GPT-6 Sol versus Luna API token pricing shown as a 20-to-1 scale, asking when Sol is worth it.

GPT-6 Sol and Luna: What Got Cheaper—and Which One Do You Actually Need?

by

By Si Fisher · Published September 22, 2026 at 12:45 PM PDT · Updated October 5, 2026

Quick answer

OpenAI has introduced two cheaper GPT-6 models: Luna for cost-conscious work, and Sol for more demanding jobs. API customers get lower token rates. Eligible ChatGPT subscribers can benefit too: inside Work and Codex, choosing Luna may stretch included usage farther before hitting the plan limit, even though the monthly plan price has not fallen.

The useful question is simple: can the less expensive model finish your job well enough, or will fixing its work eat the savings?

October 5 update — new Sol, same standard API sticker price: OpenAI released GPT-6.1 Sol on September 29. Its listed standard API input and output rates match original Sol ($2 and $10 per million tokens), while OpenAI reports better results on demanding tasks. For complex work, I would test 6.1 Sol first where available; Luna is still the cheaper first try for bounded jobs.

I have not run a head-to-head test of these versions. API prices do not set your Work or Codex allowance, so check the models and limits your account actually shows.

Confirmed · checked September 22, 2026 at 12:33 p.m. Pacific: Sol and Luna are rolling out in ChatGPT Work and Codex, but are not yet available in ordinary ChatGPT chat. Both are available through the API. OpenAI’s announcement.

How to make your ChatGPT subscription go further

ChatGPT plans and API usage have separate billing, so the new API prices do not reduce a monthly subscription fee. But eligible subscribers use shared Work and Codex allowances. Choosing an efficient model can let that included usage cover more work before the plan limit, potentially reducing the need for extra credits.

A useful Plus example, checked September 22: OpenAI estimates roughly 15–150 local Codex messages per five hours with GPT-6 Sol and 350–3,000 with GPT-6 Luna, compared with 10–100 for GPT-5.6 Sol. These are estimates, not fixed limits. Task size, reasoning level, context, tools, and cloud execution can change usage, and the account dashboard is the current source for remaining allowance. The 20-to-1 API price ratio is not a promise of 20 times as many subscription messages.

For included usage, start with Luna on bounded, reviewable tasks. For complex work, try GPT-6.1 Sol if your Work or Codex account offers it; otherwise consider original Sol. Compare the result and check your account’s displayed allowance. The matching API price does not establish matching subscription usage.

At the September 22 launch, Sol and Luna rolled out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users; Free and Go users could access Luna in the desktop app. OpenAI’s current Work and Codex guide says 6.1 Sol, original Sol, and Luna are separate from regular Chat. Model access depends on plan, workspace settings, and rollout.

What the new API prices mean in actual dollars

The API lets software send requests directly to a model. It charges for tokens: small pieces of text and other data. Input is what you send; output is what the model generates. Think of it as a usage meter, separate from a ChatGPT subscription.

At standard API rates, Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 input and $0.50 output. For the same quantities, Luna’s token bill is one-twentieth of Sol’s. These are the rates in OpenAI’s Sol and Luna documentation.

One substantial request

Assume 100,000 uncached input tokens and 10,000 output tokens.

Sol: $0.20 input + $0.10 output = $0.30
Luna: $0.01 input + $0.005 output = $0.015

A month’s worth of requests

Assume 10 million uncached input tokens and 1 million output tokens.

Sol: $20 input + $10 output = $30
Luna: $1 input + $0.50 output = $1.50

These are illustrative token totals, not measured workloads. They use standard rates, without discounts or tool fees, and assume no individual request exceeds 272,000 input tokens; larger requests carry higher rates. The $0.015 Luna example is 1.5 cents, versus 30 cents for Sol. Neither figure includes the cost of checking and correcting the result.

Paper arrows contrast a direct route to a finished document with a loop of review and revision; text reads Price per token is not price per task.
Token rates are one part of the bill. Finished-task cost also depends on how much context you send, how much the model returns, and how often the workflow needs another pass. Conceptual illustration.

When GPT-6 Luna makes sense

Luna is the sensible first cost test for high-volume or budget-sensitive work with a clear prompt, an objective check, and a human review path. Examples could include classifying routine requests, extracting structured fields, summarizing internal material, or producing first-pass drafts. Those are possible use cases, not TPTK test results.

The main question is whether Luna completes your actual job reliably enough. Test a representative batch, record retries and review time, and compare completed-task cost. The model with the lowest token rate can still be the expensive choice if people spend longer repairing its output.

When GPT-6.1 Sol may be worth testing

For complex, costly-to-redo work, I would test GPT-6.1 Sol first if it is available to you. Original Sol remains a useful comparison if that is the version your account offers. OpenAI reports improvements on coding, professional work, and computer use, but has not tested your particular job. Run the same representative tasks through Luna and the available Sol version, define a pass in advance, and compare accuracy, review effort, latency, and completed-task cost.

For example, a complicated software change may justify trying GPT-6.1 Sol if it reduces debugging and rework. A routine document-labeling job may not. Paying for a bigger model to do every small task is a little like booking a moving truck to collect one office chair.

What actually got cheaper

Compared with OpenAI’s listed GPT-5.6 promotional rates, Sol’s input and output rates were halved. Luna’s input rate also fell by 50%, while output dropped from $1.20 to $0.50 per million tokens: a 58.3% reduction. That is why the headline “50% cheaper” does not describe every workload’s exact saving.

What remains unknown

We have not independently benchmarked these models. OpenAI’s performance claims are vendor evidence, and your results may differ. Start with a small set of real tasks before moving an entire workflow. Rollout timing and plan limits may also change after launch.

This announcement does not establish a retirement date for GPT-5.6 or a guaranteed support period for GPT-6. If your software depends on a particular model, check its current lifecycle documentation before migrating.

Frequently asked questions

Did ChatGPT subscriptions get cheaper?

No monthly subscription price cut was announced. Eligible subscribers can still benefit: Work and Codex share included usage, and choosing Luna may allow more work before the plan limit, potentially reducing the need for extra credits. OpenAI’s message estimates are not fixed limits, and the 20-to-1 API price ratio is not a subscription-quota multiplier.

Is GPT-6 Luna always the cheaper choice?

Its standard token rates are 20 times lower than Sol’s, but the cheaper completed task depends on results. Retries, longer outputs, tool calls, and human review can outweigh the initial token-price advantage.

Can I select GPT-6 Sol or Luna in a normal ChatGPT conversation?

Not yet, according to OpenAI’s launch announcement. The first rollout covers ChatGPT Work and Codex on eligible plans, with Luna also available to Free and Go users in the desktop app. Ordinary ChatGPT chat access is still to come.

Should I choose GPT-6.1 Sol instead of original Sol?

For complex work, test GPT-6.1 Sol first if your Work or Codex account offers it, or use API model ID gpt-6.1-sol. OpenAI lists the same standard API input and output rates as original Sol and reports better results on several demanding tests. Availability and subscription allowance can vary; check your account, then compare finished results on your own tasks.

How should a business compare Sol and Luna?

Use the same representative tasks, define a pass before testing, and record total tokens, retries, latency, review time, and correction effort. Compare the cost of an acceptable finished result rather than the price of one request.

Sources and method

TPTK compared OpenAI’s September 22 launch announcement and September 29 GPT-6.1 Sol announcement with the official model pages, then calculated the original Sol/Luna examples from the published standard rates. GPT-6.1 Sol uses the same listed standard input and output rates as original Sol. Calculations exclude caching, batch or priority pricing, tool charges, retries, and human review unless stated. For corrections, email hello@theprotoolkit.com.

Leave a Comment