GPT-6 Sol and Luna: OpenAI Cuts API Prices in Half

A Three-Tier Family

In OpenAI's messaging, Astra remains the pinnacle: the smartest, most aligned model, the one you pick when you want zero compromise. Sol and Luna don't claim to replace it, but rather to occupy the two tiers below, where most of the day-to-day work happens.

The central argument isn't "smarter" but "same generation, cheaper." OpenAI claims to have trained Sol and Luna using methods similar to Astra's, carrying over its advances in professional work, factuality, code, computer use, and alignment into faster models.

The language used is telling: the company talks about "pushing the frontier on cost efficiency." The competition is no longer just about the maximum achievable score, but about where a model sits along the cost/intelligence curve.

The War of Cost Per Task

The price cut is significant: $2 for input and $10 for output per million tokens for Sol, $0.10 and $0.50 for Luna — half the promotional rate of the corresponding GPT-5.6 models. OpenAI attributes this margin to infrastructure gains in caching and inference, passed on to customers.

But the metric highlighted throughout the post isn't the token price — it's cost per task. On AutomationBench, a test of business workflows involving 47 tools across sales, marketing, operations, support, finance, and HR, Sol at xhigh effort scores 33.2% for $0.27 per task, while Claude Opus 5 at max caps out at 26.9% for 11.1 times that cost.

OpenAI even notes that Claude Fable 5.1's score understates its real cost, since it omits fallback escalations to Opus 5, which occurred on roughly 40% of tasks. This is competitive comparison in full swing, complete with the usual level of methodological detail — and the standard caveat that competitor scores come from public reports.

A tasty detail: Sol at xhigh even beats Astra at low effort (33.2% versus 30.3%), for nearly four times less money. The model hierarchy is no longer linear — it depends on the chosen effort level.

Code, Factuality, Computer Use

On code, OpenAI justifies its logic with its own internal numbers: valued at API prices, daily token consumption exceeds $600 for the median researcher and $7,000 at the 90th percentile. When coding agents work for long stretches, the cost of sustained use becomes the main constraint.

Results follow this logic: on DeepSWE v1.1, Sol at max hits 68.8%, just 1.1 points behind Claude Fable 5's best score in the evaluation, for about 80% less per task. Luna at max scores 66.6%, comparable to Opus 5 and Fable 5 at medium effort, but 93% and 96% cheaper respectively.

On factuality, OpenAI evaluates its models on de-identified ChatGPT conversations where users had flagged an error. Sol makes roughly half as many errors as its predecessor there; Luna at high effort matches GPT-5.6 Sol for about one-hundredth of its cost. The company notes that these conversations, chosen because they induce errors, don't reflect typical usage.

On computer use, Astra is still presented as the best model in the world. Sol at xhigh scores 60.5% on OSWorld 2.0 offline, tying Opus 5 at medium (60.3%), for about 80% less cost per task.

Caching, the Agents' Quiet Weapon

The least flashy part of the post may be the most structurally important. OpenAI announces better default cache hit rates on GPT-6, with a 90% discount on reading cached input tokens — a decisive lever when an agent re-injects the same context on every turn.

Three new tools support developers: a Prompt Caching Dashboard to track the share of input being cached, a diagnostic tool explaining missed opportunities, and explicit breakpoints to choose where the cached prefix ends. Most importantly, changing reasoning effort or toggling tools on/off no longer breaks the cache.

Social proof comes from GitHub: over recent months, these improvements reportedly cut the share of prompt tokens requiring fresh processing by more than 50%, across billions of requests, speeding up Copilot's responses.

Style, Alignment, and Availability

OpenAI also carried Astra's communication style over to Sol and Luna: more clarity, less jargon, fewer strange turns of phrase, less low-value detail, and slightly shorter responses. The company illustrates this with a React front-end prompt example, explaining why it prefers GPT-6 Sol's answer — clearer about what has and hasn't been verified, less chatty about its own tools.

On alignment, both models improve over their GPT-5.6 equivalents, notably on the rate of misleading claims about their own coding work. OpenAI reminds readers that these evaluations deliberately test difficult situations and don't measure failure rates under normal use; the full results are in the system card.

Rollout is immediate but partial: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu, with Luna in the desktop app for Free and Go. The models aren't yet available in Chat, and rollout is staggered throughout the day to preserve service stability.