OpenAI has expanded the GPT-6 family with Sol and Luna, two lower-cost models intended for workloads that do not require the full capability or price of GPT-6 Astra.
The clearest change is API pricing. 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.
OpenAI presents both models as 50 percent cheaper than the promotional pricing for their GPT-5.6 predecessors. The exact earlier prices were $4 and $20 for Sol input and output, and $0.20 and $1.20 for Luna.
The September 22 launch also illustrates how model availability now depends on the product, subscription, rollout stage and organisation policy rather than one universal release switch.
API prices fall, but workload cost still varies
The official OpenAI announcement positions Sol as a balance of capability and cost, while Luna is the smallest and cheapest GPT-6 option. Astra remains the company’s top tier for its most demanding workloads.
Developers can call the new models through the API using gpt-6-sol and gpt-6-luna.
List prices alone do not determine the cost of an application. Output length, reasoning effort, tool calls, retries, cached context, failure rates and the number of steps required to finish a task can all affect the final bill.
A cheaper model can cost more per successful task if it needs additional calls or human correction. Conversely, a lower-priced model that meets a workflow’s accuracy threshold may make previously expensive high-volume processing practical.
Teams should therefore compare complete workflows rather than token rates in isolation.
Caching is part of the economics
OpenAI says inference and caching improvements helped it reduce prices. GPT-6 cached input-token reads receive a 90 percent discount.
The company also says developers can change reasoning effort or available tools without invalidating earlier reusable context. Explicit cache breakpoints provide more control over which prompt prefixes are retained.
Those changes are relevant to coding agents and long conversations, where repositories, instructions, prior messages and tool descriptions can create large repeated prefixes.
OpenAI cites GitHub data showing that caching changes reduced the share of prompt tokens needing fresh processing by more than 50 percent across billions of requests to OpenAI models. That is a reported infrastructure result, not evidence that every application will receive the same reduction. Cache performance depends on how stable and reusable its prompt prefixes are.
Performance and factuality results remain company claims
OpenAI reports improvements across professional work, coding, computer use, factuality and alignment. It also publishes comparisons with Anthropic models at different reasoning-effort settings and estimated costs per task.
These results were produced in OpenAI’s research environment or through its API. The company notes that production ChatGPT can behave differently because system prompts and available tools vary.
Its factuality evaluation uses de-identified conversations in which users had previously flagged mistakes. OpenAI says GPT-6 Sol made about half as many mistakes as GPT-5.6 Sol on that test. The company also states that these error-inducing conversations are not representative of typical usage and that the scores were not controlled for response length.
The result is useful evidence about OpenAI’s test set, but it is not independent proof of a general 50 percent factuality improvement. Teams considering a migration still need evaluations based on their own prompts, tools, data, error costs and review requirements.
TechCrunch’s launch report similarly describes the lower pricing and factuality improvements as company claims.
Availability depends on where the model is used
At launch on September 22, OpenAI said GPT-6 Sol and Luna were available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users.
Free and Go users could access Luna through the desktop app. OpenAI said the models were not yet available in Chat and described the broader ChatGPT rollout as gradual. Users who could not see a model immediately were told to try again later.
That launch-time distinction matters because ChatGPT Work, Codex, desktop access, Chat and the API do not expose models on identical schedules.
GitHub announced a separate Copilot rollout. Sol is available for Copilot Pro+, Max, Business and Enterprise plans. Luna also includes Copilot Pro.
GitHub says both models are offered through model pickers across supported development surfaces, including Visual Studio Code, Visual Studio, Copilot CLI, the cloud coding agent, GitHub’s Copilot app, github.com, mobile clients and several IDEs.
Copilot Business and Enterprise administrators can control GPT-6 access through model policies. Under GitHub’s default enablement behaviour, a new model is enabled automatically unless an administrator has disabled the global default or the model itself. The Copilot rollout is gradual and uses usage-based billing.
Migration should start with a bounded evaluation
Sol and Luna give developers two cheaper GPT-6 price points, but the right choice depends on task difficulty and failure tolerance.
A practical comparison should use representative inputs, the intended reasoning effort, identical tools and measurable success criteria. Teams should track total cost per accepted result, latency, retries, factual errors, tool failures and human-review time.
They should also verify access in the actual deployment surface. API availability does not guarantee that a model is already visible in Chat, Codex or a managed Copilot organisation.
The pricing change is immediate and concrete. The wider performance advantage remains something each team must test.
