GPT-6 Sol vs Luna: Which Tier for Which Task

1. The lineup, with verified prices

Earlier this month OpenAI released GPT-6 Astra, its most capable model ever broadly deployed — and its first to reach the Critical level of cybersecurity capability under the Preparedness Framework. Sol and Luna inherit Astra’s training advances (factuality, coding, computer use, alignment) and trade peak capability for speed and price.

The headline is the price cut: GPT-6 Sol costs $2 input / $10 output per 1M tokens (vs $4/$20 for GPT-5.6 Sol) and Luna costs $0.10/$0.50 (vs $0.20/$1.20). OpenAI attributes the drop to better caching and inference, and passes the savings straight through.

2. Astra vs Sol vs Luna

Think of the three as depth settings on the same intelligence. Astra is full depth for the most demanding projects: hard reasoning, autonomous computer use, safety-critical review. Sol is the general-production default: multistep coding, long-context analysis, tool workflows. Luna is efficient intelligence for extraction, summarization, routing and routine interactions.

Microsoft’s Foundry framing matches: Astra for demanding reasoning, Sol as the starting point for production workloads and migrations off legacy models, Luna for high-volume data and preparatory work. Reserve deep reasoning for the steps that need it instead of applying one model to everything.

The three tiers

3. Evidence and availability

On factuality, OpenAI’s internal evaluation — built on de-identified real conversations where users flagged mistakes — finds GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost. Coding error rates are down as well. These are vendor numbers; treat them as directional until independent evals land.

Availability is broad but uneven. ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; Luna also in the desktop app for Free and Go users. API from day one. GitHub Copilot: Sol on Pro+ and up, Luna from Pro. Azure Foundry: Standard across 28 regions plus US/EU data zones. Not in Chat yet — rollout is gradual, and Enterprise admins must enable the models.

4. The five-line router

The practical move is a task-tier router in front of your calls: classify the step, pick the cheapest tier that can do it, escalate only on low confidence. Below is the shape — adapt the thresholds to your eval data, not to launch marketing.

Start with Luna as the default for summarization, extraction and routing; Sol for anything that writes or edits code; Astra for planning, review and anything autonomous. Log tier, tokens and outcome per call for two weeks, then tune.

def pick_model(step):
    if step.needs_depth or step.is_autonomous:
        return “gpt-6-astra”      # review, planning, agents
    if step.writes_code or step.uses_tools:
        return “gpt-6-sol”        # $2/$10 coding default
    return “gpt-6-luna”             # $0.10/$0.50 volume tier

# escalate on low confidence:
# if result.confidence < 0.8: retry one tier up

5. What to do this week

If you run OpenAI workloads, this is a pricing event first and a capability event second. Audit where your tokens go by tier, then move the cheap-shaped work down before touching anything else.

Verdict

Sol and Luna are exactly what OpenAI claims: Astra’s training advances repackaged into faster, dramatically cheaper tiers, with Sol halving factuality mistakes and Luna resetting the floor to ten cents. The lineup finally has a sensible default for every workload shape.

Adopt the router, not the hype: Luna for volume, Sol for code, Astra for depth — and run your own three-way eval against Opus 5.5 and Grok 4.7 before declaring a winner.