GPT-6 Sol vs Claude Opus 5.5: Cost per Task and Benchmarks

OpenAI released GPT-6 Sol and Anthropic released Claude Opus 5.5 on the same day, September 22, 2026. Sol lists at $2 per million input tokens and $10 per million output tokens, half of Opus 5.5’s $4 and $20. Neither vendor compared the two: OpenAI’s charts show Opus 5, and Anthropic’s table shows GPT-5.6 Sol.

The independent evaluator Artificial Analysis has now run both models on the same ten tests at every effort level. On its index, Sol is the cheaper way to reach any score up to about 44. Above that, Sol runs out of headroom. Its best result, 47.5 at max for $1.06 a task, is below Opus 5.5 at its default medium setting, 51.2 for $1.34.

Every Artificial Analysis figure below was read from its GPT-6 Sol and Claude Opus 5.5 model pages on September 22, Intelligence Index version 4.3.2. It measures cost per task at list prices, including cache reads and writes.

The Intelligence Index averages ten evaluations: agentic coding, business workflows, knowledge-work documents, science, long-context reasoning and factual knowledge. Both models default to medium, and both offer low through max. Each cell shows the index score and the average cost per task.

Sorted by cost, the two models barely overlap. Every Sol setting below max costs less than the cheapest Opus 5.5 setting, and Sol at xhigh (44.1 for $0.53) already beats Opus 5.5 at low (42.3 for $0.55). Everything Opus 5.5 does from medium upwards scores higher than anything Sol can reach. The choice is less “which model is better” than “which price band does this task belong in.”

One result goes against intuition. Sol with reasoning switched off scored 28.1 for $0.33 a task, below Sol at low (33.9) and more than twice as expensive. Almost all of the extra cost is input. One likely reason is that a model without reasoning takes more steps to finish, and each step re-sends the conversation. Turning reasoning off is not a reliable way to save money on agent work.

The fairest pairing is each model at the setting a team is most likely to use for demanding work: Opus 5.5 at its default medium, against Sol at max, which costs about the same. Sol at xhigh is added as the cost-conscious option.

Coding is the widest gap. On Terminal-Bench 4.0, Opus 5.5 at medium scores 52.5% against Sol’s 43.9% at max, and Opus 5.5 reaches 59.6% at xhigh. Sol falls away fast at lower settings: 30.3% at xhigh and 26.3% at high. For coding agents that run in a terminal, Sol’s lower price buys noticeably weaker results.

Knowledge work also favours Opus 5.5. On GDPval-AA and AA-Briefcase, which rate work documents in head-to-head comparisons scored as Elo, Opus 5.5 at medium leads Sol at max by about 90 and 160 Elo points. At max, Opus 5.5 reaches 1846 on GDPval-AA.

Business workflows are level. On AutomationBench-AA, Artificial Analysis’s partial-credit run of business tasks across apps, Sol at xhigh (61.7%) matches Opus 5.5 at medium (61.2%) at 40% of the average cost per task. Long-context reasoning is also level, and at these settings Sol leads on CritPt, a physics research test.

Each vendor published results for the benchmarks both models share. They point the same way as the independent data, but they are not a clean comparison.

• AutomationBench (Zapier). Anthropic reports Opus 5.5 at 40.0% at max, from Zapier’s early-access run. OpenAI’s chart puts Sol’s best at 33.2%, at xhigh. These are Zapier’s own scores, which run far lower than Artificial Analysis’s partial-credit version in the table above.
• FrontierCode 1.1 Main (Cognition). Anthropic reports Opus 5.5 at 54.4%; OpenAI’s chart puts Sol at 49.3%, both at max. But the two vendors disagree about the same model: Anthropic lists Opus 5 at 48.0%, OpenAI’s chart at 53.4%. A five-point gap between vendors’ runs is larger than the gap between these two models, so treat this row as inconclusive.
• OSWorld 2.0. Anthropic reports 81.8% on the partial-credit score; OpenAI reports 64.4% for Sol on the offline set. These are different task sets and cannot be compared.

Each launch also claims cost wins, but against other models: OpenAI against Opus 5, Anthropic against GPT-6 Astra and GPT-5.6 Sol. Our GPT-6 Sol and Luna launch analysis and Opus 5.5 launch analysis read those claims against each vendor’s own chart data.

Two lines break the “half price” rule. Cache reads cost $0.20 per million tokens on both models, and for agents, cache reads are usually the largest share of the bill. And Anthropic charges the full 1M-token window at standard rates, while OpenAI bills any request over 272K input tokens at twice the input and cache rates and 1.5 times the output rate.

The table below reuses the illustrative agent task from our Opus 5.5 launch post. The token counts are invented for the example and kept the same for both models; the rates are the published ones.

At equal token counts, Sol costs 57% of Opus 5.5 on this cache-heavy task, not 50%. If every request goes past 272K input tokens, Sol costs slightly more than Opus 5.5. Real token counts differ between the models, and Artificial Analysis’s measured figures already include that difference. At their defaults, Sol costs $0.25 a task against $1.34 for Opus 5.5, about a fifth, but for an index score of 39.8 against 51.2.

The cache behaviour matters most for routers. A system that raises effort on hard turns and lowers it on easy ones keeps its cache on Sol. On Opus 5.5 it needs the per-message effort beta, or it pays to rebuild the cache after every change. Opus 5.5 also routes most cybersecurity tasks to Opus 4.8 behind the scenes, so security teams should test that workload directly before choosing.