OpenAI's GPT-6.2 Sol Cuts Browser Agent Steps by 36% and Costs by 35%
Browser Use reports GPT-6.2 Sol hits 82% on its agent benchmark, finishing tasks twice as fast and 35% cheaper than its predecessor.
Takeaways
- GPT-6.2 Sol hits 82% on Browser Use's benchmark, up from 81% for GPT-6.1 Sol.
- Tasks finish in 12 minutes instead of 22, a roughly 2x speedup at high reasoning.
- Median steps per task drop from 105 to 67, a 36% reduction.
- Cost per task falls 35%, with Browser Use calling it the fastest top-scoring model.
- The claim is a tweet; Browser Use's public leaderboard still shows no GPT-6.2 Sol row.
- LiveBench independently puts GPT-6.2 Sol at 88.2 on browser tasks, $0.52 per success.
OpenAI’s newest mid-tier model, GPT-6.2 Sol, scored 82% on the Browser Use benchmark, narrowly ahead of GPT-6.1 Sol at 81%. Browser Use reports that the new model reduced median task length from 105 to 67 steps, cut median elapsed time from 22 to 12 minutes, and lowered cost per task by 35%. Those operating gains can reduce inference spend, browser runtime, and queue latency in production agents.
Fewer decisions drive the speedup
Browser Use attributes the speed gain to fewer actions per task and faster decisions at each step. A browser agent repeatedly observes a page through its DOM or a screenshot, adds that state to the model context, chooses an action, and waits for the browser to execute it. Every cycle consumes tokens and adds model, network, and browser latency, so removing 38 steps compounds across a run.
| Measure | GPT-6.2 Sol | GPT-6.1 Sol | Change |
|---|---|---|---|
| Benchmark score | 82% | 81% | +1 percentage point |
| Median steps | 67 | 105 | 36% fewer |
| Median elapsed time | 12 minutes | 22 minutes | 45% less time |
| Indexed cost per task | 65 | 100 | 35% lower |
The comparison used high reasoning effort, 180 shared tasks, and the same harness and judge for every model. Browser Use also describes GPT-6.2 Sol as the fastest model among the top scorers on this harness. The one-point accuracy increase is small for a benchmark that operates against live third-party sites, where page content, login flows, latency, and anti-automation systems can change between runs.
The leaderboard has not caught up
Browser Use’s public leaderboard still lists GPT-6.1 Sol at 81.3 and contains no row for GPT-6.2 Sol. At the time of the announcement, the results page showed a last-modified date of 29 October 2026, and the expected GPT-6.2 Sol model page returned a 404. The X post remains the only available source for the new result because Browser Use has not published a leaderboard entry or task-level results.
Live-site variability adds another limitation to the comparison. A shared harness and judge reduce methodological differences across models, but they cannot control every change on external websites. Repeated, interleaved runs are needed to separate model improvements from site drift and transient failures.
Other tests support the speed claim
LiveBench reports an 80.3 overall score for GPT-6.2 Sol, including 88.2 in its Browser Use category and 83.7 in coding, at $0.523 per successful task. That cost metric combines model pricing with pass rate, so retries and failures can materially affect it.
BenchLeader’s aggregate reports roughly 70 output tokens per second for GPT-6.2 Sol, compared with 45 for GPT-6.1 Sol, an increase of about 1.6 times. It also flags a regression on one browser sub-benchmark while showing gains on others. Token throughput measures generation speed rather than complete browser-task latency, and the mixed accuracy results indicate that performance still depends on the workload and harness.
Where 38 fewer actions pay off
- Data extraction: Multi-page scraping jobs that currently require long navigation and recovery sequences.
- Form workflows: Account provisioning, onboarding, and submission flows that users wait for synchronously.
- Automated testing: QA bots and end-to-end browser agents that run repeatedly in CI.
- Parallel research: Procurement and research agents traversing many comparable sites at once.
An aggregate increase from 81% to 82% does not identify a newly supported class of browser tasks. Teams should retest GPT-6.1 Sol failures, especially tasks involving authentication, dynamic interfaces, file uploads, and recovery from unexpected page states.
Efficiency gains extend beyond browsing
OpenAI reports a similar pattern on its internal GDPval v2 benchmark, which evaluates economically valuable knowledge-work tasks. GPT-6.2 Astra leads that evaluation while using 22% fewer output tokens than GPT-6 Astra, and GPT-6.2 Sol gains 5.3 points over GPT-6.1 Sol. Combined with Browser Use’s lower step count, the reported results point to a generation optimized for shorter trajectories and fewer model outputs per completed task.
Test before changing the default
- Build a representative set. Select at least 20 to 30 production tasks across easy, difficult, and historically unreliable workflows.
- Hold the environment constant. Use the same prompts, tools, browser image, reasoning effort, timeouts, retry policy, and judge.
- Interleave repeated runs. Alternate models and repeat tasks to reduce the effect of changing websites and transient network errors.
- Measure the full distribution. Record pass rate, median and tail latency, step count, retries, input and output tokens, and total cost per successful task.
- Inspect failures before rollout. Use sandboxes for workflows that modify accounts or submit data, then compare failure categories and recovery behavior.
Promote GPT-6.2 Sol when the shadow test preserves the required pass rate and produces meaningful reductions in latency or cost. Expand the sample when the accuracy difference is small or failures cluster around high-value tasks.