Sol, Not Terra, Is the Value Model for Agent Pipelines: LiveBench 2.0's Agentic Gap
Two OpenAI value-tier models sit in LiveBench 2.0’s top 10, and the cheaper one is the wrong choice for most agentic workloads.
GPT-6.1 Terra costs $0.665 per successful task — the lowest of any model in the top 10. GPT-6.1 Sol costs $0.819. That $0.154 gap looks trivial until you compare their Agentic scores: Sol at 68.7, Terra at 48.6. Terra’s 48.6 is the lowest Agentic mark in the top 10 by 11 points.
The Full Terra Profile
LiveBench 2.0’s eight categories reveal a model optimized for discrete, tool-use tasks rather than multi-step orchestration:
| Category | Terra | Sol | Fable 5.1 |
|---|---|---|---|
| Agentic | 48.6 | 68.7 | 77.9 |
| Coding | 94.3 | 96.0 | 97.7 |
| Browser Use | 93.1 | 93.3 | 89.8 |
| Reasoning | 88.7 | 92.6 | 91.7 |
| Data Analysis | 82.3 | 82.7 | 80.3 |
| Design | 82.3 | 82.1 | 83.4 |
| Research | 73.9 | 83.5 | 82.0 |
| Overall | 77.8 | 84.1 | 85.7 |
| $/task | $0.665 | $0.819 | $7.406 |
Terra leads Sol on zero categories outright. Its Coding (94.3) and Browser Use (93.1) are within 1.7 and 0.2 points of Sol respectively. Its Agentic score is 20.1 points behind Sol. Its Research score trails Sol by 9.6 points.
What “Agentic” Measures
LiveBench 2.0’s Agentic category evaluates multi-step task execution: planning across tool calls, maintaining context across turns, recovering from partial failures, and coordinating sub-goals without per-step human guidance. These are the load-bearing capabilities in autonomous agent pipelines.
Terra’s 48.6 on this dimension does not mean the model fails at agentic tasks outright. It means, on contamination-free evaluation, Terra resolves roughly half of what Claude Fable 5.1 (77.9) resolves. Against Sol (68.7), the gap is 20.1 points — meaningful if agents run hundreds or thousands of tasks.
Where Terra Actually Wins
Terra is the right model when the workload is:
- Coding tasks: 94.3 is 0.3 points off Sol and one of the top three scores in the table
- Browser automation at volume: 93.1 Browser Use at $0.665/task versus Astra’s 91.3 at $4.828/task is a 7x cost advantage on a task where Terra slightly trails Astra in quality
- Single-turn or lightly chained tasks: Where orchestration complexity is low and per-task cost dominates
Where Terra is the wrong choice: any pipeline requiring the model to track state across more than a few steps, plan over uncertain tool outputs, or self-correct across sub-tasks. For those workloads, Sol at $0.819 captures most of Terra’s cost advantage while delivering 20 more Agentic points.
The Real Cost of Getting This Wrong
At 1,000 tasks per day, the difference between Terra and Sol is $154. At the same scale, replacing Sol with Claude Fable 5.1 costs an additional $6,587 per day. The value opportunity is in Sol, not the step below it.
Claude Opus 5.5 scores 72.8 on Agentic — Sol’s closest competitor in that dimension — at $4.820 per task, nearly 6x Sol’s rate. Claude Sonnet 5.5 scores 63.4 Agentic at $4.140 per task, 5x the cost of Sol with lower agentic performance.
For teams running autonomous pipelines at any meaningful volume, Sol’s combination of 68.7 Agentic and $0.819/task is currently unmatched in the value tier. Terra’s cost floor is real, but the use case for it narrows to specialized coding and browser automation work where orchestration depth is shallow.