GPT-6 Sol and Luna cover artwork. Product claims below are checked against the linked sources.
Should you move to GPT-6 Sol or Luna? The strongest immediate case is lower API cost. Whether either model improves your actual work enough to justify a migration still needs testing. This review checks OpenAI's launch post, its API pricing, and independent results from Artificial Analysis, all reviewed on September 23, 2026.
Methodology: Mengbi has not run its own side-by-side API evaluation. The performance figures here belong to their named sources. Our recommendations are editorial inferences, not first-hand benchmark scores.
GPT-6 Sol vs Luna: which jobs fit each model?
| Model | Consider it for | Check before production use |
|---|---|---|
| GPT-6 Sol | Complex coding, analysis, and multistep agents | Total output tokens, retries, test pass rates |
| GPT-6 Luna | Extraction, summaries, classification, and repeatable subtasks | Omission rates and failures on tightly constrained tasks |
| GPT-6 Astra | The hardest tasks with low error tolerance | Higher price and task-level return |
This is a suggested division of work based on OpenAI's model positioning, not a claim that Luna cannot code. The official Sol and Luna pages each list a 1,050,000-token context window, 128,000 maximum output tokens, text and image input, and text output. Prompts above 272K input tokens move to a higher pricing tier for the entire request.
GPT-6 Sol and Luna API pricing
As of September 23, OpenAI's Standard, short-context list prices per million tokens are:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Sol | $2 | $0.20 | $10 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
| GPT-6 Astra | $10 | $1 | $50 |
These figures come from the official pricing table. OpenAI's launch post compares Sol with its GPT-5.6 promotional prices of $4 input and $20 output. Luna moves from $0.20 to $0.10 input and from $1.20 to $0.50 output. That last change is about 58%, so “half-price across every line item” would be imprecise.
For a request with 100,000 uncached input tokens and 10,000 output tokens, the short-context model charge would be $0.30 for Sol or $0.015 for Luna: 0.1 × input rate + 0.01 × output rate. This example excludes cache writes, paid tools, and retries. An input over 272K tokens would use different rates. The useful budget metric is cost per successful task, including any repair work.
What do OpenAI's benchmarks actually establish?
OpenAI reports that Sol made roughly half as many errors as GPT-5.6 Sol on its internal factuality evaluation. It also reports 68.8% for Sol at max effort and 66.6% for Luna at max effort on DeepSWE 1.1, and 60.5% for Sol at xhigh effort on the OSWorld 2.0 offline set. These are vendor-reported results under specific settings, not measured failure rates for your repository or support queue. Read the launch post and its methodological notes before comparing them with other models.
The factuality sample comes from conversations in which users had already flagged an error, so OpenAI says it is not representative of routine use. Competitor scores were taken from public reports, and OpenAI notes that its research or API outputs can differ from production ChatGPT. These qualifications are part of the result, not small print to discard.
Independent results and early reporting
On the Artificial Analysis comparison page, Sol high scores 43 and Luna high scores 32 on Intelligence Index v4.3.2. The same evaluation lists weighted average costs of about $0.37 and $0.03 per index task, respectively. Those are costs for the organization's task mix and weighting, not a quote for an arbitrary coding ticket. The comparison holds reasoning effort at high; different settings or tools may change the result.
The independent data supports a practical trade-off: Luna can be dramatically cheaper, while Sol performs better on that shared set of demanding evaluations. It does not establish that every GPT-6 setting beats its GPT-5.6 counterpart or every competitor. TechCrunch's launch coverage similarly separates the price reduction from OpenAI's performance claims. It is reporting and analysis, not a controlled head-to-head benchmark.
How to decide whether to migrate
Start with 20–50 tasks from your real workload whose outcomes can be checked. That is a pilot size suggestion, not a statistically powered sample. Run the old model, Sol, and Luna with matching inputs, tools, reasoning settings, and time limits. Track first-pass success, human repair time, end-to-end latency, and the full bill including retries.
- Coding: count passing tests and review-ready changes, not just patches produced. Try Luna for bounded subtasks and Sol for more complex implementation or review.
- Extraction: sample Luna's output for omitted fields and incorrect citations. Valid JSON is not proof that the extracted fact is right.
- Agents: measure tool calls and cache hits separately. OpenAI announced caching changes, but your hit rate depends on your prompts and workflow.
Keep human review for high-impact changes and a way to return traffic to the previous model. If the old model is already reliable, lower cost alone may justify a limited trial. If rare failures are expensive, expand the regression set before any broad switch.
FAQ
Is GPT-6 Sol or Luna better for coding?
Start with Sol for complex or multistep work and test Luna on clear, repeatable subtasks with automatic checks. OpenAI's DeepSWE figures show that Luna can handle coding tasks, but they do not replace tests and review in your own repository. For the more capable tier, see our GPT-6 Astra API review.
Where are GPT-6 Sol and Luna available?
At launch, OpenAI listed both in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Free and Go users could access Luna in the desktop app. Ordinary Chat had a different rollout schedule; check your current model picker and release notes.
What are the API model IDs?
Use gpt-6-sol and gpt-6-luna. Both support the Responses API. Check the official model documentation for tools, endpoints, and reasoning settings before integrating.
Verdict
The price improvement is clear; workload-wide quality gains are still something to measure. Sol is a strong candidate for complex tasks, and Luna is attractive when volume and cost matter and outputs can be checked. Run a small controlled comparison before shifting all traffic.
Sources and scope: OpenAI launch post (announcement and vendor evaluations), OpenAI API pricing (checked September 23, 2026), Artificial Analysis comparison (independent high-effort measurements), and TechCrunch (news analysis).