For most of the generative AI boom, the competition has been easy to describe: build the smartest model.
On Monday, that race started looking much more like a price war.
Anthropic launched Claude Opus 5.5, promising performance close to its highest-end models at a significantly lower cost. Roughly 90 minutes later, OpenAI announced GPT-6 Sol and Luna, cutting API prices for both models by 50 percent compared with their GPT-5.6 equivalents.
Two rival labs. Two major model releases. One increasingly similar message: frontier-level AI needs to get cheaper.
Anthropic Brings Premium Performance Downmarket
Anthropic is positioning Opus 5.5 as its strongest Opus model yet, with improvements across coding, agentic work and other complex tasks. More interestingly, the company says its performance can approach its more powerful Fable models while requiring substantially less compute.
The economics have moved with it.
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 respectively for Opus 5. Anthropic also says the model is faster, reflecting a broader reduction in the computing required to run it. Reuters reports that Anthropic describes the model as delivering performance comparable with Fable 5.1 while costing 40 percent less to run than its predecessor.
A few months ago, getting closer to frontier performance generally meant moving up the model stack and accepting the bill that came with it.
Opus 5.5 starts to compress that relationship.
Then OpenAI Cut Prices In Half
OpenAI followed shortly afterward with GPT-6 Sol and GPT-6 Luna, extending the technology behind its flagship GPT-6 Astra model into cheaper tiers.
Sol is positioned for more demanding professional and coding work, while Luna is built for higher-volume jobs such as summarization, extraction and straightforward everyday tasks. OpenAI says improvements in inference and caching allow it to charge 50 percent less than the previous GPT-5.6 versions.
GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens. Luna drops to just $0.10 for input and $0.50 for output.
Those numbers matter more as AI shifts from answering occasional prompts to performing sustained work.
Ask a chatbot ten questions and token economics can feel abstract. Run an AI agent across thousands of customer interactions, software tasks, research workflows or internal processes and every fraction of a cent starts multiplying.
OpenAI is explicitly pitching the new models through that lens. The company says Sol can perform certain professional automation tasks at a fraction of the cost of competing Claude models, while Luna can approach the performance of more expensive systems on some coding workloads for dramatically less money. Those are OpenAI’s own benchmark comparisons and should be treated accordingly, but the way the company is selling the models is revealing.
The argument is no longer simply: ours is smarter.
It is: ours gets enough of the job done for less.
The Benchmark War Is Becoming An Economics War
Model companies will keep publishing benchmark charts. Every launch will still arrive with claims about coding, reasoning, factuality and whatever new test has become important that month.
But those numbers are beginning to share the stage with another metric: cost per completed task.
That is a much more consequential measure for companies actually deploying AI.
A model that scores slightly higher on a benchmark but costs five times more to operate may not be the better model for a customer processing millions of requests. Conversely, the cheapest model is not particularly useful if it fails often enough that humans have to redo the work.
The commercial race is increasingly happening between those two points.
How much intelligence does a task actually require? How quickly can the model complete it? How reliably? And what does it cost when multiplied across an entire organization?
We saw this beginning with Claude Opus 5 earlier this summer, when Anthropic started talking less about general intelligence and more about specific jobs, particularly coding and agentic workflows. OpenAI has followed a similar path with multiple model tiers and reasoning levels designed to let customers match compute to the complexity of the work.
Now those tiers are getting cheaper surprisingly quickly.
AI Is Moving From Scarce Intelligence To Abundant Intelligence
There is a bigger shift underneath all of this.
The first phase of generative AI was defined by access. Powerful language models felt scarce because every jump in capability was expensive, technically difficult and concentrated among a handful of labs.
Efficiency changes that equation.
OpenAI says GPT-6 Luna can match its previous-generation Sol model on some factuality tasks at roughly one-hundredth of the cost. It also says improved caching can reduce the cost of repeatedly processing the same context by up to 90 percent.
Whether every benchmark translates perfectly into real-world performance is another question. But the direction is difficult to miss.
Yesterday’s expensive intelligence keeps becoming today’s affordable intelligence.
That is how AI begins to spread into workflows where it previously made little financial sense. Customer service interactions that were too cheap to automate intelligently become viable. Agents can run for longer. Software teams can let models iterate more. Businesses can process documents, analyze data and generate content at volumes that would have made API costs uncomfortable a year earlier.
The model does not simply become cheaper.
Entire use cases become possible.
The Winner May Not Have The Smartest Model
The AI industry has spent years treating intelligence as the ultimate competitive advantage.
It probably still is at the absolute frontier. OpenAI continues to position Astra as its most capable model, while Anthropic maintains specialized systems above Opus for particularly demanding work.
But most people and most businesses do not operate at the absolute frontier all day.
They need enough intelligence, delivered quickly, reliably and affordably.
That creates a different competitive landscape. OpenAI, Anthropic, Google, Meta and others are not only racing to create the most capable models. They are racing to make yesterday’s frontier cheap enough to become infrastructure.
And September 22 may end up being a useful snapshot of that transition.
Anthropic released a stronger Opus for less money.
Ninety minutes later, OpenAI cut prices in half.
The AI race is not slowing down.
It is getting cheaper.