Understand model access, reasoning effort and what benchmark results actually tell you.
Loading cost explorer…
Official OpenAI documentation says GPT-6 Astra supports low, medium, high, xhigh and max reasoning effort. GPT-5.6 Sol supports none, low, medium, high, xhigh and max, with medium documented as the default. Low generally asks the hosted model to spend less reasoning compute; higher settings allow more. Exact quality, latency and cost effects are workload-dependent and should be measured with task-specific evaluations.
Open-weight means model parameters are downloadable under a stated license, which can enable self-hosting, inspection and fine-tuning. It does not automatically mean the training data, training code or full development process are open. Closed-weight models are accessed through company products or APIs. License terms, hardware needs and safety obligations still vary by model.
Record the model version, effort setting, tool harness, prompt, number of trials, dataset availability and whether the evaluator is internal. Vendor results can be useful, but comparisons are weak when harnesses or private datasets differ. Prefer independently reproducible public evaluations and treat internal sets as product evidence rather than universal rankings.
Try the same real task across models. Compare correctness, speed and price at a stated effort setting. A leaderboard alone cannot choose for you.
Start with reported GPU-hours. Change the rental price, then see how the GPU count changes the calendar.
Assumptions: fixed reported GPU-hours, ideal parallelism and your hypothetical hourly price. This models compute rental, not total development spending.
December 2024 · 2.788m reported H800 GPU-hours
Double the GPUs → half the modeled time. The GPU-hour bill stays $5.58m when price and total GPU-hours stay fixed.
The report counts pretraining, context extension, and post-training. Its $5.576m calculation assumes $2 per H800 GPU-hour and excludes prior research and ablations.
Time = reported GPU-hours ÷ GPUs ÷ 24. Cost = reported GPU-hours × your assumed hourly rate. Ideal parallelism keeps total GPU-hours fixed; real runs have communication overhead, capacity limits, downtime, and other costs. These historical runs do not establish a performance ranking or current frontier budget.