OpenAI’s half-price GPT-6 Sol and Luna aren’t a promo

OpenAI on September 22 released GPT-6 Sol and GPT-6 Luna at half the API price of their GPT-5.6 predecessors. An OpenAI spokesperson told The New Stack that this time the lower price is the default and not a promotion. Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 and $0.50.

Both models build on GPT-6 Astra. In OpenAI’s announcement, they carry much of Astra’s strength into “faster and more affordable models to support work at scale.” Astra stays the flagship above them, and for now there is no GPT-6 Terra, The New Stack noted.

The price list

GPT-5.6 Sol cost $4/$20 and GPT-5.6 Luna $0.20/$1.20, so the headline “half” understates Luna’s cut: its output price falls 58%. The GPT-5.6 Sol rate was a temporary discount to begin with. OpenAI’s model page lists it as promotional pricing, available at least through November 21. Astra reached the API at $10 in and $50 out in early September, so per token Sol costs a fifth of the flagship.

OpenAI gives the reason in its own words: “Improvements in caching and inference let us serve these models at lower cost, and we’re passing those savings directly on to users and customers.” The company also promises higher usage limits along with the lower cost.

Cost per task, against which rival?

OpenAI’s benchmark charts are built around price per task more than raw token prices. On DeepSWE v1.1, a software engineering benchmark, Sol scores 68.8% at max reasoning effort against 69.9% for Anthropic’s Fable 5 at xhigh effort, at 20% of the cost. Luna at max effort lands near Claude Opus 5 and Fable 5 at medium effort, at what OpenAI calls significantly lower cost. On Zapier’s AutomationBench, which tests models on business workflows, Luna gains 5.4 percentage points over GPT-5.6 Luna.

The gains over GPT-5.6 are mostly not extreme, The New Stack said. It also reported that, unlike with earlier releases, OpenAI did not hand over a full set of benchmark results ahead of launch.

Anthropic released Claude Opus 5.5 earlier the same day, which dates OpenAI’s comparisons against Opus 5 within hours. Opus 5.5 drops to $4/$20 from $5/$25, still twice Sol’s per-token price. Anthropic says Opus 5.5 also uses fewer tokens per task, which it puts at 40% lower costs than Opus 5 on typical workloads.

Nobody has run Sol and Opus 5.5 head to head yet. The New Stack’s reading is that Sol likely stays cheaper per task on OpenAI’s AutomationBench numbers, while Opus 5.5 posts higher scores on shared benchmarks in Anthropic’s own testing. The outlet adds that it is almost impossible to know how many tokens an agent will need to finish a task, so neither price cut makes budgeting much easier.

Caching for agents

For anyone running agents, the caching changes may matter more than the list price. OpenAI says GPT-6 gets higher prompt cache hit rates by default, with discounts of up to 90% on cached input tokens. Changing reasoning effort or the set of available tools no longer invalidates the cache, so an agent can switch effort between steps and keep its cached prefix. Explicit breakpoints let developers decide where the cached prefix ends, and a new dashboard and diagnostics tool shows what is getting cached and what isn’t.

GitHub says these improvements cut the share of prompt tokens that need fresh processing by more than half over the past several months, across billions of requests to OpenAI models. Anthropic went the same way with Opus 5.5 and cut the price of cache reads by 60% for token-billed usage, on top of the 20% per-token cut.

Answers should also get shorter. With Astra, OpenAI moved its models toward more direct replies, and it tells users to “expect to see more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall without losing substance.”

Alignment, test by test

OpenAI says both models improve on their GPT-5.6 counterparts across its alignment evaluations, including fewer misleading claims about their own coding work. On an internal coding deception test, Sol’s rate fell to 1.3% from 10.4%. Given a deliberately broken search tool and graded on whether it reported the problem or guessed, Sol failed to disclose it 5.4% of the time, down from 77.8%.

On respecting an explicit warning such as an “access denied” message, Sol barely moved: it still tried to work around the restriction in 64.4% of runs, against 68.2% for its predecessor. Luna fell further, to 42.4% from 78.5%. OpenAI says these tests mostly cover low-stakes situations and run without the system-level safeguards used in its products.

In a simulated message board seeded with unauthorized instructions, including requests to disclose private information, Sol followed them in 11.3% of the runs where it found the board, down from 51.9%. Luna and Astra followed none, though OpenAI notes Luna also found the board less often.

Anthropic makes its own case for Opus 5.5. It says the model performs best on its most comprehensive alignment test, and it names METR and Frontier Design as external testers before release. Opus 5.5 also reroutes some requests: most cybersecurity tasks go to Opus 4.8, and anything flagged by Anthropic’s biology or frontier LLM development classifiers goes to Opus 5.

Who gets it

Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and both are in the API. Free and Go users can try Luna in the desktop app.

Neither model is in Chat yet. OpenAI says it is rolling them out gradually through the day to keep the service stable, so they may not show up right away.