GPT-6.1 Sol delivers Astra-level accuracy for one-fifth the cost

OpenAI introduced GPT-6.1 Sol in late September as a cost-efficient alternative to its premium GPT-6 Astra model, claiming comparable performance at a fraction of the price. The new offering slashes both input and output token costs to $2 and $10 per million respectively—five times cheaper than Astra's $10 and $50—while halving cached input pricing to $0.10.

To verify whether GPT-6.1 Sol genuinely delivers on these promises, the model was put through three demanding real-world scenarios alongside Astra. Both systems faced identical prompts, maximum reasoning effort, and a 64,000-token output ceiling.

Testing on production workloads#

The evaluation encompassed three representative tasks: CI triage required triaging 40 failed CI job logs using conditional rules; incident logs demanded precise analysis of 3,664 log lines from a simulated outage, including exact counting and timezone conversions; and resolver spec asked for writing a dependency resolver against a fictional package manager spec, later graded by 120 hidden test cases.

On CI triage, both models achieved perfect accuracy across all five runs. Astra completed runs in an average 24 seconds compared to GPT-6.1 Sol's 31 seconds, though token usage remained nearly identical at around 3,867 input tokens and roughly 1,400 output tokens. The cost gap emerged sharply: 2 cents per run for GPT-6.1 Sol versus 11 cents for Astra.

The incident logs test illustrated where GPT-6.1 Sol gained ground. Both models answered all seven questions flawlessly on every attempt, but GPT-6.1 Sol completed runs 19% faster (2 minutes 20 seconds versus 2 minutes 53 seconds) while consuming 113,966 input tokens and roughly 8,300 output tokens. Crucially, each run cost just $0.31 for GPT-6.1 Sol against $1.57 for Astra—over five times cheaper.

The resolver spec test showed the widest performance separation. GPT-6.1 Sol again passed all 120 hidden tests, finishing 30% faster than Astra (7 minutes 7 seconds versus 10 minutes 6 seconds). Astra generated 28% more output tokens (25,207 versus 19,637), pushing per-run costs to $1.28 against GPT-6.1 Sol's $0.20.

Bottom line#

Combining all three tests across 15 total runs, GPT-6.1 Sol maintained perfect accuracy while costing just $2.66 total—18% of Astra's $14.77. Execution time favored the cheaper model on longer tasks. The results align closely with OpenAI's published pricing claims, suggesting the gap between the models may be narrower than performance tiers suggest for many applications.