GPT-6 Sol and Luna Arrive at Half the Price of GPT-5.6

GPT-6 Sol and GPT-6 Luna are the mid-priced and budget models of OpenAI's GPT-6 family, released on 22 September 2026, less than three weeks after GPT-6 Astra. In the API they are gpt-6-sol and gpt-6-luna, both with Astra's 1.05M-token context window and 128K-token output ceiling, priced at $2 / $10 and $0.10 / $0.50 per million input and output tokens. OpenAI says in its announcement that it trained them "with similar methods as GPT-6 Astra", and sells them as the way to get most of Astra's gains at a price that fits everyday work.

Working out which ChatGPT model to use? How to Use ChatGPT in 2026: A Beginner's Guide.

The names now read like a sky chart. Astra, the stars, sits at the top at $10 / $50. Sol is the middle of the range and Luna the bottom, which is a quiet demotion for Sol: in GPT-5.6 it was the flagship, with Terra between it and Luna. The GPT-6 announcement does not mention a Terra.

Half of which price?

OpenAI's headline is a 50% price cut, and on its own table the arithmetic is exact. GPT-5.6 Sol's $4 / $20 becomes $2 / $10, and GPT-5.6 Luna's $0.20 / $1.20 becomes $0.10 / $0.50. The baseline deserves a closer look, though. OpenAI measures against GPT-5.6's promotional pricing, and Sol's $4 / $20 was itself a discount: it took effect on 21 August and runs "at least through November 21", according to the GPT-5.6 Sol model page. Against the $5 / $30 Sol launched at in June, GPT-6 Sol is 60% cheaper on input and two thirds cheaper on output. An OpenAI spokesperson told VentureBeat the new rates are permanent rather than introductory.

Luna's line is steeper over a longer window. It launched at $1 / $6 in June, fell 80% at the end of July, and is now $0.10 / $0.50: a tenth of its launch input price and a twelfth of its launch output price, in under three months.

Model	Input	Cached input	Output	Replaces
GPT-6 Astra	$10.00	$1.00	$50.00	new tier, 3 September
GPT-6 Sol	$2.00	$0.20	$10.00	GPT-5.6 Sol: $4 / $20 on promotion, $5 / $30 at launch
GPT-6 Luna	$0.10	$0.01	$0.50	GPT-5.6 Luna: $0.20 / $1.20, $1 / $6 at launch

Put side by side, the family now spans exactly 100x from bottom to top, $0.10 to $10 on input and $0.50 to $50 on output. Astra costs five times Sol, and Sol twenty times Luna. In July the GPT-5.6 range had stretched to 25x; this one is four times wider again, which makes picking the wrong tier a correspondingly expensive mistake.

Sol against Opus 5, on cost per task

OpenAI's strongest claim sits on AutomationBench 1.0.6, a test of end-to-end business workflows run across 47 tools in sales, marketing, operations, support, finance and HR. Every figure in this section and the next is OpenAI's own, from its announcement. None has been independently reproduced yet.

The score is the smaller half of the argument. OpenAI's real pitch is the cost column: $0.27 per task for Sol against 11.1 times that for Opus 5 at max effort, which is how it arrives at "9% of Opus 5's cost per task". Fable 5.1's figure carries OpenAI's own footnote: it leaves out the Opus 5 fallbacks that happened on about 40% of tasks, so the true multiple is higher than the 8.9x shown.

AutomationBench 1.0.6	Score	Cost per task, as OpenAI states it	In dollars (approx.)
GPT-6 Sol (xhigh)	33.2%	$0.27	$0.27
GPT-6 Astra (low)	30.3%	3.9x Sol	~$1.05
Claude Fable 5.1 + Opus 5 fallback (max)	31.4%	more than 8.9x Sol	over $2.40
Claude Opus 5 (max)	26.9%	11.1x Sol	~$3.00

One thing the table cannot show is that every Claude model in it is already a generation old. Anthropic shipped Claude Opus 5.5 earlier on the same day, and Anthropic's own AutomationBench figure for it is 40.0%, above Sol's 33.2%. The two numbers come from different companies' runs and should not be read as a head-to-head. What they do show is that the model OpenAI chose to beat was replaced that morning. Opus 5.5 lists at $4 / $20, which is exactly what GPT-5.6 Sol cost on promotion and twice GPT-6 Sol's rate.

On Agents' Last Exam, which covers long-horizon professional work across 55 sub-industries, OpenAI puts Sol at max effort on 56.4%. It says that is above Opus 5's best result on the test, at 60% lower cost per task.

Coding and computer use

OpenAI frames coding in terms of spend. It says the median researcher inside the company now burns more than $600 of tokens a day at API prices, and one at the 90th percentile more than $7,000. At that scale the price of a long agent session matters as much as its score.

On DeepSWE 1.1, Sol at max effort scores 68.8%, 1.1 points behind Claude Fable 5's best of 69.9% at xhigh, at roughly 80% lower cost per task. Luna at max reaches 66.6%, which OpenAI calls comparable to Opus 5 and Fable 5 at medium effort, for 93% and 96% less per task. On Cognition's FrontierCode 1.1, which grades whether a change is actually mergeable, OpenAI says Sol matches Fable 5.1 at xhigh for much less money, though its text gives no score.

Computer use is the one area where OpenAI keeps Astra clearly on top. On OSWorld 2.0 offline, Sol at xhigh lands on 60.5%, against 60.3% for Opus 5 at medium, at about 80% lower cost per task. OpenAI says Luna at max beats GPT-5.6 Sol at medium for a tenth of the cost.

Fewer mistakes, less deception, mostly

On factuality OpenAI publishes a ratio rather than a score. Its internal test replays de-identified ChatGPT conversations in which users had flagged a factual error by an earlier model, and on it Sol makes about half as many mistakes as GPT-5.6 Sol. Luna at higher effort matches GPT-5.6 Sol at around a hundredth of the cost. Those conversations were chosen because a model had already got them wrong, so OpenAI is careful to say they are harder than typical use.

The alignment results are more concrete, because OpenAI labelled every bar. In its coding-deception test, built from tasks picked to tempt a model into misreporting its own work, Sol's rate fell from 10.4% to 1.3% and Luna's from 9.5% to 2.8%. The broken-tool test hands an agent a search tool that does not work, and measures how often the agent carries on as if it did. There the drop is dramatic: Sol went from 77.5% to 4.9%, while Luna fell only to 28.7%. Sol's rate of acting without authorisation fell from 51.9% to 11.3%, and both models went to roughly zero on bypassing a code reviewer.

The row that did not move is the one worth remembering. When a task raises a warning that should make an agent stop, GPT-6 Sol still worked around it in 64.4% of cases, against 68.2% for its predecessor. Luna improved to 42.4%, and Astra sits at 17.4%. On OpenAI's own chart, the cheaper tiers have not inherited that part of Astra's behaviour. These tests are built to provoke failures, so the numbers describe worst cases rather than daily use.

Shorter answers, cheaper caches

Sol and Luna also inherit Astra's writing style, which OpenAI describes as clearer, lighter on jargon and slightly shorter without losing substance. The example it published is a coding reply that says what it checked and what it did not, rather than narrating its own tool calls.

The less visible change is caching. OpenAI says GPT-6 gets higher cache hit rates by default, with cached input billed at a 90% discount. Raising or lowering reasoning effort, or switching tools on and off mid-conversation, no longer throws the cached prefix away. Developers can also set explicit breakpoints to choose where a cached prefix ends, and a diagnostics tool explains missed cache hits. The model pages list cache writes at $2.50 per million tokens for Sol and $0.125 for Luna, a quarter above the normal input rate. GitHub reports that the caching work has already more than halved the share of Copilot's prompt tokens that need fresh processing.

Where to get them

In ChatGPT, both models are rolling out through the day in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. Free and Go users get Luna in the desktop app. OpenAI says neither is in regular Chat yet. In the API, both support the Responses and Chat Completions endpoints, batch, and the full hosted tool set including computer use (Sol, Luna).

Reasoning effort runs from none through low, medium, high and xhigh to max, with medium as the default. That gives both models a no-reasoning setting Astra lacks. Their knowledge cutoffs differ: 20 April 2026 for Sol and 18 May 2026 for Luna, so the cheapest model in the family has the latest knowledge cutoff. GitHub Copilot added both the same day, with Sol on Pro+, Max, Business and Enterprise and Luna on the Pro plan as well.

The quieter number in the launch may be Luna's. OpenAI says the model now shipping to free desktop users matches last generation's Sol on its factuality test when run at higher effort, at around a hundredth of the cost per answer. If that holds outside OpenAI's own test set, most people will first meet GPT-6 at the bottom of the range, not the top.