GPT-6 Sol and Luna
Introducing GPT-6 Sol and Luna
OpenAI is expanding the frontier of accessible intelligence. Following the recent debut of GPT-6 Astra—currently the world's most aligned and capable model—we recognize that professional requirements vary. Not every task requires the massive depth of Astra; different projects operate on different scales, rhythms, and budgets.
To address this, we are launching GPT-6 Sol and GPT-6 Luna.
While Astra set the new benchmark for intelligence, Sol and Luna are designed to democratize that power by maximizing cost efficiency. By applying the same training methodologies used for Astra, we have infused Sol and Luna with state-of-the-art capabilities in coding, computer interaction, factuality, professional workflows, and alignment, all while ensuring they remain fast and affordable.
📉 Optimized Economics
Thanks to breakthroughs in inference and caching, we have significantly lowered the overhead of serving these models. We are passing these gains to our users via a 50% price reduction compared to the promotional rates of GPT-5.6.
API Pricing Comparison (per 1M tokens):
| Model Transition | Input Cost | Output Cost | Total Reduction |
|---|---|---|---|
| \text{\4} \rightarrow \text{$2}$ | \text{\20} \rightarrow \text{$10}$ | 50% | |
| \text{\0.20} \rightarrow \text{$0.10}$ | \text{\1.20} \rightarrow \text{$0.50}$ | 50% |
Pro Tip: Use GPT-6 Astra when you require an uncompromising, top-tier experience. Use Sol and Luna to scale complex work efficiently.
💼 Professional Work & Automation
GPT-6 Sol and Luna provide a significant leap in intelligence for the models users rely on for daily productivity.
AutomationBench Performance
In tests involving business workflows across various applications, GPT-6 Sol (at xhigh effort) outperforms Claude Opus 5 (at max effort) while costing only 9% of the price per task.
What is AutomationBench 1.0.6? It is a rigorous evaluation of AI agents performing end-to-end workflows using 47 different tools across HR, Finance, Support, Operations, Marketing, and Sales.
Note: Reported costs for Claude Fable 5.1 are often understated because they exclude the cost of Opus 5 fallbacks, which were required for of tasks.
Comparative Performance Table:
| Model (Effort Level) | Score | Cost per Task |
|---|---|---|
| GPT-6 Sol (xhigh) | 33.2% | $0.27 |
| GPT-6 Astra (low) | 30.3% | |
| Claude Opus 5 (max) | 26.9% | |
| Claude Fable 5.1 w/ Fallback (max) | 31.4% |
Additionally, on Agents’ Last Exam (which tests long-horizon tasks across 55 sub-industries), GPT-6 Sol at max effort achieved a 56.4% score, surpassing the best Claude Opus 5 result at a 60% lower cost.
✅ Factuality and Reliability
Accuracy is the cornerstone of utility. We have made substantial strides in reducing hallucinations:
- GPT-6 Sol: Reduces mistakes by approximately 50% compared to its predecessor, nearing Astra's reliability.
- GPT-6 Luna: At high effort, it matches the performance of GPT-5.6 Sol but at of the cost.
Methodology Note: These evaluations utilize de-identified ChatGPT logs where users previously flagged errors. While these "edge cases" are more difficult than typical prompts, our verbosity sweeps indicate that answer length does not significantly impact these scores.
💻 Advanced Coding Capabilities
As coding agents evolve to handle larger scopes, the cost of sustained token usage has skyrocketed. For the median researcher, daily API spend has hit 7,000.
GPT-6 Sol and Luna allow developers to be more ambitious with Codex by combining high performance with lower costs.
Benchmarks:
- FrontierCode 1.1 Main: This test measures "mergeability" (style, test quality, and codebase standards). GPT-6 Sol significantly beats GPT-5.6 Sol and matches Claude Fable 5.1
xhighat a fraction of the cost. - DeepSWE v1.1: Testing complex software engineering in real-world repositories:
- GPT-6 Sol (max): 68.8% (only 1.1% behind Claude Fable 5's 69.9%) at lower cost.
- GPT-6 Luna (max): 66.6%, performing similarly to medium-effort Claude Opus 5 or Fable 5.
# Example: The efficiency of GPT-6 Sol allows for
# more iterative loops in agentic coding workflows.
def optimize_workflow(task):
while not task.is_merged():
response = gpt6_sol.generate(task.context)
task.apply_changes(response)
# Lower cost per token = more iterations possible
