On September 22, 2026, OpenAI released GPT-6 Sol and GPT-6 Luna and cut their token prices roughly in half. Sol dropped from $4.00/$20.00 per million input/output tokens to $2.00/$10.00. Luna went from $0.20/$1.20 to $0.10/$0.50. An OpenAI spokesperson confirmed to VentureBeat that these are permanent rates, not promotional or introductory pricing. A week and a half earlier, Anthropic had launched Claude Opus 5.5 at prices 20% below Opus 5. The price-performance race is real, and it is moving fast.
Here is the part that gets lost in the headlines: a 50% cut to the list price almost never becomes a 50% cut to your AI bill. Finance leaders who pencil the new rate straight into the 2027 forecast are building a number they will miss. The gap between what a provider cuts and what you actually save is where most AI budgets quietly come apart.
What actually changed on September 22
GPT-6 arrives as a three-model family. Luna handles high-volume jobs like summarization, Sol tackles coding and multi-step workflows, and Astra is the flagship. The reductions land on the two workhorse tiers, not the flagship.
| Model | Before (per 1M in / out) | After Sept 22 (per 1M in / out) |
|---|---|---|
| Sol | $4.00 / $20.00 (GPT-5.6 Sol) | $2.00 / $10.00 |
| Luna | $0.20 / $1.20 (GPT-5.6 Luna) | $0.10 / $0.50 |
| Astra (flagship) | new tier | $10.00 / $50.00 ($20 / $100 fast mode) |
OpenAI credited “improvements in caching and inference” for the cut, and the caching story is doing a lot of work: cached input tokens now run at a 90% discount, and The New Stack reports that prompt-caching gains cut fresh processing by more than half across the billions of requests behind GitHub Copilot. That matters, because it tells you where the savings are concentrated and where they are not. If your workloads do not reuse prefixes, the caching half of the discount is not yours to bank. I wrote earlier about how caching economics flipped from a guaranteed win to a conditional one; GPT-6 widens that gap rather than closing it.
Why half-price tokens rarely halve the bill
Four things sit between the published rate and your invoice.
You have to migrate first, and migration is not free. A price cut on GPT-6 Sol does nothing for a workload still calling GPT-5.6. Moving means re-running your evaluation suite, re-tuning prompts, and re-checking output quality on real traffic. Newer models also tend to reason more, and reasoning tokens bill as output. A model that is cheaper per token but spends more tokens to answer can land you flat or higher. The only number that matters is cost per completed task, not cost per million tokens, which is exactly why analysts keep pushing teams toward cost per outcome rather than cost per token.
Your committed capacity does not reprice itself. If you reserved Azure OpenAI provisioned throughput, a public list-price cut does not touch the rate you locked. More on that below, because it is the trap most likely to bite a mature FinOps team.
The rest of the AI bill is moving the other way. Tokens are one line. Compute underneath the models is the other, and it is rising. AWS raised EC2 Capacity Block rates about 15% effective October 7, 2026, taking a reported H100 accelerator-hour from $5.191 to $5.970 and a B200 from $12.355 to $14.208, with larger increases outside the United States. If you train, fine-tune, or self-host, the hardware line can swallow the inference savings whole. The capacity scarcity that reset the commitment playbook has not eased.
Cheaper tokens invite more tokens. Every time inference gets cheaper, teams send more of it: longer contexts, more retries, more agents calling other agents. That is not a reason to avoid the price cut, but it is the reason the cut shows up as more usage rather than a smaller bill. It is the same pattern I described in the forecasting paradox, where the better a team gets at optimizing AI spend, the less its forecast is worth.
The commitment you signed last quarter may now be underwater
This is the one worth putting in front of your CFO.
Azure OpenAI sells provisioned throughput units, or PTUs, as reserved capacity. You can buy them hourly with no commitment, or lock a monthly or annual reservation for a steep discount. CloudZero’s breakdown puts the minimum at 15 PTUs for Global deployments, with annual reservations carrying the deepest discount and PTUs only penciling out “for sustained, predictable, high-volume inference” where throughput “looks like a plateau, not a mountain range.”
A reservation makes sense when the committed rate beats pay-as-you-go at your expected utilization. The problem is that a reservation is a bet on the gap between those two prices staying roughly where it was the day you signed. When the metered rate drops 50% and your reservation does not, the bet moves against you. CloudZero names the mechanism directly: after OpenAI’s earlier July 30 cut, “a workload that justified a PTU commitment against last quarter’s rates may no longer justify it, because the metered alternative got cheaper while the reservation didn’t.” Their advice is to “re-run the crossover math before every renewal, not just the first commitment.”
If that sounds familiar, it is the same failure mode as a cloud reserved instance going underwater when on-demand prices fall. Anyone who has managed a commitment portfolio has watched a once-smart three-year reservation turn into dead weight because the market repriced beneath it. The discipline that keeps reserved instances and savings plans honest is the discipline AI commitments now need: measure utilization continuously, model the break-even against the current metered price, and treat every renewal as a fresh decision rather than an autopilot.
The twist with AI is velocity. Cloud on-demand prices move once or twice a year. Token prices moved twice in two quarters in 2026, and the providers are telling you the cuts are permanent. A commitment that assumed stable metered pricing is working off an assumption the market just invalidated.
Treat a price cut as a renegotiation trigger
The right response to a 50% list cut is not to celebrate a saving you have not captured. It is to run a short, unglamorous sequence.
Start by finding where the cut actually applies. List every AI workload, the model it calls, and whether it bills metered or against a commitment. Metered Sol and Luna traffic gets cheaper the moment you point it at GPT-6, but only after validation. Committed traffic does not move until the reservation expires, so flag every PTU or enterprise agreement with its renewal date and current effective rate.
Next, re-run the crossover math on every commitment against the new metered price, not the price you signed at. If a reservation is now underwater, decide whether to ride it to term, renegotiate, or shift the workload back to metered at renewal. If you have an enterprise agreement, a public list-price cut is your opening to reopen the conversation; the vendor just told the whole market what the new floor looks like.
Then validate before you migrate. Run the new model against your real evaluation set and watch output-token volume, because reasoning behavior is where a cheaper model quietly gets expensive. Measure the delta in cost per completed task, not per token.
Finally, hold the total line, not the unit line. Watch the GPU and egress side of the bill as closely as the token side, and expect usage to rise as inference gets cheaper. The State of FinOps 2026 found 98% of practitioners now manage AI spend, up from under a third two years ago. The ones who will look good in next year’s review are not the ones who quoted the new list price in a forecast. They are the ones who treated the cut as a trigger to re-examine commitments, migrations, and the half of the bill that tokens never covered.
A 50% token price cut is a gift. It is just not a gift that unwraps itself.