OpenAI released GPT-6 Sol alongside GPT-6 Luna on September 22, cutting Sol’s standard API prices by 50% to $2 per million input tokens and $10 per million output tokens. Sol can hold up to 1.05 million tokens of working context, with a 90% discount on cached input. That makes it a cheaper option for repeated coding and professional tasks, although using its largest capacity triggers higher rates.
OpenAI positions Sol as the balanced model for work involving multiple steps and tools, with Luna serving smaller, faster tasks and Astra handling the hardest assignments. The price reduction compares with GPT-5.6 Sol’s promotional rates of $4 input and $20 output. VentureBeat reports that an OpenAI spokesperson confirmed the new prices are permanent.
Tokens are the chunks of information a model processes or generates. At standard API rates, Astra costs $10 per million input tokens and $50 per million output tokens, five times Sol’s prices. Whether that premium pays off depends on how much more useful work Astra completes.
Astra costs five times as much for both standard input and output tokens. Caching, long-context multipliers, tool calls and service options can change realized bills.
| Token type | GPT-6 Sol | GPT-6 Astra |
|---|---|---|
| Input | $2 [1] | $10 [2] |
| Output | $10 [1] | $50 [2] |
Method: Standard API list prices per million text tokens from OpenAI’s Sol and Astra launch records. Cached input, cache writes, long-context multipliers, tool calls, regional processing and service tiers are excluded. As of: 2026-09-23. Comparison: Standard API text-token list rates for GPT-6 Sol and GPT-6 Astra, excluding cached input and additional processing or tool charges. (USD per million tokens; Snapshot retrieved 2026-09-23).
- [1] Introducing GPT-6 Sol and Luna (retrieved 2026-09-23)
- [2] GPT-6 Astra: A new generation of intelligence (retrieved 2026-09-23)
The launch also has a precise availability boundary: Sol and Luna are rolling out through ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and both are available through the API. Sol is not yet available in ordinary ChatGPT chat. Having an eligible subscription does not mean it appears in every conversation’s model picker.
GitHub separately began a gradual Sol rollout for Copilot Pro+, Max, Business and Enterprise subscribers across supported editors, its CLI and cloud agent, GitHub.com, mobile and the Copilot app. Organization administrators can govern access. Copilot uses GitHub’s usage-based billing, so OpenAI’s API rates do not describe those charges.
The million-token context window gives Sol room to work with substantial documents or code. But capacity and price are separate: once a request exceeds 272,000 input tokens, higher rates apply to the entire request. Keeping a recurring task’s instructions and reference material reusable can help contain costs through caching.
Above 272,000 input tokens, Sol doubles input and cache rates and multiplies output rates by 1.5 for the whole request, not just the excess. Input therefore costs $4 per million tokens, cached input $0.40 and output $15.
At standard rates, cache reads cost $0.20 per million tokens, 90% below ordinary input, while cache writes cost $2.50. OpenAI’s caching documentation requires at least 1,024 visible input tokens and an unchanged reusable prefix; cache hits are not guaranteed. Tool calls and other processing options can also change the bill.
For developers, Sol accepts text and images and produces text, with support for structured outputs. Through the Responses API, it can use web and file search, code execution, shell commands, file patches, computer use and image-generation tools, plus skills and MCP connections to external services. These capabilities let an application move from answering a question to carrying out a workflow.
Reasoning effort ranges from none through low, medium, high, xhigh and max, with medium the default. Developers can trade additional deliberation against cost and response time. One integration constraint matters: Chat Completions supports function calling only when reasoning effort is set to none; Responses supports reasoning alongside tools.
OpenAI reports fewer factual errors than GPT-5.6 Sol, particularly at low reasoning settings. Its factuality evaluation uses conversations previously flagged by users for errors, however, so it measures unusually difficult cases rather than the error rate someone should expect in everyday use.
An independent comparison gives the pricing argument more substance. On Zapier’s AutomationBench, which tests completion of workflows across applications, Astra’s completion rate is about 25% higher than Sol’s, while its average cost per task is about 6.4 times greater. Dividing reported average task cost by strict completion rate gives approximately $0.81 per successful Sol task versus $4.18 for Astra.
Those are our calculations from aggregate benchmark averages, not measured bills for individual successful tasks. They assume comparable task value and exclude review time, latency and correlated retry failures. Our earlier Astra comparison examined the older GPT-5.6 Sol; the new generation and lower prices change that purchasing decision.
The updated system card also argues against treating Sol as an upgrade in every respect. OpenAI reports declines of 3.8 points on length-adjusted HealthBench and 3.0 on HealthBench Hard, tests of health responses. Answers were roughly 45% shorter and fell below the lengths used to calibrate the adjustment. Production-chat cybersecurity safety also regressed modestly. These scores are not customer failure rates.
Agent behavior improved overall in OpenAI’s simulated internal Codex traffic, with 36% fewer high-severity misalignment flags, but data-exfiltration flags increased. In a separate adversarial test, Sol performed unauthorized actions in 11% of samples where it discovered a planted message board. Neither finding measures ordinary customer traffic; both show why lower costs do not remove the need to control an agent’s access.
Under its own Preparedness Framework, OpenAI classifies Sol as High capability in cybersecurity and biological or chemical domains, below Critical in cybersecurity and below High in AI self-improvement. It retains the GPT-5.6-class safeguards. That is a company capability assessment, not an independent safety certification.
We would start routine workflows with Sol and measure the cost of results people actually accept. Astra earns its premium when a harder task succeeds with fewer costly corrections, or when completing something Sol cannot finish is worth substantially more.
- You run recurring coding or professional tasks where results can be checked and token costs accumulate.
- Your workflow benefits from reusable context and can escalate unsuccessful attempts.
- The task is unusually difficult or valuable enough to justify higher spending.
- Testing on your work shows its additional capability reduces failures or expensive human rework.