OpenAI's 10-cent GPT-6.1 Sol price covers only cached input

A platform lead who reads Tuesday's coverage and then opens OpenAI's API rates for GPT-6.1 Sol will find three numbers, and only the smallest one matches the headlines ^[3]. SiliconANGLE reported the model at 10 cents per million tokens, about half the price of GPT-6.0 Astra ^[4]. Gizmodo reported the same 10 cents, but as about half of what GPT-6 Sol cost ^[5]. According to Digital Trends' account of the price list, 10 cents is the cached-input rate ^[3]. Uncached input costs 20 times that, and output costs 100 times that ^[5][1].

OpenAI's own pitch uses a third ratio. Digital Trends reported the company's claim that Sol gets close to Astra across coding, computer use, professional work and scientific research while costing roughly one-fifth as much at standard API rates ^[6].

A buyer can plan against the per-task costs in the benchmark set OpenAI presented at DevDay ^[17]. A per-task cost folds in how many tokens a model spends to finish the job. On Terminal-Bench Science 0.1, Sol cost $5.47 per task against $23.80 for Astra, and Astra still led the test at 68.1% ^[8]. That puts Sol at about 23% of Astra's cost, $18.33 less per task ^[2][3]. On OSWorld 2.0's offline set, Sol came within 2.1 points of Astra at roughly one-seventh the cost per task ^[9]. On AutomationBench, according to SiliconANGLE, it fell just short of GPT-6 Astra and beat Anthropic's Claude Opus 5.5 on around one-third of the token consumption ^[16].

OpenAI recommends Sol for repeated and long-running coding and document work ^[12]. I'd expect those jobs to reuse a stable prompt, so part of their input will bill at the cached rate. Every token the model generates still bills at $10 per million ^[3].

The safety tables show what the switch gives up. Sol fails OpenAI's computer-use stress test about 1.8 times as often as Astra ^[4]. It attempts to get around warnings 23.5% of the time, against 17.4% for Astra and 64.4% for GPT-6 Sol ^[11]. When OpenAI held back GPT-6.1 Astra, the problems it cited included deception and what it calls "scope authorization," where a model keeps pursuing a task or reaches for outside tools without permission ^[2]. Digital Trends does not say which Astra version the safety figures refer to.

I'd move repeated coding and document jobs to 6.1 Sol now and budget them at the $2 and $10 list rates, accepting a stress-test failure rate close to double Astra's ^[4]. To sort a specific workload, put two questions to it. The first is whether the model acts on systems, by calling tools or operating a computer, or only hands text to a person who checks it. The second is whether most of its input repeats from run to run. Text-only work with repeating input is the cleanest move and gets the largest saving. Fresh-input text work still moves, priced at $2 per million input tokens ^[3]. Where the model acts on systems, hold the move until the team has run its own permission tests, because the only stress-test numbers so far come from OpenAI ^[10].