For years, the frontier AI race was measured by one question: which company built the smartest model?
That is changing.
On September 22, Anthropic introduced Claude Opus 5.5, cutting its API price while claiming performance comparable to its more expensive flagship models. Shortly afterward, OpenAI released GPT-6 Sol and GPT-6 Luna, cutting the API prices of both models by 50% compared with their GPT-5.6 promotional rates.
The timing was striking. But the more important story is not who announced first.
It is what both companies are now competing on:
the cost of completing useful work.
Anthropic Cuts the Cost of Its Flagship
Claude Opus 5.5 arrives as Anthropic’s latest high-end model for coding, agentic tasks and knowledge work.
Its API price is $4 per million input tokens and $20 per million output tokens, compared with $5 and $25 for Opus 5. Anthropic says typical workloads can be around 40% cheaper overall, partly because Opus 5.5 can accomplish tasks with fewer tokens. Cache-read pricing also fell from $0.50 to $0.20 per million tokens.
That distinction matters.
The headline price is only one part of the economics of AI.
If a model needs fewer tokens, fewer reasoning cycles or fewer interactions with tools to complete the same task, its effective cost can fall much faster than its published token price suggests.
Anthropic is therefore competing on cost per completed task, not simply cost per token.
OpenAI Responds With Sol and Luna
OpenAI’s response arrived with two models rather than one.
GPT-6 Sol is positioned for more demanding professional work, coding and agentic workloads, while GPT-6 Luna is designed for higher-volume, more routine tasks.
OpenAI priced Sol at $2 per million input tokens and $10 per million output tokens, compared with $4 and $20 for GPT-5.6 Sol.
Luna is priced at $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 for GPT-5.6 Luna.
Both represent a 50% reduction against those previous promotional prices.
OpenAI also says GPT-6’s improved caching can reduce the cost of repeated context by 90%, an important change for agents that repeatedly work with the same instructions, documents and tools.
The message is clear.
OpenAI is not simply trying to make GPT-6 more capable.
It is trying to make advanced intelligence cheap enough to run continuously.
The Benchmark Race Is Becoming an Economics Race
This is where the latest releases differ from earlier generations.
The first phase of the frontier AI race was dominated by benchmarks.
Companies competed over reasoning tests, coding evaluations, mathematical problems and increasingly complex academic exams.
Then AI agents changed the question.
A model that can write code, operate a browser, update a CRM, analyze a spreadsheet or complete a business workflow has economic value beyond its benchmark score.
At that point, the relevant metric becomes:
How much does it cost to finish the job?
OpenAI reports that GPT-6 Sol reaches 33.2% on AutomationBench at its highest reasoning setting, at an estimated $0.27 per task. It also reports 68.8% on DeepSWE v1.1 for software engineering, within 1.1 percentage points of Claude Fable 5’s highest reported score in that evaluation, while claiming roughly 80% lower cost per task. These are OpenAI’s own evaluations and comparisons, so they should not be treated as independent head-to-head results.
Anthropic, meanwhile, says Opus 5.5 delivers performance comparable to its Fable 5.1 model while reducing typical operating costs by 40%.
The companies are increasingly selling the same idea from different directions:
frontier intelligence is becoming more economically usable.
Why This Matters for Companies
Consider a software company using AI coding agents.
At a high enough price, developers may use an advanced model selectively — for difficult bugs, architecture decisions or major features.
At a much lower cost, the economics change.
The agent can run more often.
It can test more ideas.
It can analyze larger codebases.
It can remain active for longer.
The same principle applies to research, customer support, finance, legal work, marketing and operations.
Lower inference costs do not merely make AI cheaper.
They make new categories of AI usage economically possible.
The Infrastructure Behind the Price Cuts
There is another layer to the story.
Lower prices do not mean the underlying technology has become inexpensive to build.
Quite the opposite.
The companies are investing enormous resources in compute, specialized hardware, data centers, networking and inference optimization.
OpenAI says improvements in caching and inference are helping it serve GPT-6 models more efficiently. It also says its caching improvements have reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models over recent months.
Anthropic is pursuing a similar objective: produce more useful output while consuming fewer computational resources.
The frontier is therefore moving in two directions simultaneously:
more intelligence and lower cost.
That combination is far more consequential than either one alone.
The New Competitive Equation
The AI industry once had a relatively simple hierarchy.
Build the best model.
Attract users.
Charge for access.
The economics are becoming more complicated.
Now companies must balance at least four variables:
Capability. Cost. Speed. Scale.
A model that is marginally smarter but several times more expensive may not be the best choice for an automated business workflow.
A cheaper model that is slightly less capable may be deployed millions of times.
This creates a new competitive curve in AI.
The winners in specific applications may not necessarily be the models with the highest benchmark scores.
They may be the models that deliver the best intelligence per dollar.
The Race Is Getting Cheaper — and More Intense
Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Sol and Luna arrived within hours of one another.
That does not prove that one launch was directly caused by the other. But the sequence illustrates how quickly the frontier market is responding to competitive pressure.
And the direction is unmistakable.
The most capable AI models are moving toward lower prices, greater efficiency and wider deployment.
The frontier is no longer only about building intelligence.
It is about making that intelligence economically viable at scale.
That may ultimately be the more important race.
Because once advanced AI becomes cheap enough, the question changes again.
Not:
“What can AI do?”
But:
“What should we now ask it to do?”
