Best AI Model for Computer-Use Agents 2026 (Ranked)
The short answer
Claude Opus 5.5 is the best model for computer-use agents as of September 2026: 81.8% on OSWorld 2.0 at $4/$20 per million tokens, ahead of every model and cheaper than every model within ten points of it. Claude Fable 5.1 (80.7%) is a rounding error behind at 2.5x the price. GPT-6 Astra (72.6%) is OpenAI’s best and costs the same as Fable. Below 70%, Meta’s Muse Spark 1.3 (66.9%) is the value option and Gemini 3.8 Flash (59.0%) the multimodal budget option. If your agent clicks through real software unattended, the ranking below is the one to budget against, with the caveat that latency, safeguards and injection resistance decide reliability as much as the score.
The ranking
OSWorld 2.0 scores as published by vendors or tracked by independent leaderboards. Pricing is per million tokens, standard API rates.
| Rank | Model | OSWorld 2.0 | Input / output | Cached input | Context | Released |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 81.8% (partial, Anthropic) | $4 / $20 | $0.20 | 1M | Sep 22, 2026 |
| 2 | Claude Fable 5.1 | 80.7% (partial, Anthropic) | $10 / $50 | $0.25 | 1M | Sep 3, 2026 |
| 3 | Claude Opus 5 | 74.0% (partial, Anthropic) | $5 / $25 | $0.50 | 1M | Jul 24, 2026 |
| 4 | GPT-6 Astra | 72.6% (OpenAI) | $10 / $50 | $1.00 | 1.05M | Sep 3, 2026 |
| 5 | Muse Spark 1.3 (Meta) | 66.9% (trackers) | $1.25 / $4.25 | ~$0.15 | 1M | Sep 2, 2026 |
| 6 | GPT-5.6 Sol | 66.2% (OpenAI) | $4 / $20 (promo) | — | 1.05M | Jun 2026 |
| 7 | GPT-6 Sol | 60.5% offline; 64.4% best | $2 / $10 | $0.20 | 1.05M | Sep 22, 2026 |
| 8 | Gemini 3.8 Flash | 59.0% | $0.75 / $3.75 (intro; $1.50 / $7.50 from Jan 1, 2027) | Standard | 1M | Sep 2, 2026 |
Notes on reading the table. Anthropic marks its OSWorld 2.0 runs “partial,” meaning a subset of the full task set; the ordering among Claude models is consistent across Anthropic’s table, but treat the exact gap to Astra with some caution. Grok 4.7, MiMo-V2.6-Pro, Kimi K3 and DeepSeek V4.1 Flash have no published OSWorld 2.0 score as of September 24, 2026, so they are omitted rather than guessed. Qwen3.8-Max’s widely cited 86.1 is on OSWorld-Verified, the earlier benchmark, and does not belong in this column.
Why Opus 5.5 leads on value, not only score
Two numbers explain it. First, price: Opus 5.5 is the only model above 75% that costs less than $10 per million input tokens, and its $0.20 cache read is what agent loops actually pay for, since every screenshot step resends the system context and tool definitions. Second, tokens per task: Anthropic reports Opus 5.5 uses fewer tokens than Opus 5 to finish the same work (an early tester’s 200,000-line codebase audit took under three hours versus more than twenty, with 2.5x fewer tokens), and matches GPT-6 Astra on Terminal-Bench 4.0 at about 40% of the cost per task. Computer-use agents are long-horizon by definition, so per-task efficiency compounds.
The migration cost is real. Opus 5.5 rejects the older computer_20251124 tool on the API and Google Cloud; you need the computer_toolset_20260801 toolset. Thinking cannot be disabled and forced tool use is unsupported. See How to migrate to Claude Opus 5.5.
When to pick something else
GPT-6 Astra if you are OpenAI-native, need the Agents API and Codex harness, or your tasks are science-heavy (Astra leads Terminal-Bench-Science at 64.6% vs Opus 5.5’s 58.7%). Expect roughly 40 minutes per OSWorld task and the >272K-token repricing on long sessions. Comparison in GPT-6 Astra vs Opus 5 vs Muse Spark 1.3 for computer use.
Muse Spark 1.3 if cost dominates and 67% is acceptable: at $1.25/$4.25 it is roughly a third of Opus 5.5’s list price and the best score under $2 input. Its “max reasoning” mode was still in safety testing at launch, and Meta’s pricing is published via trackers rather than a prominent official page, so confirm against billing.
Gemini 3.8 Flash if you need native audio and video input alongside screen control, or Google Cloud residency. Budget at the 2027 price ($1.50/$7.50), not the intro rate.
GPT-6 Sol if you want the cheapest frontier-lab model with a published OSWorld 2.0 number. At $2/$10 and 60–64% it is a fine choice for bounded, repetitive UI tasks with a verifier, not for open-ended ones.
What the score does not tell you
- Latency. A 72% model that takes 40 minutes per task and an 82% model that takes 15 are not close in production throughput. Anthropic reports Opus 5.5 generates output 30% faster than Opus 5; measure your own step times.
- Safeguards. Anthropic’s Opus 5.5 runs were scored with production safeguards on; when the classifier intervened, cyber tasks fell back to Opus 4.8. In your deployment, an intervention is a failed step unless you build the fallback.
- Prompt injection. A computer-use agent reads whatever is on screen, including hostile web pages. Opus 5.5 ties Fable 5.1 for the lowest injection success rate in Gray Swan’s benchmark and matches or beats Opus 5 in every setting Anthropic tested. No other vendor in this table publishes an equivalent number. See How to deploy a computer-use agent safely.
- Boundary behaviour. Anthropic reports Opus 5.5 is 85% less likely than Opus 5 to attempt to circumvent boundaries it has been given. For an agent with your credentials in a browser, that is a selection criterion.
Bottom line
Rank by OSWorld 2.0 and Claude Opus 5.5 wins. Rank by accuracy per dollar and it still wins, because the only models near it cost 2.5x more. Rank by cheapest acceptable and Muse Spark 1.3 or GPT-6 Sol take it. The tier list will move: Gemini 4 leaks claim 86.8% on OSWorld 2.0 for an unreleased model, Sonnet 5.5 and Haiku 5.5 are due within weeks, and Google has not yet answered the September 22 price cuts. This page is updated when a published score changes the order.
Last verified: September 24, 2026.