OpenAI's 50 percent API price cut doubles the token volume a flat budget buys

Per-task cost and per-token price are not the same measure, and the release leans on the first. Sol at xhigh effort beating Claude Opus 5 at max effort for 91 percent less per task is a cost-per-completed-task figure on AutomationBench ^[3]. It includes how many tokens each model spent reaching an answer, and at which effort setting. Luna at high effort gained 5.4 percentage points on its predecessor while cutting cost per task 58 percent, and Sol at max effort scored 56.4 percent on Agents' Last Exam at 60 percent less per task than Claude Opus 5's best score in that evaluation ^[5][6]. The 50 percent in the headline is list price ^[2].

Anyone budgeting this can start from OpenAI's own consumption. Daily token usage inside the company, valued at API prices, ran above $600 for the median researcher and above $7,000 for researchers at the 90th percentile ^[7]. Across 260 working days that is more than $156,000 a year for the median and more than $1.8m at the 90th percentile ^[14][15]. The gap between the two is about a factor of twelve ^[16].

In practice the halved list price turns into volume. A team that holds its budget flat buys roughly twice the token volume ^[17], and token volume in an agent product is tool calls. AutomationBench 1.0.6 tests agents on end-to-end workflows using 47 tools across sales, marketing, operations, support, finance and HR ^[4].

OpenAI put the cut in terms of sustained use. "Coding agents are taking on longer and more demanding tasks, making the cost of sustained use more important. GPT-6 Sol and Luna combine strong coding performance with lower API prices. This gives developers more room to iterate and allows teams to give Codex more ambitious tasks," the company said ^[8].

OpenAI reports Sol and Luna beating their GPT-5.6 counterparts on alignment evaluations, including tests that measure misleading claims about coding work ^[11]. The company also says those evaluations deliberately used situations chosen to elicit dishonesty and therefore do not represent failure rates during typical use, and that the coding-deception test ran at maximum effort ^[11]. The publication gives no normal-use rate for agent misreporting, and it leaves aside abuse of the API and who authorizes an agent's tool calls.

Every performance and cost figure in the piece is OpenAI's. The chart note on the Claude Fable 5.1 datapoint is incomplete as published ^[13].

Caching compounds the price move. Cached input-token reads take a 90 percent discount ^[9], so at prices 50 percent lower a cached read costs about 5 percent of the pre-cut uncached read ^[18]. Changes to reasoning effort and tool availability do not break the cache, and developers can set explicit breakpoints for where a cached prefix ends ^[10]. An agent can swap its tool set mid-session and keep the discount.