GPT-6 Sol and Luna Review: Frontier Performance, Budget Pricing

GPT-6 Sol and GPT-6 Luna are out after the flagship GPT-6 Astra. OpenAI says both bring Astra's training methods to cheaper tiers, and it cut their API prices in half compared with GPT-5.6 Sol and Luna.

The rate cut is visible on the price table. We wanted to know what it does to a real workload, so we priced an agentic coding session on all four models and checked OpenAI's own cost-per-task claims.

The short version is that the 50% holds exactly for Sol and runs slightly past it for Luna. Caching does more for the bill than the cut itself. Luna costs 5% of Sol per token, cheap enough that it stays cheaper even if it uses several times the tokens.

Price changes

GPT-6 Sol

GPT-6 Luna

Prices come from OpenAI's announcement and its API pricing page. We cover the full rate card for every current model in our OpenAI API pricing guide.

OpenAI labels both models "50% cheaper," but Luna's output price falls 58%. Output is the most expensive line, so output-heavy Luna workloads save more than the headline figure.

OpenAI's announcement measures the cut against GPT-5.6's promotional pricing, so the 50% is relative to a temporary discount, not a list price.

On all four models, a cached read costs 10% of regular input and a cache write costs 125%. Luna costs exactly 5% of Sol on every line (input, output, cached input, and cache write).

Three sources of savings

OpenAI's announcement points to three changes that lower the bill. Only the first shows up on the rate card.

1. The rate cut. Every token type is at least 50% cheaper, which OpenAI credits to improvements in caching and inference.
2. Better caching. OpenAI says GPT-6 gets higher cache hit rates by default. Changing reasoning effort or turning tools on and off mid-conversation no longer breaks the cache, and explicit breakpoints let you choose where a cached prefix ends. Every token that hits the cache is billed at 10% of input.
3. Shorter answers. Sol and Luna adopt Astra's communication style, which OpenAI describes as slightly shorter answers with fewer low-value details. Fewer output tokens matter because output is the most expensive line. OpenAI hasn't published a figure for the reduction.

The first follows directly from the price table. The second depends on how much of your context repeats between calls. The third depends on the task.

Example session

To see how the rate cut applies to a workload, we priced one illustrative agentic coding session on all four models. We used the same session shape as our Claude Opus 5.5 review, so the numbers compare directly.

Assumptions:

• 80 turns
• 60,000 cached input tokens per turn (repo files, tool definitions, and history)
• 1,000 fresh input tokens per turn (new instructions)
• 2,000 cache-write tokens per turn (new context added to the cache), billed at 1.25× input
• 1,000 output tokens per turn on every model. We held this constant to isolate the price change.

Across the session, that comes to 4.8M cached tokens, 80,000 input tokens, 160,000 cache-write tokens, and 80,000 output tokens.

Sol saves exactly 50%, since every line on its rate card fell by half. Luna saves 53%, and the extra comes from the 58% output cut.

On this session, caching saves more than the rate cut. Run the same session on GPT-6 Sol with no cache hits and all 5.04M input tokens bill at $2, for a total of $10.88. Caching takes that to $2.32, a 79% reduction. The rate cut saves $2.32 and caching saves $8.56.

A few factors would change these results. A short chat with little repeated context gets the 50% cut and nothing more. A model that needs fewer turns saves more, since each turn removed also removes its cached reads. Output length varies by model and effort, which we cover under Sol or Luna below.

Compared with Claude Opus 5 and 5.5

OpenAI's benchmarks compare Sol mostly against Claude Opus 5. Claude Opus 5.5 shipped about 90 minutes before Sol and Luna, so we priced the same session on both.

All four totals include a cache-write charge on the 2,000 new-context tokens per turn. Anthropic's 5-minute cache write and OpenAI's cache write both cost 1.25× input, so the sessions compare line for line.

On identical token counts, GPT-6 Sol costs 37% less than Opus 5.5 and 60% less than Opus 5. Sol and Opus 5.5 charge the same $0.20 for cached reads, so the whole gap comes from input, output, and cache writes, where Sol is half the price. The more of a workload sits in cache, the closer the two get.

This compares price, not quality. OpenAI hasn't published benchmarks against Opus 5.5, and our Opus 5.5 review assumed it uses 25% fewer output tokens than Opus 5. On that assumption, Opus 5.5 comes to $3.28, and Sol is still 29% cheaper.

OpenAI's cost-per-task claims

OpenAI frames most of its benchmarks as cost per completed task. The only dollar figure it publishes is for AutomationBench, where GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task. The other models are shown as multiples of that, so we converted them back to dollars and divided by the success rate.

Cost per successful task is the more useful figure, since a cheaper model that fails more often costs you in retries. OpenAI notes that the Fable 5.1 figure leaves out the cost of Opus 5 fallbacks, which ran on about 40% of tasks.

The other claims in the announcement:

• Agents' Last Exam. Sol at max effort scores 56.4%, above Opus 5's best result, at 60% lower cost per task.
• DeepSWE v1.1. Sol at max effort scores 68.8%, 1.1 points behind Claude Fable 5 at xhigh, at about 80% lower cost. Luna at max scores 66.6%, comparable to Opus 5 and Fable 5 at medium, at 93% and 96% lower cost.
• OSWorld 2.0. Sol at xhigh scores 60.5% against Opus 5's 60.3% at medium, at about 80% lower cost. Luna at max beats GPT-5.6 Sol at medium for a tenth of the cost.
• AutomationBench, Luna. At high effort, Luna scores 5.4 points above GPT-5.6 Luna at 58% lower cost per task.

That last figure is close to the rate cut itself. On input-heavy agent work, the price change alone accounts for about 50%, so Luna uses at most about 16% fewer tokens per task. Most of its saving is the new price, not a leaner model.

A few caveats apply. Effort levels differ from one comparison to the next, such as Sol at xhigh against Opus 5 at medium on OSWorld. OpenAI used Fable 5 scores where Fable 5.1 scores weren't available, and took competitor results from public reports rather than running them itself. None of the comparisons include Opus 5.5.

Sol or Luna

Luna costs 5% of Sol per token, so it would need 20 times Sol's tokens to cost the same. OpenAI positions Sol for complex coding and professional work, and Luna for high-volume tasks with a clear goal, such as summarizing, extracting information, and answering quick questions.

OpenAI's announcement never compares the two directly. It does give enough to estimate how much Luna's token use varies. GPT-6 Luna costs exactly 1/40 of GPT-5.6 Sol per token, so any cost ratio OpenAI reports against GPT-5.6 Sol reveals Luna's token ratio:

• Factuality. Luna matches GPT-5.6 Sol at about 1/100 of the cost, which implies about 0.4× the tokens.
• OSWorld. Luna at max beats GPT-5.6 Sol at medium at 1/10 of the cost, which implies about 4× the tokens.

Depending on the task and effort level, Luna uses anywhere from less than half to four times the tokens. Even at 4×, our example session would cost $0.46 on Luna, a fifth of Sol's $2.32.

That makes Luna at high effort worth trying before Sol on anything it can handle. Move to Sol when Luna's success rate drops, since retries cost more than the per-token savings.

Beyond price

A few changes don't show up in cost figures but affect how the models behave.

• Factuality. On OpenAI's internal evaluation, Sol makes about half as many factual mistakes as GPT-5.6 Sol. The test set is built from conversations where users flagged an error from a prior model, so it isn't representative of typical use.
• Honesty about coding work. Both models make fewer misleading claims about their own coding work than their GPT-5.6 counterparts, according to OpenAI's alignment evaluations run at maximum effort.
• Caching controls. A new dashboard shows how much input is cached, and a diagnostics tool explains missed cache hits.

Checking it yourself

Each response reports its token usage, so you can replace our fixed 1,000 output tokens per turn with real figures. Puter.js gives free access to both models without an API key, and switching between them is a one-word change.

Script tag:

<html>
<body>
  <script src="https://js.puter.com/v2/"></script>
  <script>
    const prompt = "Refactor this function to handle retries with exponential backoff";

    for (const model of ["gpt-6-sol", "gpt-6-luna"]) {
      puter.ai.chat(prompt, { model }).then(response => {
        puter.print(`${model}: ${response.message.content}<br><br>`);
        console.log(model, response.usage);
      });
    }
  </script>
</body>
</html>

npm (@heyputer/puter.js):

import { puter } from '@heyputer/puter.js';

const prompt = "Refactor this function to handle retries with exponential backoff";

for (const model of ['gpt-6-sol', 'gpt-6-luna']) {
  const response = await puter.ai.chat(prompt, { model });
  console.log(model, response.message.content);
  console.log(model, response.usage);
}

Run each task several times per model and average the output tokens, since length varies between runs. To test caching, send the same long prompt twice and compare the cached input count on the second response. If you'd rather use the official SDK, Puter also exposes an OpenAI-compatible endpoint that returns the same usage fields.

What this adds up to

OpenAI's 50% figure holds on our example session. Sol comes in at exactly 50%, since every line on its rate card fell by half. Luna comes in at 53% because of the 58% output cut.

Caching saves more than the rate cut. On a cache-heavy agent session, caching saves more than three times as much as the rate cut, which makes the new cache controls worth setting up. A short chat with little repeated context gets the 50% cut and not much else.

Against Claude, GPT-6 Sol costs 37% less than Opus 5.5 on identical tokens, with the gap coming from input, output, and cache writes. Luna costs 5% of Sol per token, so it stays a fraction of Sol's cost even when it uses several times the tokens. The pricing claim holds. The performance claim depends on your own tasks, and OpenAI hasn't yet compared either model against Opus 5.5.

FAQ

When were GPT-6 Sol and Luna released? GPT-6 Sol and GPT-6 Luna were released on September 22, 2026, following GPT-6 Astra earlier the same month.

What do GPT-6 Sol and Luna cost? GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 and $0.50. Cached input reads are 90% off on both, at $0.20 per million for Sol and $0.01 for Luna.

Are GPT-6 Sol and Luna cheaper than GPT-5.6 Sol and Luna? GPT-6 Sol and Luna are 50% cheaper than GPT-5.6 Sol and Luna on input and output, except Luna's output, which is 58% cheaper. The comparison is against GPT-5.6's promotional pricing.

What's the difference between GPT-6 Sol and Luna? GPT-6 Sol is built for complex coding and professional work. GPT-6 Luna is for high-volume tasks with a clear goal, such as summarizing and extraction, and costs 5% of Sol per token.

Where can I use GPT-6 Sol and Luna? GPT-6 Sol and Luna are available in the API as gpt-6-sol and gpt-6-luna, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users get Luna in the desktop app. Neither model is in ChatGPT's Chat yet.

Related

• OpenAI: Introducing GPT-6 Sol and Luna
• GPT-6 Astra review
• Claude Opus 5.5 review
• OpenAI API pricing: full breakdown of costs
• Free, unlimited OpenAI API with Puter.js
• Puter.js AI documentation