OpenAI has introduced GPT-6 Sol and GPT-6 Luna, expanding its GPT-6 model family with two new models. Earlier this month, the company introduced GPT-6 Astra and said Sol and Luna were trained using similar methods to bring advances from Astra to lower-cost models.
OpenAI said improvements in caching and inference have also reduced the cost of serving GPT-6 Sol and Luna. The company is offering both models at 50% lower API prices than the promotional pricing of their GPT-5.6 counterparts.
GPT-6 API pricing
The new models have lower input and output token prices compared with GPT-5.6 Sol and GPT-5.6 Luna. The pricing is as follows:
| Model | Input | Output | Price reduction |
| GPT‑5.6 Sol → GPT‑6 Sol | $4 → $2 | $20 → $10 | 50% cheaper |
| GPT‑5.6 Luna → GPT‑6 Luna | $0.20 → $0.10 | $1.20 → $0.50 | 50% cheaper |
The prices are per 1 million tokens, with both models representing a 50% reduction from the promotional pricing of their respective GPT-5.6 counterparts. OpenAI continues to position GPT-6 Astra as its highest-capability model across the GPT-6 family.
Professional work
GPT-6 Sol targets complex professional work, with higher usage limits and lower costs for additional iterations. OpenAI evaluated the model on AutomationBench, which tests business workflows across applications.
At xhigh effort, GPT-6 Sol scored 33.2% at a cost of $0.27 per task. The comparison with other models was as follows:
- GPT-6 Astra, low: 30.3%, at 3.9x the cost of GPT-6 Sol
- Claude Opus 5, max: 26.9%, at 11.1x the cost
- Claude Fable 5.1 with Opus 5 fallback, max: 31.4%, at more than 8.9x the cost
- The cost of the Claude Fable 5.1 fallback was not reported.
OpenAI also said GPT-6 Sol at xhigh effort costs about 9% as much per task as Claude Opus 5 at max effort. In the same AutomationBench comparison, Sol exceeds Claude Fable 5.1 at a lower cost and scores above low-effort GPT-6 Astra.
| Model (and effort) | Score | Cost per task |
| GPT‑6 Sol (xhigh) | 33.2% | $0.27 |
| GPT‑6 Astra (low) | 30.3% | 3.9x GPT‑6 Sol |
| Claude Opus 5 (max) | 26.9% | 11.1x GPT‑6 Sol |
| Claude Fable 5.1 w/ Opus 5 Fallback (max) | 31.4% | >8.9x GPT‑6 Sol (fallback cost not reported) |
On Agents’ Last Exam, which evaluates agents on complex professional workflows, GPT-6 Sol at max effort scored 56.4%. OpenAI said this was above Claude Opus 5’s highest score in the evaluation, while costing 60% less per task.
For GPT-6 Luna, the company reported a 5.4 percentage-point improvement over its predecessor at high effort, along with a 58% reduction in cost per task.
Factuality
OpenAI’s internal factuality evaluation uses de-identified real-world conversations in which users had flagged mistakes made by models. On this evaluation, GPT-6 Sol makes about half as many mistakes as its predecessor, while approaching GPT-6 Astra-level reliability at a much lower cost.
GPT-6 Luna also improves on its predecessor. At higher effort, OpenAI said Luna matches GPT-5.6 Sol at about one-hundredth of its cost.
Coding
OpenAI said coding agents are now handling tasks with greater complexity, scope, and duration. Internal usage at OpenAI has grown exponentially, with daily token usage at API prices exceeding:
- $600 for the median researcher
- $7,000 for researchers at the 90th percentile
These figures are referenced in OpenAI’s Research acceleration: The view inside OpenAI. As coding agents take on longer and more demanding tasks, OpenAI said the cost of sustained use becomes increasingly relevant.
On FrontierCode, which evaluates whether coding agents produce changes ready to merge into real codebases, GPT-6 Sol improves substantially over GPT-5.6 Sol and can match Claude Fable 5.1 at xhigh effort at a lower cost.
On DeepSWE v1.1, which tests complex software-engineering tasks in real codebases, the results were:
- GPT-6 Sol, max: 68.8%
- Claude Fable 5, xhigh: 69.9%
- Difference: 1.1 percentage points
- GPT-6 Sol cost: approximately 80% lower per task
GPT-6 Luna at max effort scored 66.6%, with OpenAI saying the result is comparable to Claude Opus 5 and Claude Fable 5 at medium effort. Luna also costs 93% less per task than Claude Opus 5 and 96% less than Claude Fable 5.
Computer use
GPT-6 Astra remains OpenAI’s model for computer-use tasks, while Sol and Luna reduce the cost of computer-use workloads compared with their predecessors. On OSWorld 2.0 offline, GPT-6 Sol at xhigh effort scored 60.5%, compared with 60.3% for Claude Opus 5 at medium effort.
OpenAI reported that GPT-6 Sol achieved its result at approximately 80% lower cost per task. GPT-6 Luna at max effort also exceeds GPT-5.6 Sol at medium effort at one-tenth of the cost.
Collaboration style
OpenAI has brought the communication changes introduced with GPT-6 Astra to Sol and Luna, particularly in technical and coding conversations. The changes include:
- More clarity
- Less jargon
- Fewer unusual turns of phrase
- Fewer low-value details
- Slightly shorter answers without losing substance
Prompt caching
GPT-6 introduces higher default cache hit rates for prompt caching, allowing agents to reuse more context while receiving a 90% discount on cached input-token reads. OpenAI has also added controls to help developers monitor caching and adjust how agents use context.
Developers can use the following tools and controls:
- Prompt Caching Dashboard: Shows how much input is cached and how that changes over time.
- Diagnostics tool: Identifies missed caching opportunities and explains what developers can change.
- Reasoning and tool controls: Developers can increase reasoning effort for harder tasks, lower it for simpler follow-ups, or enable and disable tools while preserving earlier context for cache reuse.
- Explicit breakpoints: Let developers choose where cached prompt prefixes end, providing more control over cache reuse.
OpenAI said GitHub reported that these improvements reduced the share of prompt tokens requiring fresh processing by more than 50% over the past several months across billions of requests to OpenAI models. The company said this has helped Copilot respond faster.
Alignment
GPT-6 Sol and Luna build on the alignment work introduced with GPT-6 Astra, which OpenAI describes as its most aligned model to date. In alignment evaluations, both models improve over their GPT-5.6 counterparts, including lower rates of misleading claims about coding work.
Availability
GPT-6 Sol and GPT-6 Luna are rolling out today across OpenAI’s products and API. Availability is being introduced in stages:
- ChatGPT Work and Codex: GPT-6 Sol and GPT-6 Luna are available to Plus, Pro, Business, Enterprise, and Edu users starting today.
- Desktop app: Free and Go users can access GPT-6 Luna.
- Chat: GPT-6 Sol and GPT-6 Luna are not yet available.
- OpenAI API: The models are available as gpt-6-sol and gpt-6-luna.
OpenAI is gradually rolling out the models in ChatGPT throughout the day to keep the service stable. If GPT-6 Sol or Luna does not appear in ChatGPT Work or Codex yet, users can try again later.