GPT-6.1 Sol offers a meaningful improvement over GPT-6 Sol at the same standard input and output token prices. Independent testing also places it close to GPT-6 Astra, although Claude Opus 5.5 and Sonnet 5.5 achieve higher overall scores.
The table below uses Artificial Analysis results at maximum reasoning effort. The Intelligence Index combines ten evaluations, covering reasoning, coding, knowledge and professional work. It is an index score, not a percentage of questions answered correctly.
| Model | AA Intelligence Index, higher is better | Average AA cost per task | API input per 1M tokens | API output per 1M tokens |
|---|---|---|---|---|
| GPT-6 Sol | 48 | $1.05 | $2 | $10 |
| GPT-6.1 Sol | 52 | $0.72 | $2 | $10 |
| GPT-6 Astra | 53 | $3.26 | $10 | $50 |
| Claude Sonnet 5.5 | 56 | $7.60 | $2 | $10 |
| Claude Opus 5.5 | 58 | $5.98 | $4 | $20 |
Figures retrieved on September 29, 2026. Claude results use the default fallback configuration. Task costs are weighted averages across the benchmark suite, while token prices are standard API rates.
The comparison is straightforward. Sol 6.1 gains 4 index points over Sol 6 and sits 1 point behind Astra. Calculated from these figures, its average benchmark cost is approximately 31% lower than Sol 6 and 78% lower than Astra.
Claude leads this overall quality comparison at maximum effort, but uses a considerably larger budget. Sonnet’s result also shows why token prices alone do not tell the whole story, its input and output rates match Sol 6.1, yet its measured cost per task is much higher.
OpenAI’s announcement adds the following task specific comparisons:
| Benchmark | Compared with | GPT-6.1 Sol’s reported score difference |
|---|---|---|
| DeepSWE 1.1, software engineering | GPT-6 Sol’s best result | 6.4 percentage points ahead, at lower reasoning effort |
| AutomationBench, medium effort | Claude Opus 5.5 | 2.2 percentage points ahead |
| OSWorld 2.0 offline, maximum effort | GPT-6 Astra at maximum effort | 2.1 percentage points behind |
These are score differences, not absolute scores, and the reasoning settings vary between rows.
For teams choosing a model, these results make Sol 6.1 a strong candidate when both quality and operating cost matter. Compare models on the same real tasks, including their token use, completion quality and total bill.
Read the OpenAI announcement and the Artificial Analysis results. Explore AI models at opencheese.dev.