GPT-6 pricing: Astra, Sol & Luna compared

The three tiers at a glance

All figures below are USD per 1 million tokens (input / output), verified against OpenAI's official API pricing page on September 26, 2026 and cross-checked on the JetAI pricing dashboard.

The naming maps neatly to use cases: Astra is the star — maximum capability, maximum price. Sol is the workhorse sun most workloads orbit. Luna is the cheap moon you put bulk jobs on.

GPT-6 Astra — $10 / $50: when the flagship earns its keep

Astra is OpenAI's top reasoning tier, aimed at work where a wrong answer costs real money: complex multi-step agentic workflows, hard mathematical reasoning, scientific analysis, and long-horizon planning tasks. At $10 input and $50 output per million tokens, it is priced 5x above Sol — so it only makes sense when cheaper tiers measurably fail.

A crucial detail: thinking models bill their reasoning at output rates. Astra's chain-of-thought tokens cost $50/M just like final answer tokens. A long-thinking Astra call can burn several thousand output tokens before it even starts answering. Always model thinking-token volume when budgeting, not just final-answer length.

Cost-reduction paths exist: the Batch/Flex API drops Astra to $5/$25 if your workload tolerates async processing, and prompt caching discounts repeated context. If you run Astra in production, those two levers typically cut the effective bill 30–60%.

GPT-6 Sol — $2 / $10: the default for most teams

Sol is the tier most production deployments should start on. Five times cheaper than Astra, it handles customer chat, document Q&A, drafting, and mid-complexity coding work at what is usually indistinguishable-from-flagship quality on routine tasks.

The two variants worth knowing:

• Batch/Flex mode — $1/$5. Half price for non-urgent jobs: nightly report generation, bulk classification, dataset labeling. If nothing needs an answer in seconds, there is no reason to pay the interactive rate.
• Fast mode — $4/$20. A lower-latency configuration that costs double. Pay it only where latency directly affects conversion or user retention, not by default.

At $2/$10, a million-token document corpus costs $2 to read once and $10 per million tokens of generated output. For reference, that makes Sol roughly comparable in price to Grok 4.7 ($2/$6) while sitting below Claude Opus 5.5 ($4/$20).

GPT-6 Luna — $0.10 / $0.50: the bulk tier

Luna is OpenAI's cheapest current-generation tier — a full order of magnitude below Sol. It exists for one job: enormous volumes of simple work. Classification, sentiment tagging, summarization at scale, first-draft marketing copy, and data-extraction pipelines all fit.

The economics are striking: one million input tokens cost ten cents. You can run Luna over a hundred million tokens of input for $10. With Batch/Flex, that halves again to $0.05/$0.25.

The trade-off is capability. Luna is not the model for reasoning-heavy coding agents or judgment calls. The pattern that works: Luna does the first pass at scale, and only uncertain or high-value items escalate to Sol or Astra. That router architecture — cheap first pass, expensive escalation — is the single most effective cost structure in production AI, and Luna is built for the cheap end of it.

Cost per task: what the bill actually looks like

Token prices are the sticker; task prices are the bill. Here's the same job — a 50K-token document plus a 2K-token summary — across all three tiers:

Same task, 100x spread between Luna and Astra. Run 10,000 of those summaries a month and you're looking at $60 on Luna versus $6,000 on Astra. That is why the first question is always "what's the cheapest tier that handles this task reliably?" — not "what's the best model?"

Which tier should you pick?

Pick Astra when…

• The task genuinely needs frontier reasoning and cheaper tiers have been measured failing
• A wrong answer costs more than the price difference (compliance, finance, scientific work)
• You're running a small number of high-stakes decisions, not bulk volume

Pick Sol when…

• You're building production features and don't yet know your failure profile — it's the sane default
• You need a balance of quality and cost across mixed workloads (chat, docs, coding)
• You can batch non-urgent work and use the $1/$5 Flex rate

Pick Luna when…

• Volume is the story: millions of tokens of classification, extraction, or summarization
• Tasks are simple enough that quality differences don't matter
• You're building a routing architecture with Luna as the first-pass tier

And if even Luna feels expensive for your simplest jobs, compare it against the cheapest coding models — DeepSeek V4.1 Flash and Grok 4.7 compete in the same budget band with different strengths.