GPT-6 Sol and Luna Cut API Prices 50%—Benchmarks Included

OpenAI has added two models to its GPT-6 family — GPT-6 Sol and GPT-6 Luna — both live immediately in the API as gpt-6-sol and gpt-6-luna. They sit below GPT-6 Astra, which launched earlier in September, and share its training methods. The headline figure is a 50% price cut against GPT-5.6 promotional pricing for Sol; Luna's output token price drops from $1.20 to $0.50 — approximately 58%, not 50%. Both are API-only: there are no weights to self-host, so AI portability constraints remain in play for any organisation needing deployment flexibility beyond OpenAI's infrastructure.

Pricing and tier structure

The GPT-6 family now spans three tiers: Astra for the most demanding workloads, Sol for complex coding and professional tasks, and Luna for fast, high-volume everyday work. OpenAI attributes the price reduction to improved caching and inference efficiency.

Benchmark performance

OpenAI published results across professional workflow, coding, computer-use, and factuality evaluations.

Professional work. On AutomationBench 1.0.6, Sol at xhigh effort scores 33.2% at $0.27 per task, outperforming Claude Opus 5 at max effort (26.9%) at 11.1x that cost. Low-effort Astra scores 30.3% at 3.9x Sol's cost. On Agents' Last Exam, Sol at max effort scores 56.4%, beating Claude Opus 5's best score at 60% lower cost per task.

Coding. On DeepSWE v1.1, Sol at max effort scores 68.8% — 1.1 points behind Claude Fable 5 at xhigh, at approximately 80% lower cost per task. Luna at max effort scores 66.6%, 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.

Computer use. On OSWorld 2.0 offline, Sol at xhigh scores 60.5% versus 60.3% for Opus 5 at medium, at approximately 80% lower cost per task.

Factuality. Sol makes about half as many errors as its predecessor on OpenAI's internal test. Luna at higher effort matches GPT-5.6 Sol at approximately 1/100th of the cost.

Prompt caching changes

GPT-6 ships an improved prompt caching system with higher default cache hit rates. Cached input reads receive discounts of up to 90%; eligible shared prefixes reused within a 30-minute window qualify. GitHub reports these changes cut the share of prompt tokens requiring fresh processing by more than 50% for Copilot workloads.

New developer controls include: a Prompt Caching Dashboard for tracking hit rates; a diagnostics tool surfacing miss reasons (e.g. "reason": "tools_changed"); explicit breakpoints for defining where a cached prefix ends; mid-conversation reasoning effort adjustment via configuration_update without breaking the cache; allowed_tools for restricting callable tools while keeping definitions stable; and a prewarming mechanism for preparing known context before the first user request.

AI Mastery analysis

The cost ratios here are large enough to shift architectural decisions, not just line-item budgets. Luna at $0.10/$0.50 per million tokens is priced to compete with open-weight inference costs when hosting overhead is factored in — though the absence of self-hostable weights means that comparison holds only at the per-token level, not at the infrastructure-control level.

The benchmark framing warrants scrutiny. Every cost-per-task comparison uses different effort settings between Sol/Luna and competitors: Sol at xhigh against Opus 5 at max, or Luna at max against Fable 5 at medium, are not controlled comparisons. Practitioners should run task-specific evaluations before committing to a model switch — consistent with the broader pattern we've observed in production AI failing on architecture rather than raw model intelligence.

The caching changes are arguably as consequential as the pricing for agentic workloads. Agents that resend identical system prompts, tool definitions, and conversation history on every turn now have a structured path to 90% discounts on those repeated tokens, provided prefix structure stays stable. The allowed_tools control — restricting callable tools without altering definitions — directly addresses the cache invalidation problem that has made tool-heavy agents expensive. For teams building multi-turn agents with complex control layers, that single mechanism may reduce effective inference costs more than the headline price cut.

Availability is tiered: Plus, Pro, Business, Enterprise, and Edu users access both models via ChatGPT Work and Codex; Free and Go users receive Luna in the desktop app. Neither model is yet available in the standard Chat interface. Both are live in the API immediately.

Primary source

OpenAI Releases GPT-6 Sol and Luna: 50% Cheaper API Pricing and Benchmarks — MarkTechPost