Two days apart, Anthropic and OpenAI each shipped a mid-tier model built on the same pitch: most of their flagship's capability, at a fraction of the price. Claude Sonnet 5.5 launched September 28. GPT-6.1 Sol followed at OpenAI's DevDay on September 29. Both landed at the exact same headline price — $2 per million input tokens, $10 per million output tokens — which makes this one of the cleanest apples-to-apples comparisons of the year, provided you look past the sticker price to how each one actually spends its tokens.
This isn't a benchmark-table rehash of either launch piece. It's the direct, side-by-side question: now that both exist, which one actually wins, and at what?
Quick summary: Claude Sonnet 5.5 and GPT-6.1 Sol are both priced at $2 input / $10 output per million tokens. Sonnet 5.5 leads on the independent Artificial Analysis Intelligence Index (56 vs. 52) and on several task-specific benchmarks tracked by third-party aggregators — Terminal-Bench 4.0 (64% vs. 56%), SciCode (61% vs. 54%), and AutomationBench-AA (71% vs. 65%) — while GPT-6.1 Sol wins on PDF question-answering (31% vs. 26%) and is dramatically cheaper per completed benchmark task (a reported $0.72 vs. $7.67 at maximum effort, since Sonnet 5.5 uses far more tokens to reach its top scores). Context windows: 1,000,000 tokens flat for Sonnet 5.5; 1.05 million for GPT-6.1 Sol, though OpenAI applies a surcharge above 272,000 input tokens. Independent throughput testing on OpenRouter shows Sonnet 5.5 generating roughly four times faster (107 vs. 25 tokens/second) — a separate measurement from OpenAI's own 6-8x speed figure, which compares GPT-6.1 Sol's optional Ultrafast tier with its own Standard tier, not with Sonnet 5.5.
Same sticker price, different fine print
The headline $2/$10 match is real, but it's not the whole pricing story. Anthropic charges $0.20 per million tokens for cached input reads on Sonnet 5.5; OpenAI charges $0.10 for the same on GPT-6.1 Sol — half the cost, which matters a lot for anyone running long, repeated-context agentic sessions. OpenAI also applies a surcharge once a prompt exceeds 272,000 tokens, even though Sol's total context window is advertised at 1.05 million tokens; Anthropic prices Sonnet 5.5's entire 1-million-token window at one flat rate with no surcharge tier. Neither structure is objectively better — it depends on whether your workload tends to stay under 272K tokens per call or regularly blows past it.
| Metric | Claude Sonnet 5.5 | GPT-6.1 Sol |
|---|---|---|
| Input / output per 1M tokens | $2 / $10 | $2 / $10 |
| Cached input reads | $0.20 / 1M | $0.10 / 1M |
| Context window | 1,000,000 tokens (flat rate) | 1,050,000 tokens (surcharge above 272K) |
| AA Intelligence Index (max effort) | 56 | 52 |
| Terminal-Bench 4.0 (3rd-party tracked) | 64% | 56% |
| SciCode | 61% | 54% |
| AutomationBench-AA | 71% | 65% |
| PDF question-answering | 26% | 31% |
| Cost per Index task (max effort) | $7.67 | $0.72 |
| Raw throughput (OpenRouter median) | 107 tok/sec | 25 tok/sec |
Why the numbers don't all agree with each launch's own press release
Worth flagging directly: the Terminal-Bench 4.0 gap here (64% vs. 56%, per third-party comparison trackers) is narrower than Anthropic's own max-effort figure for Sonnet 5.5 (70.6%, measured against Sonnet 5 rather than Sol) and OpenAI's own figures for Sol are framed against Astra and the original GPT-6 Sol, not against Sonnet 5.5 at all. Different evaluators, different test runs, and different comparison baselines produce different numbers — which is exactly why a same-day, same-source comparison like this one is worth more than stitching together two separate vendor press releases.
The real trade-off: quality ceiling vs. cost per task
Sonnet 5.5 leads on general capability by a real but moderate margin across most third-party-tracked benchmarks, and it's meaningfully faster in raw output generation. GPT-6.1 Sol's clearest win is economic: a reported $0.72 to complete an Intelligence Index task at maximum effort, against $7.67 for Sonnet 5.5 — roughly eleven times cheaper for comparable top-end reasoning work, because Sonnet 5.5 uses substantially more tokens to reach its higher scores at that effort level. For high-volume, cost-sensitive agentic pipelines, that gap can matter more than a handful of benchmark points. For work where the ceiling on quality matters more than the per-task bill, Sonnet 5.5's lead holds up across most categories except PDF-heavy document work, where Sol actually scores higher.
- Pick Claude Sonnet 5.5 if: you want the higher general-capability ceiling, faster raw output, and your workload isn't extremely cost-sensitive per task.
- Pick GPT-6.1 Sol if: you're running high volume, cost-per-task matters more than a few benchmark points, or your work is PDF/document-heavy.
- Watch your context length: if you regularly exceed 272K tokens per call, OpenAI's surcharge tier changes Sol's effective cost — model your own typical prompt size before assuming the $2/$10 headline rate applies.
- Don't compare vendor press releases directly: Anthropic's and OpenAI's own benchmark numbers for these two models were measured against different baselines, not against each other.
If cost-per-task is your deciding factor, test both models at a lower, non-maximum effort setting on your actual workload first — the 11x cost gap reported here is specifically a maximum-effort comparison, and the gap narrows substantially at default effort levels for both models.
01Is Claude Sonnet 5.5 or GPT-6.1 Sol better?
It depends what you're optimizing for. Sonnet 5.5 leads on the independent Artificial Analysis Intelligence Index and most third-party-tracked task benchmarks, plus faster raw output. GPT-6.1 Sol is dramatically cheaper per completed task at maximum effort and wins on PDF question-answering specifically.
02Are they really the same price?
The headline per-token rate is identical: $2 input / $10 output per million tokens. The fine print differs — cached-read pricing, context-window surcharge thresholds, and real-world cost per task all diverge once you account for how many tokens each model actually uses to complete a task.
03Why is GPT-6.1 Sol so much cheaper per task if the per-token price is the same?
Cost per task depends on both the per-token price and how many tokens the model uses to reach its answer. At maximum reasoning effort, Sonnet 5.5 reportedly uses far more tokens per task than GPT-6.1 Sol does, which is why its cost per completed benchmark task is reportedly about 11 times higher despite the identical sticker price.
04Which one is faster?
Independent throughput testing on OpenRouter shows Sonnet 5.5 generating output roughly four times faster than GPT-6.1 Sol (107 vs. 25 tokens per second). That's a different measurement from OpenAI's own 6-8x speed figure, which compares GPT-6.1 Sol's optional Ultrafast tier with its Standard tier — not a claim about speed versus Sonnet 5.5.
05Which model should a developer pick for a new agentic project?
If cost per task at scale is the binding constraint, GPT-6.1 Sol's price-per-task advantage is hard to ignore. If the top priority is squeezing out the highest score on complex reasoning or coding tasks and budget is more flexible, Sonnet 5.5's benchmark lead and faster raw throughput make it the stronger default.
Spec sheets and benchmark tables are a starting point, not a verdict — the only real test is your own prompts. LumiChats lets you chat with Claude Sonnet 5.5, GPT-6.1 Sol, and 40+ other leading AI models side by side in one place, so you can see which one actually earns its price on the work you do every day.
