GPT-6.1 Sol and Terra Lead LiveBench 2.0 Browser Use at 93.3% and 93.1%, Above Astra's 91.3%
On the October 28 LiveBench 2.0 run, the two models leading the Browser Use category are not the most expensive in the table. GPT-6.1 Sol scores 93.3% and GPT-6.1 Terra scores 93.1%, ahead of GPT-6 Astra at 91.3%, Claude Fable 5.1 at 89.8%, and Claude Sonnet 5.5 at 89.3%.
LiveBench 2.0’s Browser Use category tests multi-step web navigation and form interaction tasks. Tasks are generated fresh each run to prevent training data contamination.
Browser Use, Cost, and Overall Score
| Model | Browser Use | Overall | $/task |
|---|---|---|---|
| GPT-6.1 Sol | 93.3 | 84.1 | $0.761 |
| GPT-6.1 Terra | 93.1 | 77.8 | $0.665 |
| GPT-6 Astra | 91.3 | 85.6 | $4.548 |
| Claude Fable 5.1 | 89.8 | 85.7 | $6.904 |
| Claude Sonnet 5.5 | 89.3 | 80.0 | $3.718 |
| Claude Opus 5.5 | 88.4 | 85.1 | $4.291 |
| Gemini 4 Argon | 82.8 | 79.3 | $2.073 |
GPT-6 Astra costs nearly 6x more per successful task than Sol and nearly 7x more than Terra. On Browser Use, it trails both by 2 points.
Terra’s First LiveBench 2.0 Appearance
GPT-6.1 Terra launched October 26 at $1/$5 per million input/output tokens. Its 93.1% Browser Use score is its highest published sub-category result across any benchmark to date. Terminal-Bench 4.0 placed it at 50.3% and the AA Coding Agent Index v2.0 placed it at 48, below Sol and above Luna. Browser Use is where Terra distinguishes itself from the rest of that performance profile.
Terra’s overall LiveBench 2.0 score of 77.8 ranks seventh on the full board. Its Agentic score of 48.6% is the lowest in the top seven. For workloads centered on web automation rather than multi-step agentic pipelines, the category result narrows the gap with models that lead on broader metrics.
What the Category Split Signals
Browser Use is increasingly relevant to agentic deployment. Tasks like navigating UIs to pull data, filling forms across services, and handling redirects and modal flows are common in enterprise automation. Models that lead on this category in controlled benchmarks have a stronger prior for production reliability in those contexts.
The result is also a pricing signal. At $0.665 per task, Terra completing the same Browser Use workload as Astra at $4.548 costs 85% less. Sol at $0.761 per task costs 83% less than Astra on the same category. For organizations running high-volume browser automation, the model tier with the highest Browser Use score on LiveBench 2.0 happens to be the cheapest.
Gemini 4 Argon, available only through the Fairwind Program for cyber defenders, scores 82.8% on Browser Use, the lowest in the top seven. Its strength on this run is Data Analysis (85.2%, first) and Reasoning (95.1%, first). Argon is purpose-built for adversarial reasoning, not web navigation.