GPT-6 Sol and Luna Are Now Live on VEGA!
OpenAI added two models to the GPT-6 family: GPT-6 Sol and GPT-6 Luna. Both are already available on VEGA. OpenAI trained them with the same methods as GPT-6 Astra, the flagship it introduced earlier this month, so the pair carries Astra's gains in professional work, factuality, coding, computer use, and alignment into cheaper, faster tiers.
Two Models for Different Scales of Work
Sol is the workhorse of the two. It is built for difficult professional tasks, agentic coding, and long workflows where you want strong results and room to iterate. Luna is the lightweight tier: fast and inexpensive, meant for everyday assistance and high-volume workloads, though still a clear step up from the previous Luna.
Astra stays the pick for the most demanding projects. The point of Sol and Luna is efficiency: they deliver a large share of that intelligence at roughly half the API price of the GPT-5.6 generation.
Benchmarks
| Benchmark | GPT-6 Sol | GPT-6 Luna |
|---|---|---|
| AutomationBench (business workflows, 47 tools) | 33.2% (xhigh) — beats Claude Opus 5 at max effort | +5.4 pp over predecessor (high) |
| Agents' Last Exam (professional workflows) | 56.4% (max) — above Opus 5's best score | — |
| DeepSWE 1.1 (software engineering) | 68.8% (max) | 66.6% (max) |
| OSWorld 2.0 offline (computer use) | 60.5% (xhigh) — on par with Opus 5 medium | exceeds GPT-5.6 Sol (medium) |
- Professional work. On AutomationBench, GPT-6 Sol at xhigh effort beats Claude Opus 5 at max effort while costing a fraction per task, and it edges out even low-effort GPT-6 Astra.
- Factuality. Sol makes about half as many factual errors as its predecessor, close to Astra-level reliability. Luna at higher effort matches the factuality of GPT-5.6 Sol at a tiny fraction of the cost.
- Coding. FrontierCode grades changes on whether they could actually merge into a real codebase; Sol improves substantially over GPT-5.6 Sol there and matches Claude Fable 5.1 xhigh. On DeepSWE it lands about a point below Fable 5's best score.
- Collaboration style. Sol and Luna inherit Astra's clearer way of talking: less jargon, fewer filler details, slightly shorter answers without losing substance. The difference is most visible in technical and coding conversations.
Better Caching for Agents and Long Conversations
OpenAI also reworked prompt caching for GPT-6. Hit rates are higher by default, a dashboard and diagnostics tool show where caching is being missed, and explicit breakpoints let developers control where cached prompt prefixes end. Reasoning effort and tool availability can be changed mid-conversation without invalidating the cache. GitHub says these changes cut the share of prompt tokens that need fresh processing by more than half across billions of Copilot requests.
Alignment
Sol and Luna inherit the alignment work shipped with Astra. In OpenAI's evaluations, both score lower rates of misleading claims about their coding work than their GPT-5.6 counterparts, with improvements across deception and circumvention tests as well.
Try GPT-6 Sol and Luna on VEGA
Both models are already in the VEGA model picker. Take Sol for the heavier professional and coding tasks, Luna for quick everyday questions, and compare them side-by-side with the other frontier models on the aggregator.