Analysis
OpenAI released GPT-6 Sol and GPT-6 Luna this week, cutting API prices by roughly half compared with the prior GPT-5.6 generation while claiming meaningfully fewer factual errors, according to OpenAI's own announcement and VentureBeat's coverage. Sol is priced at $2 per million input tokens and $10 per million output tokens, roughly half GPT-5.6 Sol's rate; Luna costs $0.10 input and $0.50 output per million tokens, a drop of about 58% on the output side.
From Astra's Rocky Rollout To A Pricing Reset
This launch follows a rougher stretch for the GPT-6 family. OpenAI's GPT-6 Astra rollout in early September drew criticism over cybersecurity classification gaps and prompted a public apology from Sam Altman -- a very different kind of news cycle than the one Sol and Luna are generating now. Sol and Luna succeed the GPT-5.6 family (Sol, Terra and Luna), which launched in July 2026, and the naming continuity suggests OpenAI is iterating on a stable model lineup rather than resetting the brand with every release.
“## From Astra's Rocky Rollout To A Pricing Reset This launch follows a rougher stretch for the GPT-6 family.”
A Three-Way Price War, Not A Coincidence
Three frontier labs cut API prices within days of each other this week. The comparison is stark once you line the numbers up:
- OpenAI GPT-6 Sol -- $2/$10 per million input/output tokens, roughly half GPT-5.6 Sol's price.
- Anthropic Claude Opus 5.5 -- $4/$20 per million tokens list price, with cache reads down 60% to $0.20/million and typical workloads about 40% cheaper overall.
- xAI Grok 4.7 -- $2/$0.50/$6 per million tokens (input/cached input/output) below a 200K-token prompt, with a 500K context window.
No two labs are pricing the exact same way -- Sol's cut is a flat headline number, Anthropic's steepest cut is to cache pricing specifically, and Grok 4.7 uses tiered pricing by context length -- but all three moved in the same direction in the same week, which is a market response, not a scheduling accident.
What The 'Half The Price' Headline Misses
Artificial Analysis, an independent benchmarking group, found Sol and Luna roughly flat against GPT-5.6 on its overall Intelligence Index, and actually behind GPT-5.6 Sol's best score on two of three coding and computer-use charts. OpenAI's own 41-54% fewer-errors claim is also measured only on de-identified conversations where a user had already flagged a factual mistake from a prior model -- a sample deliberately weighted toward errors, not representative of typical usage where mistakes are rarer to begin with. The headline number is real; it is not evidence of a broad capability jump.
What It Means Going Forward
For startups building on any of these APIs, the direct effect is a lower cost of goods sold for the third or fourth time this year -- a genuine, compounding margin tailwind regardless of which lab ultimately wins the capability race. For the labs themselves, none has disclosed gross margin at these new price points, so it remains an open question whether this is healthy competition passing through real infrastructure cost reductions, or a subsidized race that none of the three can sustain indefinitely.
The next signal worth tracking is whether any of the three ships a genuine capability-frontier model -- rather than a cost-and-efficiency release -- on a similarly tight calendar.
