Vals AI puts GPT‑6 Sol 8th on its index at half the cost
The evaluator said the model scores about level with GPT-5.6 Sol while OpenAI has halved its per-token price, and that it slips on finance, tax and law.
Vals AI said GPT-6 Sol ranks 8th on the Vals Index, about level with GPT-5.6 Sol at half the cost. The evaluator posted the results on Wednesday, a day after OpenAI released GPT-6 Sol and GPT-6 Luna. It gave no overall index score in the posts.
The figures are the evaluator's own, from runs it conducted, and the paper has not reproduced them. Vals AI said it ran the model on OpenAI's default provider settings, leaving temperature, top-p and top-k at their defaults and setting reasoning effort to maximum.
The gains are in code. Vals AI put GPT-6 Sol 7 points above GPT-5.6 Sol on its Vibe Code test and 5 points above on Code Migration. It said the model regresses on several of its knowledge-work benchmarks, among them finance, tax and law, which it said reward completeness of answer.
On BioMysteryBench, an open-source benchmark for agentic biology tasks, Vals AI placed GPT-6 Sol 4th at 74.8%, behind three models tied at 79.3%. It put the cost at $0.66 per test and called that cost-effective. The evaluator did not name the three models above it.
Vals AI gave the context window as 1 million tokens and maximum output as 128,000. It said OpenAI has cut per-token pricing by half, from $4 and $20 per million to $2 and $10. Vals AI has not published the task mix behind the index, so the ranking may not hold on a particular workload.