OpenAI has launched GPT-6 Sol and GPT-6 Luna at roughly half the API cost of their predecessors, leaning on price as its main competitive lever against Anthropic and a widening field of open-weight challengers.

OpenAI introduced the two models on September 22, 2026, stating they were trained with the same method as GPT-6 Astra released earlier in the month. GPT-6 Sol is positioned for demanding enterprise workloads, with relaxed usage limits and enough cost headroom for iterative experimentation, while GPT-6 Luna targets high-volume, lower-complexity deployment. Per-token pricing sets Sol at $2 per million input tokens and $10 per million output tokens, half the GPT-5.6 Sol rate. Luna lands at $0.10 input and $0.50 output per million tokens, also half of GPT-5.6 Luna. That gap is stark against Anthropic's Claude Fable 5.1, priced at $10 input and $50 output per million tokens. Google is running promotional pricing on Gemini 3.8 Flash through December 2026 at $0.75 input and $3.75 output per million tokens. OpenAI says Sol's coding performance is comparable to Claude Fable 5.1 at substantially lower cost. Analysts cited by the outlet point to Chinese open-source vendors delivering models competitive with frontier systems, plus the broader rise of small and open models, as the real source of downward pricing pressure. Omdia's Lian Jie Su frames the cuts as intended to accelerate adoption and support OpenAI's IPO path; the company filed initial paperwork in June 2026. Futurum Group's Bradley Shimmin warns that the discounts may not be sustainable when vendors ship new models nearly every other week, and that optimization rather than price will become the main differentiator. Shimmin also argues the shift toward practical enterprise value, not investor optics, is the credible signal from these releases.

Price cuts of this magnitude reframe the enterprise buying conversation. For CIOs weighing frontier API costs, the comparison is no longer just OpenAI versus Anthropic on benchmark parity, but OpenAI versus Anthropic versus a Gemini promotional rate versus a self-hosted open-weight alternative. That is a structurally different market from the one where frontier labs competed on capability alone. Open-weight compression has been eroding the price umbrella under proprietary models for over a year, and OpenAI's move reads less as a capability claim than as a defensive re-pricing to hold volume in mid-tier enterprise deployments. The pre-IPO context matters: revenue growth and usage breadth now carry more weight than nominal margin per token, and a halved price at comparable quality is a volume play. Our prior coverage noted that the September 24 reporting traced much of the touted 50% cut to retiring a mid-tier model and relabeling it rather than deep architectural cost work, a distinction worth holding onto here.

For builders, the practical implication is that model routing just got more interesting and more fragile. Teams that locked in single-vendor contracts around previous price sheets should re-run cost models against Sol, Luna, Gemini 3.8 Flash promotional rates, and their own open-weight fallbacks. For investors, the signal is that inference-level price competition is now a live line item in frontier-lab economics, and the labs that can cut cost at the architecture level, rather than through tier retirement and promotional windows, will defend margin better when the pricing war cools. Watch whether Anthropic responds, and whether open-weight vendors keep narrowing the quality gap faster than proprietary labs cut price.