OpenAI and Anthropic have put the cost of running their models at the center of their latest releases. OpenAI has unveiled GPT-6 Sol and GPT-6 Luna with API prices up to 50% below GPT-5.6’s promotional rates, while Anthropic has launched Claude Opus 5.5, which the company says costs roughly 40% less to run than Opus 5 on typical workloads. Both releases come after public debate over whether the development of advanced models should slow down.
OpenAI and Anthropic’s price cuts in 20 seconds
- GPT-6 Sol drops from $4 to $2 per million input tokens and from $20 to $10 for output.
- GPT-6 Luna falls from $0.20 to $0.10 for input and from $1.20 to $0.50 for output.
- Claude Opus 5.5 cuts its rate from $5 to $4 for input and from $25 to $20 for output.
- Anthropic attributes part of Opus 5.5’s savings to greater efficiency in token usage.
- The new models arrive as labs compete with cheaper alternatives and open-weight models.
The shift points to a different phase of the AI race. Over the past few years, much of the competition has focused on boosting model capabilities. Now the cost of running those capabilities has become another battleground, especially for companies looking to use AI across thousands or millions of requests.
The Price Comparison: What It Cost Before and What It Costs Now
API rates make the shift from both companies fairly clear to see. The table below uses standard prices per 1 million tokens for the models compared, using as a reference the promotional rates OpenAI charged for GPT-5.6 before the launch of GPT-6.
| Model | Input before | Input now | Output before | Output now | Change |
|---|---|---|---|---|---|
| OpenAI GPT-5.6 Sol → GPT-6 Sol | $4 | $2 | $20 | $10 | -50% |
| OpenAI GPT-5.6 Luna → GPT-6 Luna | $0.20 | $0.10 | $1.20 | $0.50 | -50% input / -58.3% output |
| Anthropic Claude Opus 5 → Opus 5.5 | $5 | $4 | $25 | $20 | -20% |
Prices per 1 million tokens. At OpenAI, the GPT-5.6 Sol and Luna rates were promotional; GPT-6 Sol and Luna are debuting with their new standard rates. Anthropic, for its part, says Opus 5.5’s effective cost is roughly 40% lower than Opus 5’s on typical workloads, thanks partly to a reduction in the number of tokens needed as well.
The difference between price per token and cost per task matters. A model can carry a lower rate per million tokens and still need more tokens or more calls to finish a job. That’s why the real savings for a company will depend on how it uses each model.
Anthropic is putting exactly that question front and center with Claude Opus 5.5: it isn’t only cutting nominal rates, the company says the model is more efficient in its token usage too. Its price drops from $5 to $4 per million input tokens and from $25 to $20 per million output tokens, while cache reads fall from $0.50 to $0.20 per million — a 60% cut.
OpenAI is applying a more direct cut across its two new models. GPT-6 Sol costs half of what GPT-5.6 Sol charged at the promotional rate used as a reference, while GPT-6 Luna cuts its input price in half and its output price by more than half. OpenAI confirmed that both models arrive with API prices 50% below GPT-5.6’s promotional rates.
GPT-6 Sol and Luna Aim to Cover Two Kinds of Work
OpenAI has positioned GPT-6 Sol and GPT-6 Luna below GPT-6 Astra, its highest-capacity model in the new family.
Sol is aimed at more demanding work, especially coding, reasoning, and agent workflows. Luna occupies the slot for higher-volume tasks where cost per request matters most, such as information extraction, classification, or document summarization.
The strategy separates cost from the capability needed. A company can reserve a pricier model for a complex task and use Luna for millions of simple operations.
That approach matters more as AI agents spread. An agent running a single query can absorb a small cost. But if it analyzes documents, calls tools, writes code, checks results, and repeats operations for hours, the number of tokens consumed can grow quickly.
Lowering the price per token changes the economics of those systems.
OpenAI has also worked on caching and inference efficiency. GPT-6 Sol costs $0.20 per million cached input tokens, while Luna drops to $0.01. In both cases, reusing context can cut costs even further for applications that run many operations over the same information.
Anthropic Answers With a More Efficient Opus
Claude Opus 5.5 follows a somewhat different strategy. Anthropic presents it as the first model in its new Claude 5.5 family and says it matches Claude Fable 5.1’s level on most tasks, at a 40% lower cost to run than Opus 5.
The model is available in Claude for Pro, Max, Team, and Enterprise users, as well as through the API and cloud platforms such as Amazon Web Services, Google Cloud, and Microsoft Foundry. Anthropic sets its price at $4 per million input tokens and $20 per million output tokens.
The savings are especially significant for long-running agent tasks. Anthropic notes that cache reads make up a large share of the cost in this kind of work and has cut that rate from $0.50 to $0.20 per million tokens.
There’s also a fast mode for Opus 5.5, which can reach up to 2.5 times the speed, though at a rate of $8 per million input tokens and $40 per million output tokens.
The company also says Opus 5.5 was evaluated by outside organizations such as Frontier Design and METR before launch. Those evaluations are part of the information Anthropic has published about the model and don’t amount to a complete independent comparison across every available model.
The Cuts Arrive as Open Models Push Prices Down
The pricing move is happening in a market where OpenAI, Anthropic, and Google are no longer the only competitors. Companies such as Alibaba, Moonshot AI, and DeepSeek have helped raise the pressure on prices with lower-cost models or models whose weights are available to developers.
For proprietary providers, cutting rates can help keep companies from shifting certain workloads to cheaper alternatives. But it also raises another question: if the price per token falls, usage may rise.
An application that was once too expensive to run at scale can become economically viable. That can benefit providers because it generates more usage volume, even when the unit price is lower.
The battle, then, is starting to be measured in cost per task rather than just cost per million tokens.
A model that costs half as much but needs twice as many steps doesn’t necessarily deliver savings. A model that costs less and completes a task with fewer tokens, on the other hand, can cut an application’s spending more clearly.
That’s precisely why inference and reasoning efficiency are gaining weight in the way new models are being presented.
The Launches Follow the Debate Over Slowing Down AI
The timing also adds a particular dimension to the announcements. These are the first major releases from OpenAI and Anthropic since the public debate sparked by warnings over the risks of accelerated development of advanced systems.
Dario Amodei, Anthropic’s CEO, had called for greater caution in developing frontier models. The debate intensified after Jacob Coxon, a former Anthropic researcher, publicly announced his departure from the company and questioned the pace of development at the leading labs.
Sam Altman, OpenAI’s CEO, and Elon Musk also took part in the public conversation about the development of advanced systems, according to CNBC reporting used as background for this article.
But the new releases show that the safety debate and commercial development are moving forward at the same time. Anthropic hasn’t paused its releases: it has introduced a new, lower-cost Opus and says it has carried over the safeguards built for its most capable models.
OpenAI hasn’t slowed the expansion of GPT-6 either. With Sol and Luna, the company is extending the family toward models meant to be used more frequently and across more operations.
The result is a competition that’s no longer just about who has the most capable model. It’s also about who can offer enough capability at a price that makes it possible to build into products, agents, and enterprise processes at scale.
For developers, the most useful comparison is no longer simply which model scores higher on a benchmark. Cost per task, the number of tokens used, speed, reasoning capability, and the ability to reuse context are starting to matter as much as the quality of the response.
OpenAI is betting on drastically cutting unit prices with GPT-6 Sol and Luna. Anthropic is answering with a cheaper Opus 5.5 that, according to its own data, is also more token-efficient. Meanwhile, open models and Asian providers keep the pressure on an industry that is starting to treat efficiency as a core product feature.
Frequently Asked Questions
How much does GPT-6 Sol cost compared to GPT-5.6 Sol?
GPT-6 Sol costs $2 per million input tokens and $10 for output. GPT-5.6 Sol had a promotional rate of $4 for input and $20 for output, making the cut 50%.
How much does GPT-6 Luna’s price drop?
GPT-6 Luna costs $0.10 per million input tokens and $0.50 for output. GPT-5.6 Luna cost $0.20 and $1.20 respectively, according to official rates.
How much does Claude Opus 5.5 cost?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Anthropic says the typical cost of running it is about 40% lower than Opus 5, thanks partly to greater efficiency in token usage.
What’s the difference between GPT-6 Sol and GPT-6 Luna?
OpenAI positions Sol for more complex work, such as coding and agent tasks, while Luna is designed for high-volume, lower-cost operations like information extraction or document summarization.
