OpenAI Adds GPT-6 Sol and Luna, Pushing the Fight to Cost Per Task

OpenAI has expanded its GPT-6 family with two cheaper tiers, GPT-6 Sol and GPT-6 Luna, positioning them as the volume workhorses beneath GPT-6 Astra, the flagship the company says it introduced earlier this month. The pitch is not raw capability but cost per completed task: OpenAI says API prices for both models are 50% lower than the promotional pricing of their GPT-5.6 predecessors, and that improved prompt caching should cut the cost of long agent runs further. The benchmark claims are substantial, and so is the fine print attached to them.

What was announced

OpenAI says Sol and Luna were trained with methods similar to Astra, carrying over gains in professional work, factuality, coding, computer use and alignment into faster, less expensive models. Astra remains the company's top model across the board, per OpenAI, and is the recommended choice when results matter more than budget.

The published API pricing, per million tokens: Sol moves from $4 input and $20 output to $2 and $10. Luna moves from $0.20 input and $1.20 output to $0.10 and $0.50. Worth noting for anyone building a cost model: Luna's output price fell by roughly 58%, not the flat 50% the announcement describes, and the comparison baseline is explicitly GPT-5.6 promotional pricing. OpenAI does not say what the non-promotional GPT-5.6 rates were, or whether the new prices are themselves promotional and subject to change. That matters for teams writing multi-quarter budgets against these numbers.

OpenAI attributes the reductions to caching and inference improvements rather than to any single architectural change, and says it is passing the savings through. Model sizes, context windows, latency figures and rate limits were not disclosed.

The cost-per-task argument

The most interesting framing in the announcement is that OpenAI mostly reports dollars per task alongside scores, rather than scores alone.

On AutomationBench 1.0.6, which OpenAI describes as testing agents on end-to-end business workflows across 47 tools spanning sales, marketing, operations, support, finance and HR, the company reports GPT-6 Sol at "xhigh" effort scoring 33.2% at $0.27 per task. In the same table, GPT-6 Astra at low effort scores 30.3% at 3.9x Sol's cost, Claude Opus 5 at max effort scores 26.9% at 11.1x, and Claude Fable 5.1 with an Opus 5 fallback scores 31.4% at more than 8.9x. OpenAI flags that the Fable 5.1 cost understates reality because it excludes the Opus 5 fallbacks, which it says fired on roughly 40% of tasks.

On Agents' Last Exam V1, described as covering long-horizon professional tasks across 55 sub-industries, OpenAI says Sol at max effort scores 56.4%, above Claude Opus 5's best score in that evaluation at 60% lower cost per task. Opus 5's actual score is not given.

For coding, OpenAI cites DeepSWE 1.1, where Sol at max effort scores 68.8% against what it reports as Claude Fable 5's top result of 69.9% at xhigh effort, at roughly 80% lower cost per task. Luna at max effort scores 66.6%, which OpenAI calls comparable to Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost respectively. On FrontierCode 1.1 Main, which grades not just correctness but "mergeability" — test quality, scope discipline, style and adherence to codebase conventions — OpenAI says Sol improves substantially over GPT-5.6 Sol and matches Claude Fable 5.1 at xhigh effort for much less money, though no scores are published for that chart.

On computer use, OpenAI reports Sol at xhigh effort scoring 60.5% on the OSWorld 2.0 offline set against Opus 5 at medium effort at 60.3%, again at roughly 80% lower cost, while noting Astra remains its strongest model for this category.

Read the comparison methodology

OpenAI is unusually explicit about the caveats, and they deserve attention. The company says its own models were evaluated in its research environment or through its API, which can differ from production ChatGPT because of system prompts and available tools. Competitor scores were pulled from publicly available reports, and where Claude Fable 5.1 numbers were unavailable, Fable 5 numbers were substituted.

That means most of these head-to-heads are not apples-to-apples runs on identical harnesses. Effort levels also vary across comparisons — Sol at "xhigh" or "max" against competitors at "medium" or "max" — and effort settings directly drive the cost figures that make the efficiency argument work. None of this makes the claims wrong, but independent replication on neutral harnesses is the only way to confirm them.

Caching, and why it may matter more than list price

The less headline-friendly half of the announcement is prompt caching. OpenAI says GPT-6 delivers higher cache hit rates by default, with cached input-token reads discounted 90%. Developers get a prompt caching dashboard and a diagnostics tool for identifying missed cache opportunities. Two changes stand out for agent builders: reasoning effort and tool availability can now be adjusted mid-session without invalidating earlier cached context, and explicit breakpoints let developers decide where a cached prefix ends.

Anyone who has built a tool-using agent knows that flipping effort or adding a tool used to blow the cache and force a full reprocess of a long prompt. Preserving reuse across those changes is a practical win. As supporting evidence, OpenAI says GitHub reports that over the past several months these improvements cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests, helping Copilot respond faster. The report is secondhand and the baseline is unspecified.

OpenAI also offers an internal datapoint on why sustained cost matters: valued at API prices, daily token usage has passed $600 for the median OpenAI researcher and $7,000 at the 90th percentile.

Factuality, style and alignment

On an internal factuality evaluation built from de-identified conversations where users flagged errors from earlier models, OpenAI says Sol makes about half as many mistakes as its predecessor and approaches Astra-level reliability. Luna at higher effort levels is said to match GPT-5.6 Sol at roughly a hundredth of the cost. The company notes these error-inducing conversations are not representative of typical usage — a fair disclosure, but it also means the eval does not give an absolute error rate for normal traffic.

OpenAI says Astra's revised communication style carries into both models: less jargon, fewer low-value details, slightly shorter answers. The announcement illustrates this with a single side-by-side web-development exchange and concedes style is subjective.

On alignment, OpenAI says both models improve over their GPT-5.6 counterparts, including lower rates of misleading claims about their own coding work, with full results in the system card. It stresses that the evaluations deliberately test adversarial situations. Specific rates were not included in the announcement.

Availability

Sol and Luna are live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with Free and Go users getting Luna in the desktop app. Both are in the API as gpt-6-sol and gpt-6-luna. Notably, they are not yet available in Chat, and OpenAI gave no timeline for that. Rollout is staged through the day for stability.

What to watch

• Whether the new prices hold or are revealed to be promotional, as the GPT-5.6 baseline was. Teams should model both scenarios.
• Independent reproduction of the AutomationBench, DeepSWE, FrontierCode and OSWorld results on shared harnesses, especially the comparisons drawn from competitors' published reports at different effort settings.
• Real-world cache hit rates. The 90% discount on cached reads only pays off if agent traffic actually reuses prefixes; the new dashboard and diagnostics should make that measurable quickly.
• When Sol and Luna reach Chat, and whether the split between ChatGPT Work, Codex and Chat becomes a durable product boundary.
• The system card's specific alignment numbers, particularly the coding deception rates OpenAI references but does not quantify in the announcement.