OpenAI expanded its GPT-6 model family on September 22, 2026, releasing GPT-6 Sol and GPT-6 Luna as the mid-tier and budget-tier successors to the GPT-5.6 lineup. The two models sit beneath GPT-6 Astra, the flagship that launched on September 3, and are built to bring Astra-level improvements in factuality, coding, and alignment to a significantly lower price point. The release came roughly 90 minutes after Anthropic shipped Claude Opus 5.5, underscoring the accelerating rivalry between the two AI labs.
What Are GPT-6 Sol and GPT-6 Luna?
Sol and Luna are not new names in OpenAI's product line. The original versions launched earlier in 2026 alongside GPT-5.6, where Sol served as the mid-range workhorse and Luna handled lightweight, high-volume workloads. The GPT-6 upgrades retain those roles but were trained with methods similar to Astra, bringing meaningful improvements in professional work, factuality, coding ability, and computer use.
GPT-6 Sol is designed for complex tasks. Think coding, multi-step agent workflows, and deep reasoning across long contexts. OpenAI positions it as the go-to model for developers who need Astra-grade intelligence but cannot justify Astra's $10/$50 token pricing for every request.
GPT-6 Luna is the economy option, optimized for fast responses at high volume. OpenAI describes its ideal use case as tasks with a clear goal: summarizing documents, extracting structured information, answering quick questions, and handling routine queries at scale. Luna is also the first GPT-6 model available to Free and Go subscribers through the ChatGPT desktop app.
Both models share a 1.05 million-token context window with a 922,000-token input cap and up to 128,000 output tokens. They support text and image input, text output, and six configurable reasoning-effort settings from "none" through "max."
API Pricing: A 50% Cut Across the Board
The biggest headline from this launch is cost. OpenAI slashed API prices by 50% or more compared to the GPT-5.6 promotional rates, and a company spokesperson confirmed to VentureBeat that these are permanent prices, not a launch promotion. The savings come from improvements in caching and inference infrastructure.
GPT-6 Family: API Pricing Comparison
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Cached Input |
| GPT-6 Sol | $2.00 | $10.00 | $0.20 |
| GPT-5.6 Sol (old) | $4.00 | $20.00 | N/A |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
| GPT-5.6 Luna (old) | $0.20 | $1.20 | N/A |
| GPT-6 Astra (flagship) | $10.00 | $50.00 | $1.00 |
Note: Prompts exceeding 272K input tokens cost 2x on input and 1.5x on output. Batch and Flex modes offer an additional 50% discount.
Luna's output pricing saw an even steeper decline than the 50% input cut, dropping 58% from $1.20 to $0.50 per million output tokens. For teams running coding agents or document-processing pipelines that burn through millions of tokens daily, the cost difference is substantial. OpenAI notes that the median researcher at the company consumes over $600 in API-equivalent tokens per day, while the 90th percentile crosses $7,000.
Benchmark Performance
OpenAI published results across five external benchmarks and an internal factuality evaluation. The framing throughout is cost-per-task, not absolute score leadership, and that distinction matters.
Key Benchmark Results
| Benchmark | GPT-6 Sol (max) | GPT-6 Luna (max) | Competitor Ref. |
| DeepSWE v1.1 | 68.8% | 66.6% | Fable 5: 69.9% |
| OSWorld 2.0 | 60.5% (xhigh) | N/A | Opus 5: 60.3% |
| Agents' Last Exam V1 | 56.4% | 50.9% | Opus 5: 55.9% |
| FrontierCode 1.1 | 49.3% | N/A | Fable 5.1: 48.7% |
| AutomationBench | 33.2% (xhigh) | N/A | Beats Opus 5 |
Scores are vendor-reported from OpenAI's launch data. Independent reproduction is pending.
The Factuality Claim
OpenAI's strongest quality claim centers on factual accuracy. Using an internal evaluation built from de-identified ChatGPT conversations where users had explicitly flagged mistakes, GPT-6 Sol produced roughly half the factual errors of GPT-5.6 Sol at every reasoning-effort level. At the "high" setting, for instance, Sol's error rate dropped to 5.1% from GPT-5.6 Sol's 10.8%.
Luna showed a different kind of value on the same test. At higher effort levels, GPT-6 Luna matched GPT-5.6 Sol's factual accuracy while costing roughly one-hundredth as much per task. Its error rate at max effort landed at 7.6%, slightly better than GPT-5.6 Sol's 8.5% at max, at a fraction of the price.
OpenAI acknowledged that these test conversations were specifically selected because an earlier model had made a mistake, so real-world error rates during typical usage would be lower. Verbosity sweeps across both models showed virtually no correlation between response length and accuracy, suggesting the improvements come from deeper internal verification rather than shorter, evasive answers.
Coding Performance: Cheaper, Not Always Higher-Scoring
On DeepSWE v1.1, which tests models on complex, multi-file software engineering tasks in production codebases, GPT-6 Sol at max effort scored 68.8%. That is within 1.1 percentage points of Claude Fable 5's top score of 69.9% at xhigh effort, but at approximately 80% lower cost per task. However, GPT-5.6 Sol at max effort actually scored higher on this benchmark at 72.7%, though it cost $6.46 per task versus Sol's $2.74.
Luna's coding number was the surprise of the launch. At max effort, it scored 66.6% on DeepSWE, landing within 2.2 points of Sol while costing 93% less than Claude Opus 5 and 96% less than Fable 5 per task. For teams building production coding agents, that cost ratio could reshape which model gets used as the default.
On FrontierCode 1.1, which evaluates code on mergeability alongside correctness, GPT-6 Sol at max scored 49.3%, an improvement over GPT-5.6 Sol's 47.5% and on par with Fable 5.1 at xhigh (48.7%) at much lower cost.
The Anthropic Timing: 90 Minutes Apart
The competitive context is impossible to ignore. Anthropic released Claude Opus 5.5 with a 20% price cut just 90 minutes before OpenAI's Sol and Luna announcement. OpenAI's launch materials repeatedly compare Sol and Luna against Anthropic's Opus 5 and Fable 5.1, claiming substantial cost advantages at comparable performance levels.
There are caveats worth noting. Several of OpenAI's benchmark comparisons used competitors' published scores at different effort settings, which makes direct comparisons less straightforward. Additionally, the Anthropic model on most of OpenAI's launch charts was Opus 5, not the newly released Opus 5.5, which already outperforms Astra on some coding and knowledge work benchmarks according to early reports. Independent head-to-head testing between the latest models from both labs has not yet been published.
Availability and Rollout
GPT-6 Sol and Luna are available immediately in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu paid accounts. They are also live in the OpenAI API under the model IDs gpt-6-sol and gpt-6-luna. Luna has additionally been made available in the ChatGPT desktop app for Free and Go subscribers, making it the first GPT-6 model most unpaid users can access.
OpenAI expects to roll the models out to the ChatGPT website and mobile apps gradually throughout the day. Both models are also expanding to GitHub Copilot, where Sol is available to Pro+, Max, Business, and Enterprise plans, and Luna is available from Pro tier and above.
One notable absence: there is no GPT-6 Terra. The middle "Terra" tier from the GPT-5.6 lineup has no counterpart in this release, leaving a two-model structure beneath Astra.
What This Means for Developers and Businesses
The GPT-6 Sol and Luna launch is primarily a cost story, not a capability ceiling story. OpenAI is not claiming these models beat Astra. The company's own announcement states that Astra "continues to be our best model across the board." What Sol and Luna offer is Astra-generation training at dramatically lower prices, enabling developers and businesses to run agent workflows, coding assistants, and document-processing pipelines without burning through budgets at Astra's premium rates.
For high-volume API users, the 50% price cut is meaningful on its own. But the real shift may be Luna at $0.10/$0.50 per million tokens scoring within striking distance of models that cost 10x to 90x more per task. If those numbers hold up under independent testing, Luna could become the default for a significant share of production workloads where good-enough accuracy at rock-bottom cost is the priority.
The AI model landscape continues to compress on both price and performance. With OpenAI and Anthropic releasing major updates within the same hour, and OpenAI's DevDay 2026 scheduled for September 29 in San Francisco, the pace of competition shows no sign of slowing down.