For years, a major AI launch usually came with the same pitch: this model is smarter. This week, OpenAI and Anthropic put a different number near the top of the conversation: what it costs to keep an AI system working.
On September 22, OpenAI introduced GPT-6 Sol and GPT-6 Luna, while Anthropic released Claude Opus 5.5. The companies are still making capability claims, but both announcements put efficiency and long-running agent work front and center.
Lower prices for work that runs longer
OpenAI lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens for standard prompts up to 272,000 input tokens. GPT-6 Luna is $0.10 per million input tokens and $0.50 per million output tokens. Those figures compare with $4 and $20 for GPT-5.6 Sol and $0.20 and $1.20 for GPT-5.6 Luna. OpenAI says the new models are 50% cheaper than the promotional pricing of their predecessors.
The company says the cost question matters more as coding agents take on larger jobs. OpenAI put its own internal use in striking terms: valued at API prices, daily token usage exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile. That is a company-reported estimate for intensive internal work, not a measure of ordinary chatbot use.
GPT-6 Sol and Luna launched in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, while Free and Go users could access Luna through the desktop app. At launch, OpenAI said the models were not yet available in ordinary Chat. Both are available through the API.
Anthropic is cutting the bill too
Claude Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Anthropic also cut cached reads from $0.50 to $0.20 per million tokens. The company says Opus 5.5 costs about 40% less on typical workloads and generates output more than 30% faster than Opus 5.
Those comparisons come from Anthropic’s testing and should be read as vendor claims. The release still points to a real change in the market: providers are competing over the price of completed work, not only benchmark scores or the cost of a single prompt.
The agent economics test
An agent can search files, call tools, make a change, check the result and try again. Each step adds time and compute. A cheaper model can make more experiments practical, but a low token price is not enough if a system needs many more attempts to finish the job.
That makes cost per successful task a more useful question than cost per token alone. OpenAI and Anthropic are both asking developers to let their systems tackle longer, messier work. Whether the savings hold up will depend on how reliably each model completes that work.
The AI race is still about capability. These launches show that it is also becoming a contest over how much useful work a dollar can buy.
Sources
- OpenAI: Introducing GPT-6 Sol and Luna
- OpenAI API pricing
- Anthropic: Claude Opus 5.5
- TechCrunch: OpenAI launches GPT-6 Sol and Luna
