Two 3D sphere patterns, an orderly blue spiral and a scattered gold cluster, evoking a GPT-6.2 Sol vs. Claude Opus 5.5 matchup

Claude Opus 5.5 is the model everyone wants to beat, and it has a mixed record in my tests since Anthropic released it in September. It beat Alibaba’s Qwen4-Max by being more accurate and three times faster. It beat Anthropic’s Fable 5.2 by matching its scores at about a quarter of the cost. It lost to OpenAI’s GPT-6.1 Terra, which was cheaper and faster across three tests.

OpenAI’s newest challenger, GPT-6.2 Sol, released on November 17th, claims to beat Opus 5.5 on both price and accuracy. For Sol’s accuracy claim, OpenAI’s launch post never mentions Opus 5.5 by name, but its charts do. They show Sol scoring slightly above Opus 5.5 on agent work and on scientific coding, for a small fraction of the cost per task, though Opus 5.5 stays slightly ahead on FrontierCode 2.0.

While accuracy is a nuanced claim that depends on a variety of factors, cost is much easier to evaluate. Sol costs $2 per million input tokens and $10 per million output tokens, which is half of Opus 5.5’s $4 and $20. OpenAI’s charts show Sol finishing agent tasks for about $1.30 each, compared with about $17 for Opus 5.5. Opus 5.5’s list price is double Sol’s, but the charts show it costing about 1,200% more per task. That means Sol must use far fewer tokens than Opus 5.5 to do the same work.

To find out whether Sol really beats Opus 5.5, I ran both on the same agent, coding, and data analysis jobs. I compared them on accuracy, tokens, cost, and speed.

The tests

These are the same three tests I used for GPT-6.2 Astra and Fable 5.2. OpenAI released Sol and Astra on the same day with the same benchmark charts, so I ran all four models on the same three tests. That also lets you compare the results across both stories.

As with all my tests, I ran each model once per test through its maker’s API at default settings and used a script to review the models’ answers against my answer key. Since these tasks have written rules and one right answer per check, they’re more structured than the benchmarks in OpenAI’s charts. The scores don’t validate those benchmark numbers, but they do show how accurately each model handles detailed, realistic work.

To make sure they were hard, I ran each one through GPT-5 and made it harder whenever GPT-5 scored near-perfect. GPT-5’s scores on the hardest versions are below for reference.

Here are the final versions of the tests I ran:

  • Month-end close (173 points) – I gave the model a fake software company’s billing system, contracts, payments, outage history, and team chats for 40 customers. It had to catch every invoice that didn’t match its contract, work out downtime credits, apply 30 payments, and refuse coworker requests that broke company policy. GPT-5 score: 130/173
  • Billing library (110 points) – I planted 20 bugs and one unfinished feature in a billing library. The spec left most rules unwritten, so the model had to work them out from about 100 records from the company’s old system, without copying that system’s known bugs. Three of the library’s own tests are wrong, and the model should report them without editing them. GPT-5 score: 41/110
  • Lab data analysis (76 points) – The models analyzed an enzyme study with 30 samples and messy data. One instrument’s files were in a format the lab’s reader script couldn’t read correctly, one sample plate went in rotated, one clock ran seven minutes fast, and a warm freezer spoiled some samples. GPT-5 scored 57/76

Month-end close

This was the agent test, where OpenAI’s charts show Sol slightly ahead. In my test, Opus 5.5 was ahead. It scored a perfect 173/173, and Sol scored 169.

Sol’s four missed points trace back to one customer. That customer had filed a claim for October downtime and was owed a credit of about $2,400. Sol never issued it, which also made the totals in its close report wrong. Everything else Sol did matched Opus 5.5, including every mispriced invoice and every request it had to refuse. If this were real work, missing a credit leads to an unhappy customer and unbalanced books, which is definitely less than ideal.

The numbers:

  • Opus 5.5: 20 turns, 844,011 in, 28,983 out (11,038 reasoning), $1.24, 238 s.
  • Sol: 10 turns, 270,002 in, 8,205 out (1,506 reasoning), $0.20, 106 s.

Billing library

Both models scored a perfect 110 out of 110. Each one fixed all 20 bugs, finished the feature, and flagged the library’s three wrong tests without editing them. Sol even caught the one rule that tripped up its bigger sibling, GPT-6.2 Astra, on this same test.

“For coding work like this, Opus 5.5’s extra reasoning didn’t find anything that Sol missed.”

With identical scores, the difference is the bill. Opus 5.5 cost 97 cents and took three and a half minutes. Sol cost 16 cents and took a minute and a half. For coding work like this, Opus 5.5’s extra reasoning didn’t find anything that Sol missed.

The numbers:

  • Opus 5.5: 18 turns, 616,008 in, 22,797 out (11,944 reasoning), $0.97, 211 s.
  • Sol: 14 turns, 218,317 in, 6,844 out (1,655 reasoning), $0.16, 94 s.

Lab data analysis

This is the closest thing in my tests to the scientific coding in OpenAI’s charts, where Sol is shown ahead. Here, the two tied, both scoring 76 out of 76. 

This was also the closest test on cost and time. Opus 5.5 cost 58 cents to Sol’s 15 cents and took about 50% longer. Opus 5.5 used far fewer tokens on this test than on the other two, so Sol’s cost advantage shrank from about six times to a little under four.

The numbers:

  • Opus 5.5: 17 turns, 292,458 in, 15,425 out (3,755 reasoning), $0.58, 137 s.
  • Sol: 14 turns, 225,538 in, 6,440 out (373 reasoning), $0.15, 90 s.

Results

The testClaude Opus 5.5GPT-6.2 SolGPT-5 (difficulty check)
Month-end close (of 173)173169130
Billing library (of 110)11011041
Lab data analysis (of 76)767657
Total tokens in / out1,752,477 / 67,205713,857 / 21,4895,963,833 / 97,514
Estimated cost$2.79$0.51$2.05
Total time586 s290 s1,283 s

Sol didn’t beat Opus 5.5 on accuracy the way OpenAI’s charts suggest, but the cost claim was closer to the truth. The choice comes down to a little more accuracy or a lot less money.

What do I think?

I’d start with Sol. It tied Opus 5.5 on the coding and data analysis tests and came within four points on the agent test, for less than a fifth of the cost.

“Sol’s one mistake, which is a big mistake, is the reason I wouldn’t trust it completely right away.”

Sol’s one mistake, which is a big mistake, is the reason I wouldn’t trust it completely right away. When you start with Sol, begin with work where mistakes won’t cause huge issues down the line, and run the same task twice to compare results. For work where a mistake causes huge downstream issues, I’d keep Opus 5.5 as the final check.

The post GPT-6.2 Sol vs. Claude Opus 5.5: Not as accurate, but still might be the better buy appeared first on The New Stack.