OpenAI’s GPT-6 Sol and Luna Push AI Into a Price War
The frontier AI race used to be about who had the smartest model. Increasingly, it is about who can make intelligence cheapest to use.
OpenAI has expanded its GPT-6 family with GPT-6 Sol and GPT-6 Luna, two models designed to bring capabilities from its higher-end Astra model into cheaper, higher-volume workloads.
The biggest number in the announcement may not be a benchmark score.
It's 50%.
OpenAI has cut API pricing for both models by half compared with GPT-5.6 promotional pricing. GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens, while Luna comes in at $0.10 and $0.50 respectively.
That's another sign that AI competition is moving rapidly from raw intelligence toward economics.
OpenAI is building different intelligence for different jobs
The GPT-6 lineup now has a clearer hierarchy.
Astra remains OpenAI's highest-end model for the most demanding work.
Sol is aimed at complicated professional tasks such as coding, computer use and multi-step workflows.
Luna is positioned for high-volume work where speed and price matter more, including information extraction, document summaries and simpler business tasks.
That resembles the way cloud computing matured.
Businesses don't use the most powerful server for every job.
They select the amount of compute the task actually requires.
AI appears to be heading toward the same architecture.
Model cost is becoming an enterprise feature
When a person occasionally asks an AI chatbot a question, the difference between two inference prices can look trivial.
Agents change the calculation.
An enterprise agent may execute hundreds or thousands of model calls while researching information, navigating software or writing code.
Multiply that by thousands of employees and millions of tasks, and token efficiency starts showing up directly in operating costs.
OpenAI says improvements in caching and inference helped it reduce prices. It also says GPT-6 Sol makes about half as many factual errors as GPT-5.6 Sol on an internal evaluation based on conversations previously flagged for factual mistakes. Those are company-reported results rather than independent evaluations.
The strategic message is clear:
The model should not only be smarter.
It should be economically practical to keep running.
Anthropic launched cheaper intelligence on the same day
OpenAI isn't making this move in isolation.
Anthropic introduced Claude Opus 5.5 on September 22, roughly 90 minutes before OpenAI's announcement.
Anthropic says Opus 5.5 performs around the level of its higher-end Fable 5.1 model for many workloads while costing roughly 40% less to run than Opus 5. It also priced input and output tokens below the previous Opus generation.
That timing makes the competitive direction difficult to miss.
AI labs are no longer only announcing:
Our model is more capable.
They are increasingly announcing:
Our model can do comparable work with fewer dollars and fewer tokens.
Coding may be where the economics matter first
Coding agents are especially sensitive to cost because they can operate for hours.
They read files.
Generate code.
Run tests.
Inspect errors.
Make edits.
Repeat.
OpenAI says its own researchers are using large volumes of tokens through coding agents, illustrating how quickly sustained autonomous work can become expensive.
This helps explain why both OpenAI and Anthropic are emphasizing cost per completed task rather than simply cost per token.
Businesses ultimately don't care how many tokens an agent consumes.
They care how much it costs to finish the job.
The benchmark war is becoming a unit-economics war
AI benchmarks are unlikely to disappear.
But the next generation of enterprise comparisons may look different.
Instead of:
Which model scored highest?
Companies may increasingly ask:
Which model completed the workflow?
How long did it take?
How many human corrections were required?
How much did it cost?
How reliably can we repeat the result?
Those questions favor efficient models even when they are not the absolute leader on every benchmark.
What happens next?
Expect model routing to become standard enterprise architecture.
Simple request? Use a cheaper model.
Difficult coding problem? Route it to something stronger.
Mission-critical research? Escalate to the frontier tier.
The company that wins may therefore not be the one that persuades businesses to use a single model for everything.
It may be the one that gives developers the strongest intelligence-per-dollar across the entire stack.
GPT-6 Sol and Luna suggest OpenAI knows the next AI race will be fought as much on the invoice as on the benchmark.