OpenAI launches GPT-6 Sol and Luna with API prices cut in half, confirmed as permanent rates

Start with the numbers, because the entire weight of this story sits in the numbers. Per VentureBeat, GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens; the current GPT-5.6 Sol rate is $4 input and $20 output, so the cut is exactly 50% in both directions. GPT-6 Luna is priced at $0.10 input and $0.50 output per million tokens; GPT-5.6 Luna currently costs $0.20 input and $1.20 output, a 50% cut on input and roughly 58.3% on output. CNBC reports both companies attributed the reductions to improvements in inference efficiency and prompt caching. In other words, this is not a subsidy to buy share; the same work genuinely costs less compute.

The second thing worth pinning down is how the three tiers divide the work. CNBC reports Sol sits below Astra and is meant for more complex workloads such as coding, while Luna targets what the company calls high-volume tasks, for example extracting information or summarizing documents. In VentureBeat's framing, Sol aims at the complex work developers and knowledge workers perform repeatedly, such as building features, reviewing code, debugging and analyzing data; Luna covers more tightly defined jobs such as summarization, extraction and answering straightforward questions; Astra remains for the most complex, multi-faceted projects that cross several kinds of media, along with harder scientific and mathematical problems. OpenAI says both new models improve on their respective 5.6 predecessors on benchmarks while staying less capable than the flagship Astra released earlier this month.

The third angle is who these prices sit next to, and that is more informative than the size of the cut. Per VentureBeat's comparison, Sol's $2 input and $10 output land exactly on Anthropic's Claude Sonnet 5, whose introductory $2/$10 pricing Anthropic made permanent in August. Anthropic's Claude Opus 5.5, launched the same Tuesday morning, is priced at $4 input and $20 output, meaning Sol's output rate is half of the new Opus; Anthropic's top general-access tier, Claude Fable 5.1, costs $10 input and $50 output per million tokens, five times Sol. Further down, Luna's $0.10 and $0.50 undercut Sonnet 5 by 95% on both input and output. On Google's side, Gemini 3.8 Flash, introduced earlier this month, costs $0.75 input and $3.75 output under introductory pricing that rises to $1.50 and $7.50 on January 1, 2027, so even after the increase Gemini 3.8 Flash stays cheaper than Sol. The pricing structures differ, though: Google time-limited its current Flash rate, while OpenAI says Sol's $2/$10 has no promotional expiration.

The fourth thread is timing, and this backdrop matters more than the discount itself. CNBC reports this is the first release from either frontier lab since the debate over whether the industry should slow down reached a fever pitch. The report notes that former Anthropic researcher Jacob Coxon set off that debate by posting on X on September 8 that he had quit his job, warning that both labs were gambling with our lives, after which OpenAI chief executive Sam Altman and Tesla and SpaceX chief executive Elon Musk joined Anthropic chief executive Dario Amodei's call. CNBC also points out that both companies are working to satisfy customers looking for more cost-effective models and seeking to rein in AI spending, under pressure from firms offering cheaper open-weight models, naming three Chinese companies: Alibaba, Moonshot AI and DeepSeek.

The fifth thread is what this means for people actually building products. VentureBeat's framing is that for enterprises this segmentation matters because the economics of AI agents increasingly depend on how many model calls a workflow makes, how much context gets replayed and whether a company really needs a frontier model for every step. That is where the Sol and Luna tier lands: move the simple, repetitive, well-defined calls to the cheap tier, keep the steps that genuinely need complex reasoning on the expensive one, and the bill difference will come from how you allocate calls rather than from model capability gaps. Anthropic's numbers cross-check the point: cache reads on Opus 5.5 fall to $0.20 per million tokens, 60% below Opus 5's $0.50, and cache reads are exactly where the bulk of cost lives in agentic and coding workloads. In other words, how efficiently you reuse context may matter to your bill as much as which model you pick.

Two things remain to watch. First, a permanent price is not the same as a free lunch: both new models stay below Astra in capability, and which step you can downgrade and which you cannot is something only your own evaluation run will tell you, since public benchmarks are best used to rule out clearly unsuitable options. Second, the fact that calls for a slowdown coexist with this release cadence is itself the story: two labs shipping a new model each on the same day indicates the pace of competition has not slowed because of public debate. Every fact and figure here comes from VentureBeat, CNBC and The New Stack reporting, with no speculation added.