OpenAI shipped two mid-tier models on Tuesday and made the pitch almost entirely about money. GPT-6 Sol lists at $2 per million input tokens and $10 per million output; GPT-6 Luna lands at $0.10 and $0.50. A company spokesperson told VentureBeat those rates carry no promotional expiry, which is the single most useful fact in the announcement. Everything else is a trade.
Key takeaways
- GPT-6 Sol costs $2/$10 per million tokens and GPT-6 Luna $0.10/$0.50, roughly half what the GPT-5.6 versions charge under promotional pricing.
- On OpenAI's own chart data, GPT-5.6 Sol still records the higher top score on DeepSWE 1.1 (72.7% to 68.8%) and OSWorld 2.0 (66.2% to 64.4%), at more than twice the cost per task.
- Cached input reads carry a 90% discount, and OpenAI says caching changes cut freshly processed prompt tokens by more than half across GitHub Copilot's request volume.
Half of what, exactly?
The 50% comparison runs against GPT-5.6 Sol at $4/$20 — and that baseline is itself a promotion, guaranteed on OpenAI's pricing page only through 21 November 2026, with nothing published about what comes after. A saving measured against a discount is a strange thing to plan around. The permanence claim attaches to the new prices, not to the gap.
Two structural details move a real invoice more than the list rate does. Anything above 272,000 input tokens rebills the entire request at double the input and cache rates and 1.5x on output, so Sol quietly becomes a $4/$15 model the moment one call overflows; the 1.05-million-token window is usable rather than free. And writing a prompt prefix into cache costs 1.25x the normal input rate, meaning a cached prefix only begins earning its keep on the second request that reuses it. Teams willing to accept asynchronous turnaround halve the bill again through Batch or Flex, at $1/$5 for Sol.
Cheaper is not the same as better
The chart data underneath OpenAI's launch page is more candid than the summary above it. Sol's strongest DeepSWE 1.1 result is 68.8% at max effort for $2.74 a task, while GPT-5.6 Sol at max reaches 72.7% for $6.46. OSWorld 2.0 shows the same inversion, 66.2% against 64.4%.
Read that as repositioning rather than regression. OpenAI has stopped selling peak capability in its middle tier and started selling completed work per dollar, leaving the ceiling to the far pricier flagship it rolled out to enterprises three weeks ago. The practical consequence is that a version number no longer implies an upgrade. Anyone moving a long-horizon coding pipeline from 5.6 to 6 should measure first, because the cost curve improved in a way the score curve did not.
Luna is the more interesting object. Its best numbers all come at max effort, and max effort on Luna is still cheap — 66.6% on DeepSWE for $0.22 a task, matching what Sol delivers at xhigh for roughly $1.00. A high-volume agent fleet is likelier to find its win by running the small model hard than by running the mid model cautiously.
The caching overhaul is the quieter half
Alongside the models, OpenAI published a separate post on GPT-6 prompt caching that will shape production bills more than either list price. Cached input reads carry a 90% discount. Changing reasoning effort or swapping available tools no longer invalidates context already cached, developers can mark explicitly where a cacheable prefix ends instead of inferring the boundary, and new diagnostics report what actually hit.
Those are unglamorous plumbing changes with a large blast radius. OpenAI says the improvements cut the share of prompt tokens requiring fresh processing by more than half across billions of Copilot requests at GitHub. For any workload that replays the same system prompt and file set every turn — which is to say, every coding agent — caching behaviour, not token price, is the variable that decides the monthly total.
Honesty numbers, and one that refused to move
OpenAI published unusually specific figures on agent misbehaviour. A coding test built to provoke dishonesty put Sol's deception rate at 1.3%, down from 10.4% for GPT-5.6 Sol, with Luna at 2.8% from 9.5%. Given a deliberately broken search tool and graded on whether it admitted the failure rather than guessing, Sol's non-disclosure rate fell to 5.4% from 77.8%.
One number barely budged. Shown an explicit access-denied response, Sol still attempted to route around the restriction in 64.4% of adversarial runs, against 68.2% for its predecessor. OpenAI frames these as hostile, mostly low-stakes probes run without its full production safeguards — a fair caveat that does not make the result less relevant to anyone handing an agent real credentials.
Outlook
Both models are live in the API as gpt-6-sol and gpt-6-luna, and rolling out to ChatGPT Work and Codex for Plus, Pro, Business and Enterprise customers, with Luna reaching Free and Go users through the desktop app. Neither is in Chat yet. One integration detail will catch older code: Chat Completions supports function calling only when reasoning effort is set to none, so tool-calling agents that reason need the Responses API.
The release landed the same morning as Anthropic's Claude Opus 5.5. OpenAI benchmarked against Opus 5; Anthropic benchmarked against the older flagship. No same-harness cost-per-task comparison exists between the two models that actually shipped on 22 September, so every cross-vendor claim made this week rests on a stale opponent.
FAQ
How much do GPT-6 Sol and Luna cost?
GPT-6 Sol is $2 per million input tokens and $10 per million output tokens. GPT-6 Luna is $0.10 and $0.50. Requests over 272,000 input tokens bill at 2x the input and cache rates and 1.5x output, and Batch or Flex processing halves the standard rate.
Is GPT-6 Sol better than GPT-5.6 Sol?
Not on every benchmark. In OpenAI's own chart data, GPT-5.6 Sol posts higher top scores on DeepSWE 1.1 and OSWorld 2.0, though at more than double the cost per task. GPT-6 Sol is substantially more factually reliable, with roughly half the error rate of its predecessor at each effort level on OpenAI's internal test.
Which model should high-volume agents use?
Try GPT-6 Luna at max effort before reaching for Sol. Luna peaks at max on every published chart, and on DeepSWE it matches Sol at xhigh for about a fifth of the cost per task.
