OpenAI is positioning GPT-6 Sol and GPT-6 Luna as more affordable ways to use the GPT-6 family for everyday professional work. The company says the two models apply methods similar to those used for GPT-6 Astra to areas including coding, factuality, computer use, and agent alignment. That is OpenAI’s product claim, not an independent finding; the announcement does not establish that either model matches Astra across all tasks.
Placed after the introduction of Sol and Luna and the article's distinction between OpenAI's product claims and independent findings. Source: OpenAI.
The most concrete change is price. OpenAI lists Sol’s API input and output prices at $2 and $10 per million tokens, down from $4 and $20 for GPT-5.6 Sol. Luna falls from $0.20 and $1.20 to $0.10 and $0.50. Those are nominal API prices, not guaranteed application costs: caching, retries, tool calls, and the amount of reasoning used can change the final bill. OpenAI also says cached-input reads receive a 90% discount and that newer caching behavior preserves context when reasoning effort or tool availability changes.
Placed after the discussion of caching discounts, preserved context, cache dashboards, and reduced fresh-token processing. Source: OpenAI.
That caching point matters for agents and long-running coding sessions, where the same instructions and repository context may be reused across many turns. OpenAI says developers can inspect cache usage in a dashboard, diagnose missed opportunities, and choose explicit cache-breakpoint prefixes. It also cites GitHub’s report that recent improvements reduced the share of prompt tokens needing fresh processing by more than 50% across billions of requests to OpenAI models. That GitHub statement is a separate operational report, but it does not independently prove OpenAI’s model-quality claims or guarantee the same savings for every application.
The strongest reader question is not whether a headline score is higher, but whether a model completes a real workflow reliably enough to justify its cost. OpenAI points to AutomationBench, where GPT-6 Sol at xhigh effort scored 33.2% at $0.27 per task. Zapier’s independently operated benchmark page lists the same score and cost, and describes a held-out, end-to-end test spanning 47 simulated tools across sales, marketing, operations, support, finance, and HR. It says final environment states are checked against fixed criteria rather than judged by another language model.
The comparison is still narrower than a general verdict. AutomationBench’s strict metric requires every scored assertion to pass, and its page reports run-to-run variance typically within 1%. It also notes that a Claude Fable 5.1 result with an Opus 5 fallback has an understated displayed cost because the fallback handled roughly 40% of tasks. Readers should therefore treat the benchmark as evidence about a particular class of multi-application workflows, not as proof of workplace productivity or universal superiority.
Coding results require similar care. OpenAI says GPT-6 Sol scored 68.8% on DeepSWE v1.1 at maximum effort, 1.1 percentage points behind the highest Claude Fable 5 result, at roughly 80% lower cost per task. It reports GPT-6 Luna at 66.6%, with much lower per-task costs than the cited Claude configurations. These figures compare selected effort levels and benchmark-reported costs; they do not show that every repository, language, or engineering team will see the same ranking. FrontierCode, operated by Cognition, independently describes its own mergeability standard—correctness, tests, scope discipline, style, and adherence to repository conventions—and says solution-bearing internet use is detected and zeroed. Its September 22 changelog records the addition of GPT-6 Sol and Luna, corroborating that the models entered that benchmark ecosystem, but the supplied FrontierCode text does not provide the complete Sol-versus-Luna leaderboard numbers needed to independently reproduce OpenAI’s coding comparison.
OpenAI also reports gains in factual reliability and computer use, while explicitly warning that its internal factuality test uses conversations in which users had already flagged an error. Those conversations are not representative of ordinary use, where the company says errors are rarer. Its OSWorld 2.0 offline result compares Sol at xhigh effort with Claude Opus 5 at medium effort, so the apparent cost advantage is tied to the chosen settings and benchmark version rather than a blanket statement about all computer-use tasks.
Availability is another practical constraint. OpenAI says Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while Free and Go users can access Luna in the desktop app. It says the models were not yet available in Chat and planned a gradual rollout for stability. API users are directed to the gpt-6-sol and gpt-6-luna model names. A buyer evaluating a production system should test the actual product surface, usage limits, latency, and total cost rather than infer them from an announcement or a leaderboard.
The evidence supports a measured conclusion: Sol and Luna make lower-priced GPT-6 tiers available, and at least some benchmark results and caching claims are documented beyond OpenAI’s announcement. It does not establish that either model is universally better, that benchmark savings translate directly into lower bills, or that performance on simulated workflows predicts success in a particular organization. The useful next test is a small, representative workload with fixed success criteria, repeated runs, and a cost calculation that includes cache hits, failures, and retries.