How much do GPT-6 Sol and Luna cost?

GPT-6 Sol is $2 per million input tokens and $10 per million output. GPT-6 Luna is $0.1 input and $0.5 output — described as the lowest input price among frontier-class models. OpenAI says these are cuts to under half of the GPT-5.6 promotional rate, and emphasises the pricing is permanent policy rather than a limited-time promotion. Cached input reads dropped roughly 90%. One important clause: if a single request exceeds 272,000 tokens, the entire request is billed at long-context rates, doubling input and cache prices and raising output by 50%.

How does GPT-6 Sol compare with Claude Opus 5.5?

They launched the same morning, with OpenAI moving about 90 minutes after Anthropic. On benchmarks both companies report with the same methodology, Opus 5.5 wins — 40.0% versus 33.2% on AutomationBench, and 54.4% versus 49.3% on FrontierCode 1.1. But cost is a different matter: Sol's per-task cost on AutomationBench is about $0.27 while scoring above Claude Opus 5 at roughly 9% of its cost. Notably, OpenAI's official comparison chart showed Opus 5 and Fable 5 rather than the same-day Opus 5.5, which critics called a selective comparison.

What are the caveats in this release?

Several. The 50% figure is anchored to GPT-5.6's promotional price, which expires November 21; measured against list price the real reduction is closer to 60%, so the headline is a favourable framing. A single request above 272,000 tokens bills the whole request at long-context rates. OpenAI's comparison chart omitted the same-day Opus 5.5. Third-party runs of identical construction tasks across the three GPT-6 models ranked Astra first, Sol second and Luna third — matching price order — and while the cheaper models were 2 to 6 times faster, every finished output had visible defects. And all leaderboard figures reflect vendor self-testing, with no third-party same-methodology evaluation available yet.

GPT-6 Sol and Luna Review 2026: OpenAI Halved Prices Permanently

Q: What are GPT-6 Sol and GPT-6 Luna?

Two new models in OpenAI's GPT-6 family, released September 22, 2026. They sit below the flagship GPT-6 Astra and use the same training approach, but are tuned for cost efficiency rather than peak capability. Sol is aimed at complex coding and agent workflows; Luna targets high-frequency, large-scale, relatively contained tasks such as summarisation and information extraction. Together with Astra they form a three-model lineup — flagship, mid-tier and lightweight. Model IDs are gpt-6-sol and gpt-6-luna.

Q: Did GPT-6 Sol actually get better than the previous generation?

On peak capability, no — it regressed. GPT-6 Sol tops out at 68.8% on DeepSWE v1.1, below the previous generation's 72.7% and Claude Opus 5's 73.7%. On OSWorld 2.0 it falls to 60.5% from 65.7%. The claimed improvement is a leftward shift of the entire cost-performance curve rather than a higher ceiling: the same score band now costs far less, and OpenAI's framing is that a fixed budget can run two to three times as many tasks. Whether that is an upgrade depends entirely on whether you are bound by peak capability or by budget.

GPT-6 Sol and Luna Review 2026: OpenAI Halved Prices Permanently — and Let Peak Scores Fall to Do It

September 24, 2026 · Updated 2026 · LLM Pricing · AI Agents · Cost Efficiency · Read time: ~11 min

Verdict first: GPT-6 Sol and GPT-6 Luna (released September 22, 2026) are the clearest signal yet that the frontier competition has moved from who is smartest to who is affordable. The pricing is the loud part: Sol at $2 input / $10 output per million tokens, Luna at $0.1 / $0.5 — the lowest input price in the frontier class — cut to under half of GPT-5.6's promotional rate and explicitly labelled permanent rather than a limited-time deal. Both carry 1.05M context and 128K output. But the quiet part is more interesting: peak capability regressed. Sol tops out at 68.8% on DeepSWE v1.1, below the previous generation's 72.7% and Claude Opus 5's 73.7%; on OSWorld 2.0 it falls to 60.5% from 65.7%. OpenAI is not claiming a better model — it is claiming a cheaper curve, arguing the same budget now runs two to three times as many tasks. That is a real and defensible trade, but it is a trade, and it should be named as one. Watch two clauses: a single request above 272K tokens bills the entire request at long-context rates, and the 50% headline is anchored to a promo price that expires November 21.

GPT-6 Sol and Luna review 2026 — OpenAI cut API prices permanently to $2/$10 and $0.1/$0.5 per million tokens while conceding peak benchmark scores

GPT-6 Sol & Luna — $2/$10 and $0.1/$0.5 per million tokens with 1.05M context, at the cost of a lower peak benchmark ceiling

What Are Sol and Luna?

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, completing a three-tier GPT-6 lineup alongside the flagship GPT-6 Astra:

Model	Positioning	Best for
GPT-6 Astra	Flagship — best in every respect	Maximum capability
GPT-6 Sol	Mid-tier, cost-tuned	Complex coding, agent workflows
GPT-6 Luna	Lightweight	High-frequency, large-scale, contained tasks — summarisation, extraction

Both new models use the same training approach as Astra but are tuned for cost efficiency rather than peak capability. They are not replacements for Astra — they are faster, cheaper siblings built on its advances. Model IDs are gpt-6-sol and gpt-6-luna.

Availability: Plus, Pro, Business, Enterprise and Edu subscribers on ChatGPT Work and Codex, plus the API. Free and Go users get Luna on desktop — but neither model has been connected to the regular ChatGPT chat interface yet. Both support text and image input, with 1.05M-token context and up to 128K output, and six adjustable inference-intensity gears.

Pricing: Permanent, Not Promotional

Model	Input (per 1M)	Output (per 1M)	vs GPT-5.6 promo
GPT-6 Sol	$2	$10	−50% (was $4/$20)
GPT-6 Luna	$0.1	$0.5	−50% (was $0.2/$1.2)

Luna's $0.1 input is described as the lowest input price in the frontier class — an entire million tokens of input for a dime. Cached input reads dropped roughly 90%.

The word that matters here is "permanent." OpenAI is explicitly framing this as policy rather than a limited-time promotion. That commitment is what makes it credible: announcing permanence publicly forecloses any quiet return to higher rates after the promotional period.

Two clauses to read before you budget. First, the 50% figure is anchored to GPT-5.6's promotional price, which expires November 21 — measured against list price, the real reduction is closer to 60%, which sounds better but means the "50%" framing was chosen conservatively while the baseline was chosen generously. Second, and more consequential at scale: if a single request exceeds 272,000 tokens, the entire request is billed at long-context rates — input and cache prices double, output rises 50%. A 1.05M context window is not 1.05M tokens of cheap context; it is cheap up to 272K and repriced above it.

The Uncomfortable Number: Peak Scores Went Down

This is the detail most launch coverage skipped, and it should be the first thing you read.

Benchmark	GPT-6 Sol	Previous gen / rival
DeepSWE v1.1	68.8%	GPT-5.6 Sol 72.7% · Claude Opus 5 73.7%
OSWorld 2.0	60.5%	GPT-5.6 Sol 65.7%

Sol is measurably worse than the model it replaces on the headline agentic benchmarks. That is not a footnote — it is the central fact of the release. Calling this an "upgrade" fails on the peak dimension, and any review that reports only the price cut is telling half the story.

What OpenAI is actually claiming is a leftward shift of the entire cost-performance curve rather than a higher ceiling. "The number of tasks that can be run with the same budget," in the company's framing, "is multiplied by two or three times."

Whether that is good news depends entirely on which constraint binds you:

If you are budget-bound — firing thousands of agent calls a day — a 2–3× throughput multiplier at a lower ceiling is straightforwardly better, and the regression is irrelevant.
If you are capability-bound — a handful of very hard tasks where the answer must be right — you have just lost ground, and the cheaper model will not help you.

One genuine capability improvement is worth noting: Sol's factual error rate is roughly half the previous generation's. So it is worse at peak reasoning while being more reliable at the level it does operate. That combination is less contradictory than it sounds — reliability and peak reasoning are different axes.

Benchmarks and the Cost Curve

Artificial Analysis Intelligence Index:

Model	Score
Claude Opus 5.5 (max)	58
GPT-6 Sol (max)	48
GPT-5.6 Sol	47
Xiaomi MiMo-V2.6-Pro (best open)	46
GPT-6 Luna (max)	37

Sol at 48 is one point above the GPT-5.6 Sol it replaces. So on this composite, the capability regression is small — the DeepSWE and OSWorld declines are offset by gains elsewhere. It is a step sideways at the top and a large step down in price.

The cost-efficiency numbers are where the release earns its case:

Metric	Result
Sol — AutomationBench (xhigh)	33.2% at ~$0.27 per task
Sol — Agents' Last Exam	56.4%
Sol — DeepSWE v1.1	68.8% — 1.1 points behind Claude Fable 5 at ~80% lower cost
Sol — OSWorld 2.0	60.5%
Luna — DeepSWE v1.1 (max)	66.6% at ~93% below Opus 5 and ~96% below Fable 5 per task

Read that Luna line again. A model priced at $0.1/$0.5 scoring 66.6% on DeepSWE v1.1 — within a few points of frontier flagships — at roughly a twentieth to a twenty-fifth of the per-task cost is the genuinely disruptive number in this release. It is not that Luna is better; it is that the floor of usable capability has dropped through the price floor.

Sol vs Opus 5.5: The Same-Morning Fight

Anthropic released Claude Opus 5.5 the same morning; OpenAI shipped Sol about 90 minutes later. That timing was not accidental, and the comparison is worth doing head-on.

On capability, where both companies report with the same methodology, Opus 5.5 wins:

Benchmark	Claude Opus 5.5	GPT-6 Sol
AutomationBench	40.0%	33.2%
FrontierCode 1.1	54.4%	49.3%

On price, Sol is in a different category. Its per-task cost on AutomationBench is about $0.27 while scoring above Claude Opus 5 at roughly 9% of its cost. Opus 5.5 is $4/$20; Sol is $2/$10 and Luna is $0.1/$0.5. The two companies are not really fighting over the same buyer.

The credibility problem: OpenAI's official comparison chart showed Opus 5 and Fable 5 rather than the same-day Opus 5.5. The company's explanation was that "Fable 5 is used as a substitute when there is no sub-item data." Whatever the internal reason, comparing against last-generation rivals while omitting the model launched that morning is a selective comparison, and it is the kind of choice that makes the rest of a benchmark table harder to take at face value.

The Strategy: Concede the Peak, Own the Default

Step back from the individual numbers and the strategy is visible — and it is a deliberate shift from discount-driven competition to what analysts are calling "positioning pricing": controlling the developer's default choice.

The release cadence tells the story: xAI's Grok 4.7 on September 21 at $2 input, Anthropic's Opus 5.5 on September 22, then Sol at the same $2 input price 90 minutes later. One product family now spans enterprise inference down to high-volume document processing, and the purchasing metric is migrating from price-per-token to cost-per-task.

Committing publicly to permanent pricing is the smart part. A promotion can be withdrawn; a public commitment to permanence cannot be reversed without a reputational cost. By saying it out loud, OpenAI has effectively foreclosed its own option to quietly raise rates later — which is exactly what makes the signal credible to the teams deciding where to build.

One gap worth noting: there is currently no GPT-6 Terra — the mid-tier name from the GPT-5.6 line has no GPT-6 counterpart. Either it arrives later, or OpenAI has decided the "Astra plus Sol/Luna" bipolar structure is sufficient.

The independent check that does exist is not flattering. Third-party runs of identical construction tasks across all three GPT-6 models ranked Astra first, Sol second, Luna third — exactly matching price order. The cheaper models were 2–6× faster, but every finished output had visible defects. That is the practical texture of buying down the cost curve: you get speed and volume, and you get defects that someone downstream has to catch.

The Honest Take

Score: 8.0/10 for cost-sensitive workloads. 5.5/10 as a capability upgrade. Both numbers are correct, and knowing which one applies to you is the entire decision.

What genuinely impresses:

The permanence commitment. Publicly forswearing the option to raise prices later is a credible signal, not marketing — it costs OpenAI something.
Luna at $0.1/$0.5 is the lowest input price in the frontier class, and it scores 66.6% on DeepSWE v1.1 — within a few points of flagships at roughly a twentieth of the per-task cost.
Sol's cost-efficiency is real: 68.8% on DeepSWE v1.1, 1.1 points behind Claude Fable 5 at about 80% lower cost; ~$0.27 per AutomationBench task while beating Opus 5 at 9% of its cost.
1.05M context and 128K output at this price point, with six adjustable inference-intensity gears.
Sol's factual error rate is roughly half the previous generation's — more reliable at the level it operates, even with a lower ceiling.
Cached input reads down ~90%, which matters enormously for repetitive agent loops.
Free and Go users get Luna on desktop — the cheap tier is not paywalled.
Same alignment technology as Astra, with lower misconduct rates than the previous generation under adversarial prompting.
Output style was tightened — less "Absolutely" filler, more direct answers.

What should temper your excitement:

Peak capability regressed. DeepSWE 68.8% versus 72.7% previously; OSWorld 60.5% versus 65.7%. This is a cheaper model, not a better one.
The 50% headline uses a promo-price baseline that expires November 21. Against list price the cut is ~60%, which means the framing was chosen to be favourable rather than precise.
The long-context clause is a trap at scale. A single request over 272K tokens reprices the entire request — input and cache doubled, output up 50%.
The comparison chart omitted the same-day Opus 5.5, substituting Opus 5 and Fable 5. On head-to-head benchmarks Opus 5.5 wins both AutomationBench and FrontierCode.
Independent testing found defects in every cheap-model output on identical construction tasks, with quality ranking in exact price order.
All leaderboard figures are vendor self-tests. As of writing there is no third-party, cross-vendor, same-methodology evaluation of Sol or Luna.
Not yet on the regular ChatGPT chat interface — access is via ChatGPT Work, Codex and the API.
The GPT-6 Terra tier is missing, leaving a gap between Sol and Astra with no announced replacement.

Who should use it: teams firing thousands of agent or batch calls a day, where a 2–3× budget multiplier matters more than the top few benchmark points; high-volume summarisation and extraction pipelines, which is exactly what Luna targets; cost-sensitive deployments that can tolerate occasional defects in exchange for 2–6× speed; and anyone building on the API who wants a credible, publicly committed long-term price.

Who should skip it: capability-bound work where the hardest few percent decides success — Sol is a regression there and Opus 5.5 or Fable 5.1 is the better call; workloads that routinely push single requests past 272K tokens, where the reprice clause bites; and production pipelines without downstream quality checks, given the observed defect rate.

FAQ

What are GPT-6 Sol and GPT-6 Luna?

Two cost-tuned models in OpenAI's GPT-6 family, released September 22, 2026, sitting below the flagship GPT-6 Astra. Sol targets complex coding and agent workflows; Luna targets high-frequency, large-scale, contained tasks like summarisation and extraction. Model IDs are gpt-6-sol and gpt-6-luna. Both use Astra's training approach but are tuned for cost rather than peak capability.

How much do they cost?

Sol: $2 per million input, $10 per million output. Luna: $0.1 input, $0.5 output — the lowest input price in the frontier class. That is under half of GPT-5.6's promotional rate, and OpenAI calls it permanent rather than promotional. Cached input reads fell about 90%. Watch the clause: any single request above 272,000 tokens bills the entire request at long-context rates, doubling input and cache and raising output 50%.

Is Sol better than the previous generation?

On peak capability, no — it regressed. DeepSWE v1.1 falls to 68.8% from 72.7%, and OSWorld 2.0 to 60.5% from 65.7%. On the Artificial Analysis Intelligence Index it scores 48 versus 47 before, so the composite is roughly flat. The claim is a cheaper curve, not a higher ceiling — the same budget runs two to three times as many tasks. Sol's factual error rate did roughly halve, so it is more reliable even at a lower peak.

How does it compare with Claude Opus 5.5?

On benchmarks where both report with the same methodology, Opus 5.5 wins: 40.0% versus 33.2% on AutomationBench and 54.4% versus 49.3% on FrontierCode 1.1. On price they are in different categories — Sol's AutomationBench cost is about $0.27 per task, beating Opus 5 at ~9% of its cost, and Luna is a fraction of that again. Note that OpenAI's official comparison chart showed Opus 5 and Fable 5 rather than the same-day Opus 5.5.

What did independent testing find?

Third-party runs of identical construction tasks across the three GPT-6 models ranked Astra first, Sol second and Luna third — exactly price order. The cheaper models were 2–6× faster, but every finished output contained visible defects. Separately, all published leaderboard figures for Sol and Luna are vendor self-tests; there is no third-party, cross-vendor, same-methodology evaluation yet.

Where can I use them?

Plus, Pro, Business, Enterprise and Edu subscribers can use both in ChatGPT Work and Codex. Free and Go users get Luna on desktop. Developers can call both via the API. Neither model has been connected to the regular ChatGPT chat interface.

Sources

All benchmark figures reported here are vendor-run; as of publication there is no third-party, cross-vendor, same-methodology evaluation of GPT-6 Sol or Luna. The third-party construction-task ranking and defect finding are reported by 36Kr. Prices and specifications change frequently, and the long-context repricing clause applies above 272K tokens per request — verify against OpenAI's official documentation before making purchasing decisions.

搜索此博客

ring