Claude Sonnet 5.5 vs GPT-6.1 Sol: Same $2/$10 Price, Different Bill

TL;DR: Anthropic released Claude Sonnet 5.5 on September 28, 2026, and OpenAI followed with GPT-6.1 Sol on September 29, both priced at $2 per million input tokens and $10 per million output tokens. With identical sticker prices, the real differences show up in the details: how each handles very long prompts, and how each behaves on tricky tasks. Cached input is also priced the same at $0.10 per million tokens on both, according to each company’s own pricing page. Independent tracking currently scores Sonnet 5.5 higher on general capability. Neither is a clear winner for everyone.

A Quick Note on How This Was Put Together

This is a research roundup, not a hands-on benchmark we ran ourselves. Pricing and specs come from vendor announcements as reported by MarkTechPost, DataCamp, and LLM Stats, and we rechecked the prices on October 16 against Anthropic’s and OpenAI’s own pricing pages; the hands-on test results come from DataCamp’s own comparison. The two companies published very few overlapping benchmarks, so most head-to-head claims below come from third-party tracking. We’ve linked every source so you can check the underlying numbers yourself rather than relying on our framing.

The Quick Spec Comparison

	Claude Sonnet 5.5	GPT-6.1 Sol
Release date	September 28, 2026	September 29, 2026
Input / output price (per 1M tokens)	$2 / $10	$2 / $10
Context window	1,000,000 tokens	1,050,000 tokens
Max output	128,000 tokens	128,000 tokens
Cached input (per 1M tokens)	$0.10	$0.10
Training data through	June 2026	April 2026

Where the Price Is Not Actually the Same

The base rate matches, and so does cached input, but one billing detail can change what you actually pay:

• Cached input: Both are $0.10 per million tokens on each company’s own pricing page. An earlier version of this post said Sonnet’s cached reads were $0.20, which was wrong; we’ve corrected it.
• Very long prompts: OpenAI’s pricing lists higher long-context rates for GPT-6.1 Sol, 2x on input and 1.5x on output, and DataCamp reports the cutoff at 272,000 tokens, with the whole request repriced once a prompt passes it. Anthropic’s pricing page says Sonnet 5.5 bills its full 1M window at the standard rate. For prompts that regularly run past that threshold, Sonnet can come out meaningfully cheaper, even with the same headline price.

This only matters at scale. If you’re sending short prompts, the price really is the same.

Want to see what this means for your own usage? Plug your request size and volume into our free LLM API cost calculator.

Benchmarks: Mostly Apples to Oranges

The vendors mostly benchmarked against other models rather than each other, which makes direct comparison tricky:

• Anthropic’s claims for Sonnet 5.5: 70.6% on Terminal-Bench 4.0 (versus 66.4% for Claude Opus 5.5), 80.1% on OSWorld 2.1 for computer use, and roughly 30% lower cost per task than Sonnet 5 through reduced token usage.
• OpenAI’s claims for Sol: about 2.2 points above Claude Opus 5.5 on AutomationBench, and matching GPT-6 Astra on DeepSWE.

On independent tracking, Artificial Analysis’ Intelligence Index, as cited by DataCamp, scores Sonnet 5.5 at 56 and GPT-6.1 Sol at 51.8. LLM Stats separately reports Sonnet 5.5 winning three of the four benchmarks the two models share. All vendor figures are self-reported, so treat them as directional. We covered how this tier of models compares with the flagship tier in our Claude Opus 5.5 vs GPT-6 breakdown and our GPT-6 Astra explainer.

A Real-World Test

DataCamp ran both models on a task of building a visualization of Dijkstra’s shortest-path algorithm:

• GPT-6.1 Sol scored 5.0 out of 5. Notably, it rejected a graph with negative edge weights, which breaks the algorithm.
• Claude Sonnet 5.5 scored 4.7 out of 5. It warned about the negative-weight problem but still returned an answer. On the other hand, it finished in fewer turns (4 versus 5) and fewer tool calls (3 versus 7).

That’s one task, so it’s an illustration of style rather than a verdict: Sol was stricter about bad input, Sonnet was faster and more willing to proceed with a warning.

So Which Should You Use?

Pick Claude Sonnet 5.5 if your prompts are very long (past roughly 272K tokens), you want faster iteration, or you need to run across multiple clouds (it’s reported live on Claude’s own platform plus AWS, Google Cloud, and Azure).

Pick GPT-6.1 Sol if you want stricter handling of invalid input, or you’re already deployed on Azure.

If neither describes you, the practical answer is to run your own real task through both. With identical list prices, the cost of testing is low and the result will tell you more than any leaderboard.

FAQ

Are Sonnet 5.5 and GPT-6.1 Sol the same price? At the base rate, yes: $2 per million input tokens and $10 per million output tokens. Cached input is also $0.10 per million tokens on both; long-prompt billing differs.

Which is more capable? On Artificial Analysis’ index as cited in coverage, Sonnet 5.5 scores higher (56 versus 51.8), though vendor benchmarks overlap very little.

Is GPT-6.1 Sol the same as GPT-6 Astra? No. Sol is OpenAI’s cost-optimized tier below the Astra flagship, which we explain in our GPT-6 Astra guide.

Is either one better than Gemini 4 Argon? Argon is a different tier and isn’t broadly available yet; see our Gemini 4 Argon explainer.

Final Thoughts

The story here is that pricing has converged at the mid-tier, so the decision moves from the sticker price to the fine print: long-prompt billing and how each model handles messy real tasks. If you’re building something at scale, model out your actual token patterns before you pick.

Sources: Anthropic pricing, OpenAI API pricing, MarkTechPost, DataCamp, LLM Stats