GPT-6 Sol and Luna Halve OpenAI's Mid-Tier Token Price

What changed on 22 September

OpenAI expanded the GPT-6 family with two new models, arriving after GPT-6 Astra landed earlier in September. The additions are not a flagship and a research preview. They are the two tiers where the overwhelming majority of production traffic actually runs: the workhorse mid tier and the cheap tier underneath it.

• GPT-6 Sol — $2 per million input tokens, $10 per million output tokens. Its predecessor, GPT-5.6 Sol, was $4 and $20.
• GPT-6 Luna — $0.10 per million input tokens, $0.50 per million output tokens. Its predecessor was $0.20 and $1.20.
• These are permanent prices. An OpenAI spokesperson confirmed to VentureBeat that neither rate is promotional or introductory.
• API identifiers are gpt-6-sol and gpt-6-luna.
• Availability covers ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers. Free and Go users get GPT-6 Luna through the desktop application.
• Capability claims cited by OpenAI cover coding, professional work, computer use, factual accuracy and autonomous agents. No verified figures were attached to any of those claims, and we are not going to invent any.

The pricing table, with the caveat attached

Here is the full comparison. Read the last column, then read the note underneath it, because the note is the part most coverage will skip.

That caveat cuts in an unexpectedly favourable direction for planning. A promotional rate is a rate you cannot build a twelve-month forecast on, because it can revert. A permanent rate is one you can. The durability is arguably worth more to a finance team than the percentage is, and it is the part of this announcement that deserves to survive into your next budget paper.

Why the mid tier is the tier that matters

Flagship launches take the headlines. Mid-tier price moves take the budget. In most production systems the flagship handles a thin slice of genuinely hard requests, the cheap tier absorbs classification, extraction, routing decisions and short rewrites, and the mid tier carries everything in between — which in practice is most of the volume and most of the spend.

So a halving at the mid tier is not a 50% saving on a line item. For a workload that sits mainly on Sol, it is close to a halving of the whole bill. That is a materially different event from a flagship price move, and it lands on teams that had already done the hard work of demoting traffic off the frontier model. We wrote up the frontier end of the same family in GPT-6 Astra's real price and its 272K context cliff; this release is the part of the ladder underneath it, and it is the part your invoice notices.

Luna's steeper cut is the interesting number

Sol fell by a flat 50% on both sides. Luna fell by 50% on input and roughly 58% on output. The asymmetry is small in percentage terms and significant in consequence, because of what it does to the gap between the two tiers.

Work it through on output tokens, which is where the money usually is. Before: $1.20 for Luna, $20.00 for Sol — a gap of $18.80 per million. After: $0.50 for Luna, $10.00 for Sol — a gap of $9.50 per million. In absolute terms, the distance between the cheap tier and the mid tier has roughly halved.

Be precise about what that does and does not mean. As a ratio, Sol is now 20 times Luna's output price, up from about 16.7 times, so the cheap tier is relatively cheaper than it was. But a cascade is not tuned on ratios. It is tuned on the absolute saving you earn by keeping a request on the cheap model — and that saving has halved. The cost of being wrong and escalating has fallen by the same amount.

What that does to an escalation threshold

A cascade works by running the cheap model first, scoring its confidence or verifying its answer, and escalating when the score falls below a threshold. That threshold encodes a trade: how much quality risk you accept in exchange for how much money saved. When the money saved per avoided escalation halves, the same threshold is buying you less, and the rational setting moves — you should escalate more readily than you did, because escalation now costs half what it did.

Put bluntly: a cascade tuned at the old prices is now mis-tuned in the conservative direction. It is holding requests on the cheap model to protect a saving that is no longer worth what it was. The mechanics of setting the threshold in the first place are covered in model routing and cascades, and the measurement discipline for checking whether the cheap model deserved the request is in evaluating your router. Both are worth re-reading with the new numbers in hand.

An illustrative monthly cost, with assumed volumes

The figures below are illustrative. The volumes are assumed for the purposes of showing the arithmetic; they are not measured, not typical of any particular company, and not a forecast. Substitute your own numbers before you quote anything to a finance team.

Assumed monthly volume: 500 million input tokens and 100 million output tokens.

Now read the third row against the first. At the old prices, running a cascade instead of putting everything on the mid tier saved about $2,646 a month on these assumed volumes. At the new prices, the same cascade saves about $1,330. The cascade still pays — it is still roughly a third of the all-Sol bill — but the prize for operating it has halved, while the engineering cost of maintaining a router, a confidence scorer and two sets of evals has not moved at all.

That is the uncomfortable question this release puts to small teams. If your cascade was saving a few hundred pounds or a few tens of thousands of rupees a month and costing you a day of engineering attention a fortnight, the maths may no longer favour it. Simplicity has a price too, and the price of simplicity just fell. The framework for deciding is cost per successful task rather than cost per token, which we set out in LLM unit economics.

The same-week competitive backdrop

OpenAI was not moving alone. Anthropic released Claude Opus 5.5 in the same week at $4 per million input and $20 per million output, down from Opus 5's $5 and $25. Two frontier labs cutting prices within days of each other is a market signal rather than a coincidence, and it is a signal about where competition is now happening: not on a benchmark leaderboard, but on the per-million rate that appears on a monthly invoice.

Resist the temptation to read the two announcements as a like-for-like race. Opus 5.5 at $4 and $20 and GPT-6 Sol at $2 and $10 are not competing for the same request; they sit at different points on the capability ladder, and comparing their headline rates without comparing what each actually completes for you is exactly the error that makes a spreadsheet look decisive and a production system look expensive.

What to do this week

Four moves, in the order that pays back fastest.

1. Replay a week of real traffic at the new rates. Not a sample, not a synthetic benchmark. Your own logged requests, with their actual token counts, priced at $2/$10 and $0.10/$0.50. You will get the honest number in under an hour, and it will be different from the one you guessed.
2. Recompute the escalation threshold, do not just lower it by instinct. For each candidate threshold, measure two things on the replayed set: what fraction of requests escalate, and what the quality cost is of the ones that do not. Then price both at the new rates. The threshold that was optimal at $18.80 of avoided output cost per million is not the threshold that is optimal at $9.50.
3. Re-examine whether the cascade still earns its complexity. Compare three options on the same replayed traffic: all mid tier, all cheap tier, and the cascade. If the cascade's advantage over all-Sol has shrunk below the cost of maintaining it, simplifying is a legitimate engineering decision rather than a defeat.
4. Update the forecast you gave finance. A committed spend estimate built on the old rates is now wrong in your favour, which is the pleasant kind of wrong — but only if someone tells the budget holder. The permanence of the new rates is the part worth flagging, because it is what makes the revised forecast defensible. The method is in forecasting your LLM bill before launch.

What this means for builders in India and the UK

For Indian teams, two specifics matter beyond the headline. First, per-token pricing interacts with tokenisation, and Indic-script prompts consume more tokens per unit of meaning than English ones do, so the saving measured in rupees will not track the headline percentage precisely — measure it, do not assume it. Second, a permanent rate rather than a promotional one is what makes a fixed-price delivery contract safe to sign, which matters to the services and GCC teams who quote a per-transaction price to a client and absorb the model cost themselves.

For UK teams, the availability list is the operational detail to check first. ChatGPT Work and Codex access for Plus, Pro, Business, Enterprise and Edu subscribers is a procurement fact as much as a technical one, and a change of model tier inside an enterprise agreement is usually a conversation with someone in IT before it is a conversation with anyone in engineering. Data residency and processing terms do not change because a price did, so confirm nothing in your data protection paperwork is tied to a specific model identifier before you swap gpt-5-6-sol for gpt-6-sol in production.

Both markets share the same underlying lesson, and it is not a new one. Model prices at the mid tier have moved repeatedly this year, and every move invalidates a set of thresholds somewhere. The teams that benefit are the ones that treat routing configuration as a tuned parameter with a review date, rather than a constant someone set once and nobody has touched since. Write the replay script once. Run it every time a price changes. That habit will outlast GPT-6.

Every figure in this article — the per-million rates, the percentage changes, the availability tiers and the model identifiers — comes from OpenAI's launch of 22 September 2026 and the spokesperson confirmation reported by VentureBeat. The monthly cost table is arithmetic on explicitly assumed volumes. We have attached no numbers to OpenAI's capability claims, because none were verified.