OpenAI Releases GPT-6 Sol and Luna - API Rates Half of GPT-5.6's Promotional Pricing, and Caching Breaks Differently

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. The two models follow GPT-6 Astra from earlier in the month, and the company says it trained them with methods similar to Astra’s, carrying the gains in professional work, factuality, coding, computer use, and alignment over to “faster, more affordable models”¹. It also restates that Astra remains its best model overall and is the one to pick when you do not want to compromise.

The per-token rates draw the eye, but there is a second change that lands on the actual bill: the default behavior of prompt caching.

The Baseline for “Half Price” Was Itself a Time-Limited Rate

The official pricing table reads as follows, per million tokens¹.

Model	Input	Output
GPT-6 Sol (vs. GPT-5.6 Sol)	$4 → $2	$20 → $10
GPT-6 Luna (vs. GPT-5.6 Luna)	$0.20 → $0.10	$1.20 → $0.50

OpenAI writes that gains in caching and inference let it serve these models more cheaply, and that it is handing the savings back as a 50% reduction measured against GPT-5.6’s promotional pricing¹. The table labels both rows “50% cheaper,” though Luna’s output moves from $1.20 to $0.50, which works out to roughly 58%.

The phrase “promotional pricing” deserves a second look. The GPT-5.6 Sol baseline of $4 input and $20 output came from the price cut announced on August 21, and at the time OpenAI stated it was a rate offered at least through November 21, 2026. So this round of halving is measured against a rate that had an end date attached.

What about the new rates themselves? The announcement post says nothing about a term. VentureBeat reports on this point that an OpenAI spokesperson confirmed to the publication that the Sol and Luna rates are permanent prices, not promotional or introductory ones³. That is not something the announcement page alone settles, so it is worth knowing if you are folding these rates into a longer-range estimate.

Astra does not appear in this pricing table. Its standard rate at launch was $10 input and $50 output, so lined up against that figure, Sol sits at a fifth and Luna at a hundredth.

The Conditions That Break Your Cache Have Changed

If you are running agents, this may matter more than the per-token rate. OpenAI says it improved prompt caching for GPT-6 and raised default hit rates¹. Reads of cached input tokens carry a 90% discount.

Two of the changes bear directly on implementation:

• Changing reasoning effort, and enabling or disabling tools, no longer breaks the cache. Raise effort for a hard task and lower it for a simple follow-up, or add and remove tools as an agent’s needs shift — earlier context stays eligible for cache reuse through all of it
• Explicit breakpoints let developers pick where the cached prefix ends¹

Measurement tooling came with it. The Prompt Caching Dashboard shows how much of your input is cached and how that moves over time, and a diagnostics tool points at what caching you missed and how to fix it. If you had been pinning effort to a single value to avoid breaking the cache, that constraint is worth revisiting. Whether you assemble agents yourself or rent the Codex harness through something like the Agents API, where prompt reuse stops is where the bill starts.

OpenAI relays a report from GitHub that these caching improvements have, over the past several months and across billions of requests, cut the share of prompt tokens needing fresh processing by more than half¹.

Read the Benchmarks as Vendor-Published Numbers

The announcement page lines up scores from AutomationBench, Agents’ Last Exam, DeepSWE, and OSWorld. On DeepSWE v1.1, for instance, GPT-6 Sol at max effort scores 68.8%, placing it 1.1 points behind Claude Fable 5’s top result in that evaluation of 69.9% at xhigh effort, at roughly 80% lower cost per task¹. GPT-6 Luna at max reaches 66.6%, which the company equates to Claude Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost respectively.

The measurement caveats sit at the foot of the page. Competitor scores were drawn from publicly available reports, and where a Claude Fable 5.1 score was unavailable, a Fable 5 score was used instead¹. There is also a note that GPT evaluations ran in OpenAI’s research environment or through its API, which can produce slightly different output from production ChatGPT because of differences in system prompts and available tools. Anthropic publishes AutomationBench results of its own for Opus 5.5, run by Zapier, so the two companies’ figures were not produced under matched conditions. These are not numbers you can line up side by side and settle a ranking with.

The factuality claim carries a similar hedge. An internal evaluation puts GPT-6 Sol at about half the errors of its predecessor, but the company writes that the evaluation was built from de-identified conversations in which users had flagged a mistake, and does not represent typical usage.

It Has Not Reached Chat Yet

Availability cuts differently depending on where you look. GPT-6 Sol and Luna reached ChatGPT Work and Codex on launch day for all Plus, Pro, Business, Enterprise, and Edu users¹. Free and Go users get Luna only, in the desktop app. And Chat does not have them yet. The ChatGPT rollout was set to proceed gradually through the day, with a note to check back later if the models are not visible. The API identifiers are gpt-6-sol and gpt-6-luna.

In GitHub Copilot the eligible plans differ between the two models. GPT-6 Sol covers Pro+, Max, Business, and Enterprise; GPT-6 Luna adds Pro to that list². Both are picked from the model picker in VS Code, Visual Studio, Copilot CLI, the cloud agent, JetBrains, and other clients, with usage-based billing. For Enterprise and Business administrators, the default is the part to watch: a new model turns itself on unless an administrator has already switched off the global default or explicitly blocked that model — and the same settings screen is where six models are queued for retirement on October 19. Housekeeping on the old generation and additions of the new one are running in parallel, so the less an organization touches its policy, the more its model list changes underneath it.

Sources

1. Introducing GPT-6 Sol and Luna - OpenAI official announcement (September 22, 2026)
2. OpenAI’s GPT-6 Sol and GPT-6 Luna now available - GitHub Changelog (September 22, 2026)
3. OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more - VentureBeat (September 22, 2026; reports OpenAI’s confirmation that the rates are permanent)