- OpenAI cut GPT-6 Sol and Luna API prices.
- AI providers are focusing more on inference costs and model selection.
OpenAI expanded its GPT-6 model family with GPT-6 Sol and GPT-6 Luna. The company said API prices for the two models are 50% lower than the promotional pricing of their GPT-5.6 counterparts. The models join GPT-6 Astra, which OpenAI introduced earlier in September.
The release came on the same day Anthropic introduced Claude Opus 5.5 at lower prices than Opus 5. Anthropic cut input and output token prices by 20% and said the new model costs about 40% less to run on typical workloads because its token prices are lower and the model uses fewer tokens per task.
Under OpenAI’s standard pricing for prompts with up to 272,000 input tokens, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens.
GPT-5.6 Sol had promotional rates of $4 per million input tokens and $20 per million output tokens. GPT-5.6 Luna was priced at $0.20 per million input tokens and $1.20 per million output tokens following an earlier reduction, meaning GPT-6 Luna halves the input rate and cuts the output rate by more than half.
The latest reductions follow earlier pricing changes by OpenAI. In July, the company cut GPT-5.6 Luna prices by 80% and GPT-5.6 Terra prices by 20% as businesses scrutinised spending on AI services, Reuters reported.
OpenAI has since linked lower prices with higher usage. Chief Financial Officer Sarah Friar said at Goldman Sachs’ Communacopia + Technology Conference in September that the 80% reduction for Luna helped drive a roughly tenfold increase in usage. Friar also said enterprise customers were seeking clearer returns from AI spending and that OpenAI was experimenting with pricing tied to business outcomes.
The new models target different workload levels. OpenAI positions Sol for complex coding and agentic workflows, while Luna is the lower-cost option in the GPT-6 family.
Lower prices put cost per task in focus
The cost of running an AI workload depends on more than the advertised price per million tokens. Coding agents can make multiple model calls, carry context between turns, and use external tools while completing a single task.
GitHub said in September that reducing tokens in individual interactions does not necessarily lower the cost of completing a coding task, as agents can make additional calls to recover information removed from their context. In one Copilot experiment, removing unnecessary information from model inputs cut inference costs by about 5% in offline agentic coding benchmarks.
OpenAI also published cost-per-task measurements alongside benchmark results for its new models. These measurements account for the amount of inference used to complete a tested workload rather than comparing models only by their published token prices.
Anthropic has taken a similar approach with Opus 5.5. Its input price fell from $5 to $4 per million tokens and its output price dropped from $25 to $20, while Anthropic said typical workloads cost about 40% less than with Opus 5 because the model uses fewer tokens per task.
The cost comparisons published by OpenAI and Anthropic are based on vendor testing and can vary according to model settings, workload, and token consumption.
Caching changes the economics of agent workloads
The cost of coding and agent workloads also depends on how much information a model repeatedly processes. Long-running systems can reuse instructions, tool definitions, conversation history, documents, and other context across multiple requests.
OpenAI’s prompt-caching system allows previously processed portions of a prompt to be reused rather than processed again. The company says cached input can be discounted by as much as 90%, depending on the model.
GPT-6 Sol’s cached-input rate is $0.20 per million tokens compared with $2 for standard input, while GPT-6 Luna costs $0.01 per million cached-input tokens compared with $0.10 for standard input.
OpenAI said cached content can include developer instructions, tool definitions, conversation history, text, images, and documents, allowing repeated context to be reused across compatible requests.
Anthropic has also reduced caching prices. The company cut cache-read pricing for Opus 5.5 from $0.50 to $0.20 per million tokens, a 60% reduction, and said cache reads account for the majority of costs in agentic and coding work.
Model selection becomes another cost lever
GitHub has also made cost an explicit factor in Copilot’s automatic model selection. On September 14, the company added efficiency, balance, and intelligence settings that determine how the system weighs cost, quality, and response time.
GitHub said a straightforward request, such as adding a docstring, can still be routed to a smaller model even when a user selects the intelligence setting. Usage charges are based on the model Copilot selects, allowing different models to handle different requests according to their complexity.
OpenAI’s pricing also varies according to how workloads are processed. The $2/$10 rates for GPT-6 Sol and $0.10/$0.50 rates for Luna apply to standard processing for prompts with up to 272,000 input tokens.
Batch and Flex processing for GPT-6 Sol and Luna are priced at 50% of standard rates, according to OpenAI’s API documentation. Fast mode, which prioritises processing speed, costs twice the applicable rate.
Lower token prices do not necessarily make the total cost of an AI application predictable. Reuters reported in July that although token prices had been falling, businesses can still face variable costs because workloads consume different amounts of inference. GitHub’s Copilot testing similarly found that removing information from an individual request can increase total usage if an agent later has to retrieve the missing context.
GPT-6 Sol and Luna are available through the OpenAI API as gpt-6-sol and gpt-6-luna. Both support text and image inputs through the Responses and Chat Completions APIs.
OpenAI is also rolling out the models across its ChatGPT and coding products. The company said Sol and Luna were becoming available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while free and Go users can access Luna through the desktop app.
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post OpenAI cuts GPT-6 Sol, Luna prices as Anthropic lowers Opus 5.5 costs appeared first on TechWire Asia.