OpenAI Unveils GPT-6 Sol and Luna with API Costs Slashed by Half and New Benchmark Results
OpenAI launched GPT-6 Sol and GPT-6 Luna, two API-only models positioned below flagship GPT-6 Astra. Sol handles complex coding and professional tasks; Luna powers fast, high-volume work. Both replace GPT-5.6 promotional pricing with 50% cost cuts—Luna output drops roughly 58%. Benchmarks show Sol beating Claude Opus 5 on AutomationBench and Agents' Last Exam at far lower per-task costs, while Luna matches premium models for up to 96% less. New prompt caching delivers up to 90% discounts on cached inputs, with a dashboard and diagnostics tool for developers.
If you build with AI models and have been watching API costs climb, OpenAI's newest releases deserve your attention. The company has unveiled GPT-6 Sol and GPT-6 Luna, two additions to the GPT-6 family that slot beneath GPT-6 Astra, which arrived earlier this month. Both models were trained using approaches similar to Astra's, with the goal of delivering Astra-level improvements in packages that run faster and cost less. In this article, we'll break down what each model is for, what the benchmarks actually show, and how the new pricing and caching features could affect your bottom line.
Three Tiers, One Recipe
The GPT-6 lineup now covers three levels. Astra remains the flagship for the most demanding workloads. Sol is aimed at complex coding and professional tasks where cost still matters. Luna handles fast, high-volume everyday work.
Both models are deployable right now. They're live in the OpenAI API under the names gpt-6-sol and gpt-6-luna. Note that these are API-only releases, so there are no weights available for self-hosting.
According to OpenAI, improvements in caching and inference efficiency are what allow the company to serve these models at lower prices. As a result, Sol and Luna API prices are being cut by 50% compared to the GPT-5.6 promotional pricing they replace.
Here's the full pricing picture per 1M tokens:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-6 Sol | $2.00 (was $4) | $10.00 (was $20) |
| GPT-6 Luna | $0.10 (was $0.20) | $0.50 (was $1.20) |
One nuance worth flagging: Luna's output price actually drops from $1.20 to $0.50, which works out to roughly a 58% reduction rather than a flat 50%.
What the Benchmarks Show
OpenAI has published results across several areas, and the numbers are worth examining closely.
Professional work: On AutomationBench 1.0.6, Sol running at xhigh effort scored 33.2% at $0.27 per task. For comparison, Claude Opus 5 at max effort scored 26.9% while costing 11.1 times more per task. Even Astra at low effort managed 30.3%, but at 3.9 times Sol's cost. Meanwhile, Luna at high effort improved 5.4 points over its predecessor while spending 58% less per task. On Agents' Last Exam, Sol at max effort hit 56.4%, beating Claude Opus 5's best score at 60% lower cost per task.
Coding: On DeepSWE v1.1, Sol at max effort reached 68.8%, just 1.1 points behind Claude Fable 5 at xhigh, at roughly 80% lower cost per task. Luna at max effort scored 66.6%, which is comparable to Opus 5 and Fable 5 at medium effort — while costing 93% less per task than Opus 5 and 96% less than Fable 5 in those comparisons. On FrontierCode 1.1 Main, a benchmark that measures whether code is genuinely ready to merge, Sol matched Claude Fable 5.1 at xhigh at a much lower price point.
Computer use: On OSWorld 2.0 offline, Sol at xhigh scored 60.5% versus 60.3% for Opus 5 at medium, with per-task costs about 80% lower. Luna at max effort outperformed GPT-5.6 Sol at medium effort for one-tenth of the cost.
Factuality: OpenAI's internal testing used de-identified ChatGPT conversations where users had flagged model errors. Sol made roughly half as many mistakes as its predecessor, and Luna at higher effort matched GPT-5.6 Sol at about 1/100 of the cost.
OpenAI also carried Astra's communication style into these models. Expect clearer, slightly shorter answers with less jargon, particularly in coding conversations.
Smarter Prompt Caching for Long-Running Agents
Agents tend to resend the same instructions, tool definitions, and conversation history on every single turn, which adds up fast. GPT-6 ships with an improved prompt caching system that delivers higher cache hit rates by default. Cached input reads receive discounts of up to 90%, and eligible shared prefixes reused within a 30-minute window now qualify.
Developers also get a set of new controls:
- A Prompt Caching Dashboard tracks hit rates over time.
- A diagnostics tool explains why a cache missed — for example, `"reason": "tools_changed"`.
- Explicit breakpoints let you choose exactly where a cached prefix ends.
- Reasoning effort can now change mid-conversation via `configuration_update` without invalidating the cache.
- allowed_tools restricts which tools are callable while keeping definitions stable, preserving the cache.
- Prewarming prepares known context before the first user request arrives.
OpenAI's full prompt caching guide documents each of these patterns. GitHub, an early beneficiary, reports that these changes reduced the share of prompt tokens requiring fresh processing by more than 50%, helping Copilot respond faster.
Where You Can Use Them
- API: gpt-6-sol and gpt-6-luna are available now.
- ChatGPT Work and Codex: Available to Plus, Pro, Business, Enterprise, and Edu users.
- Free and Go tiers: Luna is available in the ChatGPT desktop app.
- Not yet in Chat: The ChatGPT rollout is gradual throughout launch day, so Chat access may take time to appear.
Key Takeaways
- Sol costs $2/$10 and Luna $0.10/$0.50 per 1M tokens.
- Sol at xhigh effort beats Opus 5 at max on AutomationBench at just 9% of the cost.
- Luna scores 66.6% on DeepSWE v1.1, costing 93% less per task than Opus 5.
- Cached input reads get discounts up to 90%, with new cache controls for developers.
- Both models are live in the API, ChatGPT Work, and Codex — but not yet in Chat.
Comments (0)