Model Card: GPT-6 Sol / Luna

🌏 中文版

Model Info

Highlights

• Prices are cut a further 50% versus GPT-5.6 promotional rates: Sol drops from $4/$20 to $2/$10, Luna from $0.20/$1.20 to $0.10/$0.50, while keeping the same 1.05M-token context window as flagship Astra
• On AutomationBench (agent tasks across 47 tools spanning business workflows), Sol scores 33.2% at xhigh effort — 6.3 points above Claude Opus 5's best of 26.9% — at roughly 9% of Opus 5's per-task cost (i.e. Opus 5 costs about 11x more)
• Alignment improved noticeably: in an internal simulated Codex deployment across 50,319 tasks, severity-3+ misalignment flags dropped from 66 (GPT-5.6 Sol) to 42; on a simulated message-board test, the rate of taking unauthorized action fell from 52% to 11%
• Prompt-caching discounts now reach up to 90%, and reasoning effort and tool toggles can be changed without invalidating the cache; OpenAI says this let GitHub cut the share of prompt tokens needing fresh processing by more than 50%

Benchmark Results

⚠️ Except where OSWorld 2.0's public version number is cited, all figures above are OpenAI's own self-reported numbers, each vendor's model run at its own highest effort setting, and not yet independently reproduced; the misalignment-flag counts come from OpenAI's internal simulated deployment, not a public benchmark.

Versus Prior Model and Competitors

Versus the GPT-5.6 generation, Sol and Luna's gains are mostly about price-performance rather than raw scores — OpenAI doesn't publish absolute GPT-5.6-generation numbers for most of the new benchmarks, but repeatedly emphasizes matching or beating them at a fraction of the cost: Luna's AutomationBench cost is down 58%, and on OSWorld 2.0 Luna beats prior-generation flagship GPT-5.6 Sol's medium-effort score at roughly a tenth of the cost. That's a different story from flagship Astra, which competes on raw score ceilings (ARC-AGI-3, FrontierMath); Sol and Luna are about bringing Astra's techniques to everyday workloads at much lower cost.

Against Claude, Sol edges out Claude Opus 5 on agent-style benchmarks (AutomationBench, Agents' Last Exam, OSWorld 2.0) at 10-20% of the cost. But on the heaviest benchmark, DeepSWE v1.1 software engineering, Sol's 68.8% still trails Claude Fable 5.1's 69.9% — it just gets within 1.1 points at roughly 80% lower cost. On safety classification, Sol and Luna are rated "High" (below Critical) for cybersecurity and biological/chemical risk, versus flagship Astra's "Critical" rating — OpenAI is deliberately keeping its strongest capabilities in Astra while Sol and Luna take the good-enough, lower-risk, mid/low-price route.

Worth noting: all three models (Astra, Sol, Luna) share the exact same 1,050,000-token context window and 128,000-token output cap. OpenAI isn't tiering by context length this time — it's tiering purely by price and benchmark quality, a departure from the usual pattern of reserving the largest context for the flagship.

What This Means for Agent Development

The biggest implication here is that choosing a model is becoming choosing a reasoning effort level: Astra, Sol, and Luna share an identical context window, input/output modalities, and API surface — the differences are mainly price and benchmark score — and reasoning effort can now be adjusted in real time without breaking the prompt cache.

• If you're building multi-agent orchestration systems: Sol can replace Astra as your "trusted workhorse" model for subtasks — its AutomationBench and OSWorld 2.0 scores now match or edge out Claude Opus 5, at 10-20% of the cost — and reserve Astra for the few nodes that genuinely need top-tier scores
• If you're running high-volume, structurally fixed tasks (classification, extraction, simple tool routing): Luna's 66.6% on DeepSWE v1.1 approaches last generation's flagship level, and its $0.10/$0.50 pricing changes the economics of large-scale repeated calls entirely
• Not a fit: workloads that need Critical-tier capability or the absolute score ceiling (research-grade math, abstract reasoning in unseen environments) — those still call for Astra or a cross-vendor comparison against Claude Fable 5.1; Sol and Luna still trail by roughly a point on heavy engineering tasks like DeepSWE
• Concrete architecture tip: take advantage of adjusting reasoning effort without invalidating the cache — route the same agent through different effort levels based on task complexity instead of switching models entirely, saving tokens while keeping cache hit rates high

Today's Takeaway

This release doesn't raise the capability ceiling — Astra remains the top scorer. But Sol edges out Claude Opus 5 on several agent benchmarks at 10-20% of the cost, and Luna closes in on last generation's flagship-level software-engineering performance at roughly 1% of the cost. Reshuffling the entire product line's price-performance within three months is the real story here: beyond watching who tops the leaderboard, what actually matters for a production agent system deciding whether to switch models is which direction the price-performance curve is moving at your specific price point — not who's currently ranked first.

References

• OpenAI: Introducing GPT-6 Sol and Luna
• OpenAI Developer Community: Announcing GPT-6 Sol and GPT-6 Luna
• OpenAI API Docs: Pricing
• OpenAI API Docs: GPT-6 Sol model details
• OpenAI API Docs: GPT-6 Luna model details
• ZDNET: OpenAI's GPT-6 Sol doubles its accuracy rate – for half the cost
• Pulse2: OpenAI Launches GPT-6 Sol And Luna With 50% Lower API Pricing
• TechJournal: GPT-6 Sol and Luna Launch With Half-Price API Rates