Briefly. OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22, 2026. Sol is designed for complex coding and agentic workflows, while Luna is aimed at narrower, high-volume tasks. Base API pricing: Sol — $2 per million input tokens and $10 per million output tokens, Luna — $0.10 and $0.50 respectively. Both models are available in the API under the names gpt-6-sol and gpt-6-luna; rollout in ChatGPT Work and Codex is taking place in phases. These rates do not automatically determine the cheapest workflow: you need to calculate the cost of the accepted result. OpenAI announcement, Sol API, Luna API.
This material is intended for product leaders, CTOs, developers, and owners of AI workflows. Data was verified on September 22, 2026. We review the official specs and tests, economics, and the selection process. Vendor results are not considered an independent measurement of your process.
Contents
- What GPT-6 Sol and GPT-6 Luna are
- Key differences between the models
- API pricing and the real cost of the task
- What OpenAI's tests show
- Context and API functions
- Where to choose Sol, Luna, or Astra
- Availability in ChatGPT Work and Codex
- Security and limitations
- How to run your own eval
- FAQ
- How AgentSunrise helps choose a model
- Conclusion
What GPT-6 Sol and GPT-6 Luna are
GPT-6 Sol is OpenAI's model for complex coding tasks and agentic workflows. GPT-6 Luna is the most cost-effective model in the family for focused, high-volume tasks. Both use the approaches applied in GPT-6 Astra and extend the lineup in terms of cost and scale. This is OpenAI's positioning; you should validate the specific task with your own tests. Announcement, Sol documentation, Luna documentation.
Unlike simple text generation, agentic work involves multiple steps: the model reads data, calls a tool, evaluates the result, and continues. In this scenario, the price of a single token matters, but the share of successfully completed tasks, the number of retries, and the time spent on human review are often more important. High-volume request classification and autonomous code migration require different models and different access constraints.
Key differences between the models
| Parameter | GPT-6 Sol | GPT-6 Luna | Practical meaning |
|---|---|---|---|
| Positioning | complex code and agents | high-volume focused tasks | route requests by difficulty |
| Input, $/1M tokens | 2 | 0.10 | Luna input is 20 times cheaper |
| Cached input, $/1M | 0.20 | 0.01 | cost-effective with repeated context |
| Output, $/1M | 10 | 0.50 | Luna output is 20 times cheaper |
| Context window | 1,050,000 tokens | 1,050,000 tokens | same size, extraction quality needs testing |
| Maximum input | 922,000 tokens | 922,000 tokens | long input may move to a higher rate |
| Maximum output | 128,000 tokens | 128,000 tokens | the limit is not equal to a reasonable answer length |
| Input modalities | text and images | text and images | both output text |
| Default effort | medium |
medium |
compare at a fixed setting |
Pricing, limits, and modalities are listed in the official Sol card and the Luna card. The “20x” ratio is basic pricing arithmetic, not a comparison of the full cost of a result.
The key difference within the family is tied to quality on harder tasks, not window size: Sol and Luna have the same stated context capacity. That does not mean they have the same ability to find a specific fact in a large collection of files. OpenAI recommends Astra for the most demanding projects, Sol for difficult work on a smaller budget, and Luna for a stream of more compact tasks. OpenAI announcement.
API pricing and the real cost of a task
OpenAI compares Sol with the advertised GPT‑5.6 Sol rate: $4/$20 → $2/$10 per million input/output tokens. For Luna, input drops from $0.20 to $0.10, and output from $1.20 to $0.50. In its general announcement, the company describes the reduction as about 50%; based on the published numbers, Luna’s output rate drops by about 58.3%. This matters when budgeting for long responses with precision. OpenAI table.
For the new models, reading cached input costs 10% of the standard input rate: $0.20 for Sol and $0.01 for Luna per million tokens. Writing cache costs $2.50 and $0.125, respectively. Requests with more than 272,000 input tokens are billed with higher multipliers for the entire request: 2x for input and cache, 1.5x for output. Regional processing, where available, adds 10%; Batch and Flex have a Standard rate at half price; Fast mode is billed at double the applicable rate. Some tools may be charged separately. Sol API, Luna API.
Calculation example. A request with 50,000 standard input tokens and 5,000 output tokens at Sol’s base pricing costs $0.10 + $0.05 = $0.15. The same volume under Luna pricing is $0.005 + $0.0025 = $0.0075. In this simplified pair, Luna is 20 times cheaper at the token level. If Sol resolves more difficult cases on the first try, the difference in the cost of the accepted result may be smaller. If the task is simple and both models provide equally acceptable answers, Luna’s pricing advantage becomes stronger.
For a fair cost model, divide expenses across all attempts by the number of results that pass a predefined acceptance check. Add review time, repeat model calls, tools, and cache costs. For long chains, calculate the total limit, not just the first request. You cannot promise a twofold budget reduction from a pricing announcement without measuring tokens and quality in your own process.
| Metric | Formula or measurement method |
|---|---|
| Call cost | input + cache + output + tools |
| Cost per accepted result | costs across all attempts / number of accepted results |
| Throughput | accepted results / person-hour of review |
| Latency | time from input to accepted result |
| Risk control | errors, rollbacks, unnecessary actions per process |
What OpenAI tests show
In the announcement , OpenAI presents results across several task categories. On AutomationBench GPT‑6 Sol at xhigh gets 33.2% at a cost of $0.27 per task in the described environment. On DeepSWE 1.1 Sol at max shows 68.8%, Luna at max — 66.6%. On the offline benchmark OSWorld 2.0 Sol at xhigh gets 60.5%; OpenAI compares this with 60.3% for Claude Opus 5 under a different effort setting. These percentages refer to different tests and do not describe your agent’s success rate.
OpenAI also reports fewer factual errors on an internal set of disputed conversations: Sol makes about half as many errors as GPT‑5.6 Sol. The company notes that the set was specifically assembled from cases where users had already flagged an error, and it does not represent normal usage. This is a useful signal for model selection, but it should not be turned into a promise of “twice as few errors” for document search or customer support. Methodology description.
Comparisons with Anthropic and other GPT versions depend on the reasoning.effort level, tools, cache conditions, and the cost of the specific test. OpenAI notes that the published evaluations were run in a research environment or via API and may differ from ChatGPT in the product. For model selection, it is useful to look at the direction of change, but any gain should be verified in the same environment on your own task set. OpenAI disclaimers.
Context and API features
Both documentation cards list 1,050,000 context tokens, up to 922,000 input and up to 128,000 output tokens. Sol has a stated knowledge cutoff of April 20, 2026, and Luna — May 18, 2026. That is not the release date: the launch took place on September 22, 2026. For information after the knowledge cutoff, connect a current source and verify the citations. Sol API, Luna API.
The models support the Responses API, streaming responses, structured output, function calling, file and web search, image inputs, and prompt caching. In the Responses API, the documentation lists tools such as code interpreter, hosted shell, computer use, and MCP. The availability of a specific tool in your project, its permissions, and billing should be checked separately. Chat Completions supports function calling for these models only with reasoning_effort=none; for a full agentic tool workflow, the documentation points to the Responses API. Sol API, Luna API.
The reasoning.effort parameter supports none, low, medium, high, xhigh and max. Increasing effort can help with a complex task, but it also changes latency and cost. The mode should be enabled according to routing rules: for example, only after a cheap pass fails, or for a preselected task type. This helps the team avoid a situation where every simple classification pays for maximum reasoning.
Where to choose Sol, Luna, or Astra
| Scenario | Starting candidate | Why | When to revisit |
|---|---|---|---|
| Classification of a large stream of routine requests | Luna | low token cost | frequent misrouting or manual corrections |
| Extracting fields from standard documents | Luna | short, verifiable answer | complex edge cases, weak quality on long files |
| Writing and modifying code across multiple files | Sol | positioning and tests in agentic code | high retry cost or insufficient acceptance |
| Long-running workflow across systems | Sol | stronger on complex actions | need a higher capability level or tighter control |
| The most difficult research and agentic tasks | Astra as the candidate | OpenAI flagship positioning | Sol delivers the same accepted result more cheaply |
This is an editorial test framework, not a quality rule for every request. OpenAI calls Astra its best model overall, but in a specific workflow Sol or Luna may be better on total cost. To avoid misrouting, store task type, model, effort, prompt version, and acceptance result in tracing. OpenAI positioning.
A convenient architecture for a company is a simple default route with explicit escalation conditions. For example, Luna handles the standard intake; if required fields are missing, the confidence score is low, or a validation rule rejects the result, the task is handed off to Sol. Actions that modify customer data, send a message, or initiate a payment require separate confirmation. This route must be tested end to end: extra calls can erase the expected savings.
Availability in ChatGPT Work and Codex
OpenAI announced the launch of GPT‑6 Sol and Luna in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users starting September 22, 2026. For Free and Go, the company says Luna is available in the desktop app. In the API, the model IDs are gpt-6-sol and gpt-6-luna. Rollout in ChatGPT is happening gradually over the course of the day, so not seeing the model in the interface immediately after the announcement does not mean the launch did not happen. Availability section.
OpenAI's wording distinguishes between ChatGPT Work and Chat: at the time of the announcement, the models were not yet available in regular Chat. Users should check the current model selector and organization settings. Do not transfer API pricing directly to subscription pricing: these are different access methods with different limits and rules. Official announcement.
Security and limitations
OpenAI says Sol and Luna showed improvements in behavior consistency tests compared with GPT‑5.6 versions, including fewer cases where the agent misrepresents the work it completed. The company notes that the tests intentionally include difficult situations and do not measure how often the issue occurs in normal use. So even a new model should confirm its actions with tool logs, tests, and reproducible artifacts. Announcement and system card link.
The test results and conditions are published in the Sol and Luna system card. For processes with access to files and external pages, this practical breakdown of prompt injection defenses will be useful prompt injection defenses.
A long context window does not mean it can read a million tokens flawlessly. For documents, use completeness checks: a list of found files, exact links to excerpts, and spot checks for omissions. For computer use or shell access, limit allowed actions; treat external pages and files as untrusted input. No benchmark removes responsibility for access rights, data retention, and rollback capability.
There are also technical limits. According to the model cards, they are not realtime or audio models and do not provide separate endpoints for video, audio, or embeddings generation. The ability to call an image generation tool through the Responses API does not mean the model itself outputs an image as its own modality. For regional processing, check specific availability: the cards indicate EU data residency is available only with Standard processing. Sol API, Luna API.
How to run your own eval
Eval is a model check on predefined tasks under identical conditions with a clear success criterion. To compare Sol, Luna, and your current model, prepare a de-identified set of typical, difficult, and high-risk cases. Do not keep adjusting the prompt after every failed answer for just one candidate: that will stop the comparison from being paired.
- Define the task. What does the model receive, what artifact must it return, and what is it forbidden to do.
- Establish the baseline. The current acceptance rate, employee time, and total process cost.
- Choose the same environment. Tool versions, context, permissions, timeouts, and number of retries should be comparable.
- Run the candidates. Record
reasoning.effort, cache, input/output, tool errors, and latency separately. - Check the result. For code, use tests and review; for facts, use source links; for operations, use the system state after the action.
- Calculate the decision. Compare the cost of an accepted result, rollback rate, human time, and worst-case scenarios.
- Release gradually. Start with reversible operations and define the rollback rule for returning to the previous model in advance.
The article is based on the primary OpenAI announcement, Sol specifications and Luna, verified on September 22, 2026. We did not run our own model benchmark. The pricing comparison is arithmetic; the routing recommendations are an engineering conclusion based on positioning and cost structure that should be validated on your process.
The tiering scheme can be supplemented with a detailed analysis of GPT-6 Astra and an AI model routing methodology. These materials provide context for the decision, but current pricing and availability should be checked against the latest API cards.
Frequently asked questions
When were GPT-6 Sol and Luna released?
OpenAI announced them on September 22, 2026; rollout in the interfaces is happening in stages. Source.
How much do GPT-6 Sol and GPT-6 Luna cost in the API?
Sol is $2/$10, Luna is $0.10/$0.50 per million input/output tokens in the Standard base mode. Cache, tools, and special modes are billed separately. Sol, Luna.
Which model names should be used in the API?
gpt-6-sol and gpt-6-luna. See access and limits in your project. Source.
Is GPT-6 Luna weaker than Sol?
OpenAI positions Luna for narrower mainstream tasks, while Sol is for complex code and agents. That does not mean Luna is worse on every individual prompt: compare the share of accepted results on your own data. Announcement.
Can you use them in ChatGPT for free?
At launch, OpenAI indicated access to Luna for Free and Go in the desktop app. Sol and Luna in ChatGPT Work and Codex are announced for Plus, Pro, Business, Enterprise, and Edu; rollout is gradual. Availability.
Why can a long prompt cost more than the standard rate?
For more than 272K input tokens, the documentation sets a higher rate for the entire prompt. Before uploading a large set of files, estimate length, cache, and whether you can search only the relevant fragments. Sol API.
When should you choose Astra?
When the task requires maximum capability and internal eval confirms that the extra quality justifies the cost and time. For high-volume workflows, first compare Luna and Sol. OpenAI positioning.
How AgentSunrise helps you choose a model
AgentSunrise can help you make a decision based on a single measurable process:
- define the baseline, types of incoming tasks, data sources, and acceptance criteria;
- prepare an enterprise AI environment or an agent with limited tools and an action log;
- compare Sol, Luna, and the current model by accepted results, total cost, and review time;
- set up integration, phased rollout, team training, and support for the chosen path.
The first step is to choose one reversible process and record the data, constraints, current metric, and success criterion. Discuss the project.
Conclusion
GPT‑6 Sol and Luna offer two levels of price and complexity within the GPT‑6 family. It makes sense to test Luna on high-volume short and structured tasks; Sol on complex code and agent workflows. The choice between them and Astra is determined by the quality of the accepted result, total cost, latency, and risk in your environment. Lock in the same test conditions, check pricing exceptions for long context, and open up agent actions gradually.