Cloud.

Gemini lists Computer Use as “Not available” on the free tier and “Charged as regular tokens per model pricing” on paid. Gemini pricing. Claude’s pricing page lists no free tier, though it says “New users receive free credits to test the API.”

Model Input / output per 1M tokens (USD)
Claude Haiku 5.5 $0.10 / $0.50 up to 100,000-token prompts ($0.50 / $2.50 above)
Claude Sonnet 5.5 $2 / $10
Claude Opus 5.5 $4 / $20
Claude Fable 5.2 $8 / $40
Gemini 3.5 Flash-Lite $0.30 / $2.50
Gemini 3.8 Flash $0.75 / $3.75 to Dec 31, then $1.50 / $7.50
Gemini 4 Hydrogen $1.50 / $7.50
Gemini 4 Carbon, 4 Argon $2 / $10 to Nov 30, then $4 / $20

Verified 2026-11-16 on Claude pricing and Gemini pricing.

Screenshots pile up in the context, so per-token prices don’t tell you what a task costs. Here is my estimate for one illustrative run, built from each vendor’s own image rules:

  • Screenshot size: a 1280x800 viewport. Claude charges ⌈1280/28⌉ × ⌈800/28⌉ = 46 × 29 = 1,334 tokens (vision docs). Gemini charges 2,240 at ultra_high, which Google calls “required for specific use cases such as computer use” (media resolution).
  • Fixed overhead: Claude’s toolset “adds about 4,500 input tokens to a request: about 4,590 on Claude Sonnet 5 and about 4,520 on the other supported models” (computer use docs). I use 4,520. Google states no overhead, so I use zero.
  • The run: 20 steps, one new screenshot per step, full history re-sent, no caching, output tokens left out.
  • Claude: (20 × 4,520) + (1,334 × 210) = 90,400 + 280,140 = 370,540 input tokens. The 210 is 1 + 2 + … + 20 screenshots.
  • Gemini: 2,240 × 210 = 470,400 input tokens.
Model My estimate per 20-step run (input only)
Claude Haiku 5.5 $0.04
Gemini 3.5 Flash-Lite $0.14
Gemini 3.8 Flash $0.35 to Dec 31, $0.71 from Jan 1
Gemini 4 Hydrogen $0.71
Claude Sonnet 5.5 $0.74
Gemini 4 Carbon or Argon $0.94 to Nov 30, $1.88 from Dec 1
Claude Opus 5.5 $1.48
Claude Fable 5.2 $2.96

My estimate, not a measured run; prices verified 2026-11-16. Haiku’s prompts stay under its 100,000-token price step until step 72.

Because Google publishes no per-request overhead and I set it to zero, this comparison leans slightly toward Gemini. Caching would cut the Claude figures a lot, since cache hits cost 0.05x on Sonnet and Opus 5.5; I didn’t find Gemini’s caching terms for Computer Use.

How do the safety defaults differ? #

Claude scans for prompt injection by default; Gemini makes you turn it on. That’s the difference I’d act on first.

Claude computer use Gemini Computer Use
Prompt-injection check Classifiers “automatically scan what the tool returns, such as screenshots”, on by default (opt out via support) enable_prompt_injection_detection, off by default
Confirmations Docs advise “a dedicated virtual machine or container with minimal privileges” and “an allowlist of domains” Per-action safety_decision; on require_confirmation, “prompt the end user”; seven default categories can be disabled (3.5 Flash+)
Admin lock Not covered in the doc I read Org policy vertexai.configurableComputerUseToolSafetyPolicies (Google Cloud)
Where screenshots live “Computer use is a client-side tool”; screenshots are “stored in your environment, not by Anthropic” New paid-tier projects store Generate Content requests by default since Oct 26 (store=true)
Compliance ZDR eligible (excluding Covered Models) and HIPAA eligible Not stated on the pages I read

Sources: Claude computer use docs, AI Studio Computer Use, Google Cloud Computer Use, Gemini changelog.

Gemini also lets you drop predefined actions through excluded_predefined_functions and add your own, such as a yield_to_user step.

Side profile of a white staircase whose dark-grained wooden treads form a sharp zigzag against a plain off-white wall
Every tread carries the same dark edge, and a safety check only protects every step the same way if it is switched on for all of them, so confirm which defaults your agent starts with before its first run. Photo: Ricardo Gomez Angel / Unsplash · Unsplash License

Doesn’t Claude have a browser tool too? #

Yes, and for browser-only jobs I’d use it instead. browser_toolset_20260801 launched the day the computer toolset left beta. It works from the accessibility tree, element refs and tabs as well as screenshots, and it “adds about 6,600 input tokens”. Browser use docs.

It has 31 members, 27 on by default; file_upload, read_console, read_network and javascript_exec are off. It is ZDR eligible but HIPAA “Not eligible”, so regulated work stays on the computer toolset.

If you’re weighing a page-reading agent against a screen-driving one more broadly, Browser Use vs Stagehand and browser MCP servers cover the DOM-based route. And if the job is reading pages, not acting on them, search vs fetch vs browser may save you the browser entirely.

Which one should you pick? #

Your job My pick Why
Cheapest screen-driving agent Claude Haiku 5.5 Lowest estimate per run on list prices
Android or phone tasks Gemini, 3.5 Flash or later Only documented mobile environment
Health, finance or other regulated data Claude computer toolset Client-side, ZDR and HIPAA eligible
Web-only tasks on Claude Claude browser toolset Element refs and tabs, not just pixels
Mid-price model, either vendor Sonnet 5.5 or Gemini 4 Hydrogen About level in my estimate ($0.74 vs $0.71)
Running on Bedrock with old code Stay put for now Bedrock still accepts computer_20251124

Where the agent runs is a separate choice; the cloud browser roundup and self-hosted vs managed browsers cover it, and Cloudflare’s Kitesurf is one recent option. The agent browsers hub collects every guide on running browsers for AI agents.

What I would do this week #

  • On Claude 4.x with computer_20251124: migrate to the toolset, deciding on zoom first, before you switch to any 5.5 or 5.2 model on the Claude API or Google Cloud. Bedrock still accepts the old tool, but don’t count on that lasting.
  • On Sonnet 4.5: move before November 30.
  • On Gemini 2.5 or 3.5 Flash IDs: pin a model from your platform’s list now instead of riding the reroutes.
  • On Gemini 4 Carbon or Argon: budget for the December 1 doubling or price out Hydrogen.
  • On any Gemini Computer Use agent: turn on prompt-injection detection. It is off by default.

I didn’t find an independent accuracy benchmark for either tool, so run your own task set on both before committing a big workload.

FAQ

Does Claude's old computer use tool still work?

On older models, yes. On Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.2 and Mythos 5.2, a request that declares computer_20251124 returns a 400 error everywhere except Amazon Bedrock, and Haiku 5.5 also rejects computer_20250124. Those models need computer_toolset_20260801.

Is Gemini Computer Use free?

No. Google's pricing page lists Computer Use as 'Not available' on the free tier. On the paid tier it is 'Charged as regular tokens per model pricing', with no separate tool fee.

How many tokens does a screenshot cost on each?

On Claude, an image costs ceil(width/28) x ceil(height/28) tokens, so a 1280x800 screenshot is 1,334 tokens. On Gemini 4 and 3 models an image is 1,120 tokens by default and 2,240 at ultra_high, which Google calls 'required for specific use cases such as computer use'.

Can Gemini control a phone?

Yes. Gemini's Computer Use tool has a mobile environment with 10 actions, including open_app and long_press. Google Cloud's docs cover mobile and desktop under 'Gemini 3.5 Flash or later'. Claude's computer use docs describe no mobile environment.

Do I need a beta header for Claude computer use now?

Not for the new toolset. Anthropic's August 19, 2026 release note says computer use is out of beta as computer_toolset_20260801 with no beta header. The older computer_20251124 version still uses the computer-use-2025-11-24 header on the models that accept it.

Sources

11 sources, accessed .

Official pricing, docs, changelogs and status pages 10

  1. Claude release notes (Aug 19 to Nov 9, 2026) platform.claude.com · accessed
  2. Claude docs: computer use tool platform.claude.com · accessed
  3. Claude docs: browser use tool platform.claude.com · accessed
  4. Claude docs: vision and image token costs platform.claude.com · accessed
  5. Claude pricing platform.claude.com · accessed
  6. Gemini API changelog aistudio.google.com · accessed
  7. Google AI Studio docs: Computer Use aistudio.google.com · accessed
  8. Google Cloud docs: Computer Use docs.cloud.google.com · accessed
  9. Google AI Studio docs: media resolution aistudio.google.com · accessed
  10. Gemini API pricing aistudio.google.com · accessed

News and press 1

  1. TNW: Gemini 3.5 Flash gets computer use as a built-in tool (2026-06-24) thenextweb.com · accessed