Overview
GPT-6 Sol and GPT-6 Luna are the two lower-cost members of OpenAI's GPT-6 family, released together on September 22, 2026. Sol targets complex coding and agent workflows. Luna is OpenAI's most efficient GPT-6 model for focused, high-volume tasks where latency and cost accumulate. GPT-6 Astra remains the highest-capability option for the hardest work.
The useful question is which model completes a representative task at the required quality and total cost. Sol and Luna lower API token rates from their GPT-5.6 counterparts, but they serve different workloads. At launch, both are available in ChatGPT Work, Codex, and the OpenAI API; neither is yet available in ordinary Chat. That channel distinction matters before treating this release as a change to a ChatGPT conversation default.
What's New
Two GPT-6 choices below Astra
OpenAI positions Sol for complex coding and agent workflows and Luna for scoped tasks, triage, and high-volume automation. These are new model IDs, gpt-6-sol and gpt-6-luna, rather than aliases for GPT-5.6. Applications can route tasks by quality requirements instead of sending every call to the flagship.
Lower API rates, with separate cache and long-context costs
For standard API processing with up to 272,000 input tokens, Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 input and $0.50 output. OpenAI's launch compares these with GPT-5.6 Sol at $4/$20 and GPT-5.6 Luna at $0.20/$1.20. Cached reads cost $0.20 for Sol and $0.01 for Luna per million tokens; cache writes have their own rates. Above 272,000 input tokens, OpenAI charges higher long-context rates for the entire request. A model's headline input price is therefore only one part of an agent's bill.
Stronger work and coding results at lower task cost
OpenAI reports GPT-6 Sol at 33.2% on AutomationBench with xhigh reasoning and $0.27 per task in its evaluation. It reports 68.8% on DeepSWE v1.1 at max effort for Sol and 66.6% for Luna. At high effort, Luna improves on GPT-5.6 Luna by 5.4 percentage points on AutomationBench at 58% lower cost per task. The launch also reports about half as many factual mistakes for Sol as its predecessor on an internal evaluation of conversations flagged for errors. These results use OpenAI's stated settings and tasks; they do not predict your completion rate or bill without a workload test.
Large context and GPT-6 reasoning controls
Both API models list a 1.05-million-token context window and a 128,000-token output limit. They support reasoning effort from none through max, unlike Astra, which does not support none. OpenAI's GPT-6 guidance recommends the Responses API for tool use and describes cache-preserving effort and tool-availability changes. That makes it possible to test a cheaper model or lower effort before paying for deeper reasoning, while keeping explicit quality checks around the result.
Compared With Previous Version
Sol and Luna follow the GPT-5.6 family; Astra was released earlier in September as the GPT-6 flagship. The joint release changes model choice and economics rather than replacing Astra as the highest-capability tier.
| Decision factor | GPT-5.6 Sol | GPT-6 Sol | GPT-5.6 Luna | GPT-6 Luna |
|---|---|---|---|---|
| Intended role | Previous flagship reasoning model | Demanding everyday coding and agent work | Previous low-cost tier | Focused, high-volume work |
| Standard API input / output, per 1M tokens | $4 / $20 promotional rate | $2 / $10 | $0.20 / $1.20 at launch comparison | $0.10 / $0.50 |
| API model ID | gpt-5.6-sol |
gpt-6-sol |
gpt-5.6-luna |
gpt-6-luna |
| Reasoning effort on GPT-6 | — | none through max |
— | none through max |
| Best reason to test | Existing integration baseline | More demanding tasks at lower listed token prices | Existing low-cost baseline | Higher-volume tasks with a smaller per-call budget |
The GPT-5.6 prices above are the launch article's comparison point. Check the live API pricing page before budgeting a migration, especially for cache writes, long context, regional processing, and service tiers. Keep Astra in the evaluation when the cost of a failed task outweighs the token-price difference.
Which GPT-6 Model Should You Use?
| If your workflow needs... | Start with | What to measure |
|---|---|---|
| Complex coding, research, or an agent that must make judgment calls | Sol | Successful tasks, retries, tool accuracy, and total cost |
| Repeated extraction, triage, small edits, or clear briefs at scale | Luna | Error rate, latency, review time, and batch cost |
| The strongest available result on an ambiguous or high-consequence task | Astra | Whether the quality gain offsets its higher token rate |
This is a routing starting point, not a universal quality ranking. Test the same task set at realistic reasoning effort and include tool calls, caching, human review, and retries in cost per successful task.
Compatibility and Migration
- API routing: Replace a GPT-5.6 model ID with
gpt-6-solorgpt-6-lunain a controlled test first. Compare output quality, latency, token use, and failure recovery before changing a production default. - Tool calling: Use the Responses API for built-in tools and for function calling with reasoning. For Sol and Luna, Chat Completions function calling is supported only when
reasoning_effortisnone; existing Chat Completions tool workflows at higher effort need an API-path change. - Request parameters: OpenAI's GPT-6 guidance says to remove unsupported sampling and log-probability parameters when reasoning effort is not
none. Preserve existing effort where supported, then evaluate alternatives on your own tasks. - Long context and residency: Both models support large prompts, but requests over 272,000 input tokens use the higher long-context rates for the full request. EU data residency is available only with Standard processing.
- Caching: GPT-6 adds improved default cache reuse and separate cache-write pricing. Measure cache hits and total spend on long conversations rather than multiplying the uncached input rate by all prompt tokens.
Availability & Access
OpenAI says Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can try Luna in the desktop app. Both are available in the OpenAI API. Neither model is yet available in ordinary Chat, and OpenAI describes a staged rollout within the supported products. Check your account's model picker or API project access before planning a team-wide switch.
Access through a ChatGPT plan and API token billing are separate. The prices below describe API usage, not a monthly ChatGPT subscription or an included Work/Codex allowance.
API Pricing
OpenAI's standard short-context API prices, per one million tokens, are:
| Model | Input | Cached input read | Cache write | Output |
|---|---|---|---|---|
| GPT-6 Sol | $2.00 | $0.20 | $2.50 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 |
| GPT-6 Astra | $10.00 | $1.00 | $12.50 | $50.00 |
These are standard rates for prompts with at most 272,000 input tokens. Above that threshold, input and cache rates double and output rates rise by 50% for the entire request. Regional processing and faster service tiers can also add charges. Check the live OpenAI API pricing before committing to a budget.
Who Should Upgrade / Who Should Wait
Test Sol now if GPT-5.6 Sol handles coding, research, or agents for you and total task cost is material. Its listed token rates are lower, while OpenAI reports gains in professional work and factual reliability. Keep GPT-5.6 as a baseline until your own success-rate and latency checks pass.
Test Luna now if you run many well-scoped requests or need a cheaper supporting model. Measure whether its error rate, review time, and retries preserve the savings at your chosen effort level.
Keep Astra for the hardest tasks when the best achievable result matters more than the lowest per-token price. Wait before changing a Chat workflow because Sol and Luna are not yet available in ordinary Chat. API teams using reasoning with Chat Completions function calls should migrate that path to Responses before expecting the same tool behavior.
