Since Anthropic launched Claude Opus 5.5 in September, the model has delivered mixed results in competitive evaluations. It outperformed Alibaba's Qwen4-Max on accuracy and speed, and matched Anthropic's Fable 5.2 at a quarter of the cost. However, OpenAI's GPT-6.1 Terra proved both cheaper and faster across multiple benchmarks.

OpenAI's latest offering, GPT-6.2 Sol, arrived on November 17th with claims to surpass Opus 5.5 in both price and accuracy. While OpenAI's launch materials avoid naming Opus 5.5 directly, their published charts compare the two models, showing Sol slightly ahead on agent work and scientific coding tasks, though Opus 5.5 maintains a slight edge on FrontierCode 2.0.

The pricing comparison is straightforward: Sol charges $2 per million input tokens and $10 per million output tokens, half of Opus 5.5's $4 and $20 rates. OpenAI's benchmarks indicate Sol completes agent tasks for approximately $1.30 each versus roughly $17 for Opus 5.5. Though Opus 5.5's list price is double Sol's, per-task costs show a roughly 1,200% difference, suggesting Sol achieves similar results using substantially fewer tokens.

To verify Sol's claims, I tested both models on identical agent, coding, and data analysis workloads, measuring accuracy, token consumption, cost, and execution speed.

The tests

I used the same three evaluation tasks applied to GPT-6.2 Astra and Fable 5.2. Since OpenAI released Sol and Astra simultaneously with matching benchmark charts, running all four models on identical tests enabled cross-comparison across both evaluations. Each model ran once per test through the vendor's API at default settings, with a script validating responses against an answer key. These structured tasks with defined rules and single correct answers differ from OpenAI's published benchmarks but demonstrate how each model handles detailed, realistic work.

To ensure adequate difficulty, I ran each test through GPT-5 and increased complexity whenever GPT-5 approached perfect scores. The reference scores from GPT-5 on the hardest versions appear below.

The final test versions were:

  • Month-end close (173 points) – Models received a fictional software company's billing system, contracts, payments, outage history, and team communications for 40 customers. Tasks included identifying every mismatched invoice, calculating downtime credits, processing 30 payments, and rejecting policy-violating coworker requests. GPT-5 score: 130/173
  • Billing library (110 points) – I introduced 20 bugs and one incomplete feature into a billing library. The specification omitted most rules, requiring models to infer them from approximately 100 records of the company's legacy system without replicating its known bugs. Three library tests contained errors that models should identify without modification. GPT-5 score: 41/110
  • Lab data analysis (76 points) – Models analyzed an enzyme study containing 30 samples with messy data. One instrument's files used a format the lab's reader script mishandled, one sample plate was inserted rotated, one clock ran seven minutes fast, and a malfunctioning freezer damaged some samples. GPT-5 score: 57/76

Month-end close

This agent task showed Sol slightly ahead in OpenAI's charts, but my testing revealed the opposite. Opus 5.5 achieved a perfect 173/173, while Sol scored 169.

Sol's four-point deficit stemmed from a single customer. The customer had submitted an October downtime claim entitling them to approximately $2,400 in credit, which Sol never issued. This omission also corrupted the totals in Sol's final close report. Sol matched Opus 5.5 on all other metrics, including every mispriced invoice and every request it appropriately rejected. In production, a missed credit represents a frustrated customer and unbalanced accounting—a significant failure.

  • Opus 5.5: 20 turns, 844,011 input tokens, 28,983 output tokens (11,038 reasoning), $1.24, 238 seconds
  • Sol: 10 turns, 270,002 input tokens, 8,205 output tokens (1,506 reasoning), $0.20, 106 seconds

Billing library

Both models achieved perfect scores of 110 out of 110. Each corrected all 20 bugs, completed the unfinished feature, and flagged the library's three incorrect tests without modifying them. Sol even identified the rule that caused its larger sibling, GPT-6.2 Astra, to fail on this identical test.

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

With identical accuracy, cost became the differentiator. Opus 5.5 cost 97 cents and required three and a half minutes. Sol cost 16 cents and took a minute and a half. The additional reasoning capacity in Opus 5.5 provided no advantage on this coding task.

  • Opus 5.5: 18 turns, 616,008 input tokens, 22,797 output tokens (11,944 reasoning), $0.97, 211 seconds
  • Sol: 14 turns, 218,317 input tokens, 6,844 output tokens (1,655 reasoning), $0.16, 94 seconds

Lab data analysis

This task most closely resembles the scientific coding work featured in OpenAI's charts, where Sol appears ahead. In my evaluation, both models tied with scores of 76 out of 76.

Cost and timing differences were narrower here. Opus 5.5 cost 58 cents to Sol's 15 cents and required approximately 50% more time. Opus 5.5 consumed significantly fewer tokens on this test compared to the other two, reducing Sol's cost advantage from roughly six times to just under four times.

  • Opus 5.5: 17 turns, 292,458 input tokens, 15,425 output tokens (3,755 reasoning), $0.58, 137 seconds
  • Sol: 14 turns, 225,538 input tokens, 6,440 output tokens (373 reasoning), $0.15, 90 seconds

Results

Sol failed to match Opus 5.5's accuracy as OpenAI's published benchmarks suggest, though the cost advantage proved largely accurate. The decision ultimately hinges on accepting slightly lower accuracy in exchange for substantially reduced expenses.

What do I think?

I would begin with Sol. It matched Opus 5.5 on coding and data analysis tasks and came within four points on the agent test, all while costing less than one-fifth as much.

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

Sol's single failure—missing a $2,400 credit—represents a critical error that warrants caution during initial deployment. Start with Sol on tasks where errors carry limited consequences, and run identical work twice to validate consistency. For applications where mistakes propagate significant downstream damage, maintain Opus 5.5 as a final verification step.