GPT-6.1 Sol cost less per task than Claude Sonnet 5.5 in every task I measured, and in the modeled agent run. Both list at $2.00 per 1M input tokens and $10.00 per 1M output tokens, so the gap comes from token counts: Sonnet 5.5 reads more tokens for the same text and writes more visible output.
Same list price, different bills
Both models list at $2.00 input, $10.00 output and $0.10 cached input per 1M tokens. Both always reason. Sol has a slightly larger context window (1,050,000 tokens against 1,000,000) and a higher LMArena Text score (1484 against 1476). Neither difference changes cost on the tasks below.
Price per token is only half the bill. Anthropic's tokenizer produces more tokens for the same English text: 1.39 times OpenAI's count, with OpenAI set to 1.00. You can see it in the PDF run, where Sonnet 5.5 received 18,004 input tokens and Sol received 12,987 for the same 20-page document.
Measured per-task costs
These four tasks were measured on both models: one run each, default settings, via OpenRouter, output capped at 16,000 tokens. See the methodology for details.
| Task | GPT-6.1 Sol | Claude Sonnet 5.5 | Median of all 326 models |
|---|---|---|---|
| Answer a chat question | 0.67¢ ($6.72 per 1,000) | 1.3¢ ($12.68 per 1,000) | 0.24¢ |
| Summarize a 20-page PDF | 3.1¢ ($31.05 per 1,000) | 4.5¢ ($44.58 per 1,000) | 0.83¢ |
| Review 500 lines of code | 4.9¢ ($48.89 per 1,000) | 8.0¢ ($79.66 per 1,000) | 1.4¢ |
| Write a 1,500-word blog post | 2.3¢ ($23.00 per 1,000) | 3.7¢ ($36.58 per 1,000) | 0.87¢ |
Both models sit above the median on all four tasks.
Chat answer
Sonnet 5.5 cost roughly twice as much as Sol. It wrote 1,256 output tokens with no reasoning tokens. Sol wrote 560 output tokens plus 103 reasoning tokens. Reasoning is billed as output, so Sonnet's bill came from the length of its visible answer.
20-page PDF summary
Both models mostly read here. Sonnet's input was 18,004 tokens against Sol's 12,987, and it wrote 857 output tokens against Sol's 482. Sol used 26 reasoning tokens and Sonnet used none. The input difference is the tokenizer effect.
Code review
Sol used 2,843 reasoning tokens and wrote 1,135 output tokens. Sonnet used 3,476 reasoning tokens and wrote 3,067 output tokens, and it also took more input (7,117 tokens against 4,557). This is the widest gap in dollar terms: 4.9¢ against 8.0¢.
Blog post
Sol used 172 reasoning tokens and 2,116 output tokens. Sonnet used 239 reasoning tokens and 3,401 output tokens. Sonnet wrote more text, and that text drives the cost.
The modeled 50-step agent
The agent scenario is modeled, not measured. It assumes an 8,000-token system prompt, 50 steps, 1,500 more context tokens per step and 300 visible output tokens per step.
| GPT-6.1 Sol | Claude Sonnet 5.5 | |
|---|---|---|
| Cost per task | 53¢ ($528.60 per 1,000) | 84¢ ($841.64 per 1,000) |
| Input tokens | 2,237,500 | 3,101,175 |
| Output tokens | 15,000 | 20,790 |
| Reasoning tokens | 0 | 10,900 |
| Without caching | $4.62 | $6.52 |
The median for this task across all 326 models is 48¢. Sol is slightly above it and Sonnet is well above it. Sonnet's extra cost comes from the larger input count, which reflects the tokenizer, plus its reasoning tokens.
Caching matters a great deal here. Without it, Sol's task rises from 53¢ to $4.62 and Sonnet's from 84¢ to $6.52. Either model's agent bill depends heavily on cache hits, so check your own prompt structure. You can test your own step counts in the calculator, and see the agent task page for the full ranking.
| # | Model | Cost per task | WebDev score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  DeepSeek V4.1 FlashDeepSeek | 2.4¢$24.06 / 1,000 | 1618 | 2,235,262 / 102,435 | modeled best value | |
| 2 |  GLM 5.3 FlashZ.ai | 2.6¢$25.63 / 1,000 | 1606 | 2,239,737 / 43,814 | modeled | |
| 3 |  GPT-6 LunaOpenAI | 4.1¢$40.71 / 1,000 | 1578 | 2,237,500 / 22,000 | modeled | |
| 4 |  MiMo-V2.6-ProXiaomi | 6.5¢$64.51 / 1,000 | 1625 | 2,241,975 / 24,380 | modeled best value | |
| 5 |  Claude Haiku 5.5Anthropic | 6.6¢$66.35 / 1,000 | 1585 | 3,101,175 / 50,340 | modeled | |
| 6 | GPT-5.4 NanoOpenAI | 8.7¢$86.92 / 1,000 | 1401 | 2,237,500 / 22,000 | modeled | |
| 7 |  Qwen3.8 FlashAlibaba Qwen | 9.2¢$92.31 / 1,000 | - | 2,259,875 / 96,000 | modeled | |
| 8 |  Mistral Small 4Mistral | 13¢$135 / 1,000 | - | 2,304,625 / 147,800 | modeled |
Verdict by task type
- Short chat answers: Sol. 0.67¢ against 1.3¢. Sol spent some tokens on reasoning, but its visible answer was less than half the length.
- Document summaries: Sol. 3.1¢ against 4.5¢, helped by the smaller input count from the tokenizer.
- Code review: Sol. 4.9¢ against 8.0¢, with less output for the same 497-line file.
- Long-form writing: Sol. 2.3¢ against 3.7¢ for the 1,500-word post.
- Multi-step agents: Sol is cheaper in the modeled run (53¢ against 84¢), but test caching behavior before committing.
My numbers cover cost only, not answer quality. The LMArena gap between the two models is small. If quality decides your choice, run your own evaluation on your own inputs. If cost decides it, the data points the same way on every task. Model pages: GPT-6.1 Sol and Claude Sonnet 5.5.
FAQ
Why do two models with the same price per token have different costs per task?
Cost depends on how many tokens a model reads and writes, not just the price per token. Anthropic's tokenizer produced 1.39 times OpenAI's count on the same English document, and Sonnet 5.5 also wrote more visible output on all four measured tasks.
Are the agent numbers measured?
No. The 50-step agent is a modeled scenario with fixed assumptions: an 8,000-token system prompt, 1,500 more context tokens per step and 300 visible output tokens per step. The other four tasks were measured, one run each.
Does the cheaper model win on quality too?
My cost data does not say. Sol scores 1484 on LMArena Text and Sonnet 5.5 scores 1476, a small gap, but that score does not measure your task. Test both on your own inputs.