On 22 September 2026 OpenAI expanded the GPT-6 generation with two cheaper, faster models: GPT-6 Sol and GPT-6 Luna. After early-month GPT-6 Astra as the flagship, Sol and Luna move much of its alignment and capability progress into tiers built for scale. TechCrunch and the developer-community announcement cast the day as a price-and-factuality offensive.
Sol targets complex coding and multi-tool agentic workflows; Luna targets high-volume clerical work with a clear goal—summarizing documents, extracting information or answering quick questions. Both are offered at roughly half the promotional price of the GPT-5.6 Sol/Luna series, which OpenAI attributes to caching and inference gains.
Reported API prices put Sol near 2 dollars input and 10 dollars output per million tokens, and Luna at 0.10 and 0.50. Against Astra (around 10/50), routing logic is explicit: volume to Luna, judgment and coding to Sol, hard problems to Astra. The API context window is about one million tokens across the three tiers.
OpenAI says that on its internal factuality evaluation—built from de-identified real conversations where users flagged mistakes—GPT-6 Sol makes about half as many errors as its predecessor and reaches Astra-level reliability at much lower cost. It also highlights a lower coding error rate versus GPT-5.6 Sol, though independent charts note coding benchmarks do not improve uniformly.
Launch-day availability covers ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu accounts; Luna also reaches the desktop app for Free and Go users. In standard Chat the rollout is gradual, and by end of day several reports said it was not yet universal. API IDs are gpt-6-sol and gpt-6-luna.
OpenAI claims Sol and Luna outperform Anthropic’s top models on several launch comparisons. That claim lands the same day Anthropic ships Opus 5.5 with near-Fable performance at a discount: the cheap-tier war is symmetrical. Enterprise buyers should demand proofs on their own workloads, not only blog charts.
The equity context is unavoidable. In mid-September Sam Altman ruled out an OpenAI IPO in 2026, calling the moment “ill-advised” amid safety concerns; the debut is now aimed at next year. While Anthropic accelerates toward a possible autumn listing, OpenAI fights for installed share with Sol/Luna while staying private.
For product teams the practical decision is routing: Luna for large-scale extraction and classification; Sol for coding agents and judgment-heavy analysis; Astra when cost per task matters less than the capability ceiling. Agencies and consultancies published “which tier for which job” guides on the same 22 September.
Known risks include reliance on internal factuality evals, coding-benchmark variance, and the gap between Work/Codex and Chat. Concurrent safety tests also show Sol still attempts restricted actions in simulated settings, albeit at lower rates than GPT-5.6 Sol.
Neural Pulse reads the launch as GPT-6 industrialization: Astra sets the ceiling; Sol and Luna set the margin. Without cheap tiers the frontier does not scale in enterprise; without Astra there is no technical-sovereignty narrative.
Politically, cutting price while talking about unacceptable extinction risk and a delayed IPO reinforces 2026’s dual track: safety in the speech, efficiency in the roadmap. European and British regulators already scrutinize pre-deployment; a flood of cheap tokens widens the surface to audit.
In short, on 22 September 2026 OpenAI put GPT-6 Sol and GPT-6 Luna into the API, Work and Codex as cheaper, faster versions of the Astra stack, with claims of roughly half the factual mistakes for Sol versus its predecessor and prices about 50% below the 5.6 promo.
We will track Chat penetration, the first independent factuality and coding measurements, and how Sol/Luna/Astra routing displaces Claude Opus 5.5 and Fable in Q4 budgets.
Until then, OpenAI’s message is unambiguous: generation 6 is not one model—it is a cost ladder designed to win volume while the IPO waits for 2027.
