Half the Price, Nine Tents of Astra: OpenAI's GPT-6 Sol and Luna Reset the Cost of Frontier Work

Ninety minutes. That is roughly how long Anthropic’s Claude Opus 5.5 held the “newest frontier model” title on September 22, 2026, before OpenAI walked out and dropped two models of its own. GPT-6 Sol and GPT-6 Luna are not a new flagship — GPT-6 Astra still owns that seat — but they may be the more consequential release, because they take a large share of Astra’s capability and staple it to a price tag that halves everything OpenAI shipped in the GPT-5.6 generation.

The move matters well beyond the benchmark-watching crowd. If 2026’s first half was about who could train the smartest model, its final months are shaping up as a war over who can serve intelligence the cheapest. With Sol and Luna, OpenAI just moved that battle to a new front line.

Two models, one strategy

Sol is the serious sibling. It slots in as OpenAI’s workhorse model for agentic coding and heavy knowledge work, and OpenAI positions it as bringing “much of Astra’s strengths into faster and more affordable models to support work at scale.” Luna is the volume play — a fast, cheap model aimed at classification, routing, extraction, and the long tail of high-volume tasks that never justified frontier pricing.

The pricing is the headline. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4/$20 for GPT-5.6 Sol. GPT-6 Luna lands at $0.10 per million input and $0.50 per million output — a staggering drop from the $1/$6 that Luna commanded just three months ago, before the July repricing took it to $0.20/$1.20. Luna has now fallen roughly 90% from its summer pricing in under a quarter. Cached input for both models runs at 10% of the uncached rate, Batch and Flex tiers halve standard rates, and prompts above 272K tokens carry long-context surcharges (2x input, 1.5x output) rather than a flat wall.

OpenAI is explicit that this is efficiency being passed downstream: the company says it made caching and inference more efficient and is “passing the savings directly to you” as a 50% cut against GPT-5.6 pricing.

What the numbers say

Both models carry the 1,050,000-token context window and 128,000-token output limit familiar from the GPT-6 family. The capability story is best told in per-task economics, where OpenAI’s own figures do the talking: GPT-6 Luna improves on its predecessor by 5.4 percentage points on OpenAI’s task benchmark while costing 58% less per task. At the aggregate level, Luna costs 93% less per task than Claude Opus 5 and 96% less than GPT-6 Astra. On effort scaling, GPT-6 Sol at “xhigh” effort posts a similar score to Claude Opus 5 running at medium effort — a comparison OpenAI clearly enjoys making.

The fine print deserves a skeptical read. Independent observers note that GPT-6 Sol’s benchmark gains are partly “benchmaxxed” — tuned for the public leaderboards — and that Anthropic’s same-day Opus 5.5 beats Astra itself on some coding and knowledge benchmarks. One independent analysis gave GPT-6 Astra the same Intelligence Index score as GPT-5.6 Sol despite Astra’s much higher price, a hint that OpenAI’s own tiers were converging from both directions. The honest summary: Sol is a genuine step up from GPT-5.6 Sol for agent work, but the frontier race at the very top remains contested between Astra, Opus 5.5, and Anthropic’s Fable line.

There are also qualitative wins that benchmarks undersell. In OpenAI’s side-by-side coding example, GPT-6 Sol reports what it checked and what it did not — desktop layout, narrow mobile, browser back navigation — where GPT-5.6 Sol tended to volunteer implementation trivia like its internal image-generation prompt. For agentic work, that kind of calibrated self-reporting is worth more than a point or two of SWE-bench.

Availability and rollout

Sol and Luna are live now in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, with Luna also rolling out to the desktop app. Both are available in the API, and OpenAI has committed to expanding free-tier access over time. The rollout is gradual by region, as has become standard practice.

Developers reacted to the pricing almost viscerally. “Luna is now so cheap it’s insane,” one Reddit commenter put it — noting the >90% cumulative output-price reduction since the original 5.6 Luna. Another thread simply declared that “intelligence too cheap to meter” is finally becoming true. The per-task cost curves are the real story: workloads that were uneconomical to agent-ify in June are routine line items in September.

The pricing war has a new logic

The strategic context is what makes this release more than a routine model drop. Three structural forces are converging:

Inference is getting cheaper faster than models are getting better. The gap between the flagship and the tier below has narrowed to the point where, for most commercial workloads, paying flagship prices is an engineering smell rather than a necessity.

Agents multiply token consumption. A single agent session can consume 100x the tokens of a chat turn. At GPT-5.6 prices that capped agent deployment economics; at Luna prices it does not. The $65 million-a-year inference deals and billion-dollar run rates reported across the industry this month only work if the marginal cost of a task keeps falling.

Anthropic set the trap, OpenAI sprung it. Opus 5.5 arrived with a 40% cost cut and record safety scores, a direct challenge on price. Ninety minutes later, OpenAI answered with a 50% cut across two tiers. Whatever your benchmark preference, buyers are the winners: frontier-adjacent capability is now priced like a commodity input.

The losers, medium-term, are anyone whose business model assumed durable per-token margins — a category that increasingly includes the model labs themselves. OpenAI’s answer appears to be volume: make every task cheap enough that the number of tasks explodes.

What to watch

Watch three things in the coming weeks. First, whether Anthropic responds with a Sonnet-class repricing — its Fable 5.1 line is now the expensive option in several workload classes. Second, whether Google’s Gemini 3.8 Flash pricing holds, since Flash-tier models are Luna’s most direct competitor. Third, whether OpenAI’s free-tier expansion changes app economics — free access to Sol-class agents in ChatGPT Work would pressure every startup charging per-seat for AI features.

For developers, the practical advice is simple: re-run your cost models. If you priced an agent workflow in the GPT-5.6 era, the same workflow on Sol costs half, and the routing tier on Luna costs a tenth. The frontier didn’t just get smarter this week — it got dramatically cheaper, and that compounding is the real news.