OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, two cheaper variants in the GPT-6 family. Sol costs $2 per million input tokens and $10 per million output, Luna $0.10 and $0.50 respectively — half the price of their 5.6-series counterparts. Flagship Astra gets no cut and stays at $10/$50.
Key takeaways
- GPT-6 Sol: $2 per million input tokens and $10 per million output, 50% below GPT-5.6 Sol
- GPT-6 Luna: $0.10 and $0.50 per million tokens, aimed at high-volume work
- AutomationBench 1.0.6: Sol 33.2% at $0.27 per task, Claude Opus 5 at 26.9% and eleven times the cost
- Cached input token reads up to 90% cheaper, shared prefixes valid for 30 minutes
- GPT-6 Astra keeps its price and remains the family flagship
Three price tiers in one family
| Model | Input | Output | Role |
|---|---|---|---|
| GPT-6 Astra | $10 | $50 | flagship, no cut |
| GPT-6 Sol | $2 | $10 | workhorse |
| GPT-6 Luna | $0.10 | $0.50 | high-volume work |
Cost per task instead of leaderboard position
OpenAI framed its benchmarks to show not the score but the score divided by the bill. On AutomationBench 1.0.6, a set of business tasks spanning 47 tools, Sol scored 33.2%. Claude Opus 5 stopped at 26.9% while costing eleven times as much.
In software engineering the gap narrows. On DeepSWE 1.1 Sol posts 68.8% against 69.9% for Claude Fable 5, but at roughly 80% lower cost. Luna reaches 66.6% on the same test — 93% cheaper than Opus 5.
Caching is meant to pay for the cut
OpenAI attributes the lower prices to improvements in caching and inference. Developers get explicit cache breakpoints?Cache breakpoint: A manually marked point in the prompt where the reusable prefix ends. It lets developers control exactly what gets cached., a dashboard showing hit rate?Hit rate: The share of requests served from the cache instead of being computed from scratch. The higher it is, the lower the bill., and a diagnostics tool that returns the reason for every miss as JSON.
In less than a week, our OpenAI model cache hit rate went from roughly 85% to consistently above 90%, further lowering inference costs in production.
Bin Fan, Agent Team Lead at Manus.
Permanent prices, but the baseline was promotional
An OpenAI spokesperson told VentureBeat that the Sol and Luna rates are permanent, not introductory. The company nonetheless benchmarks them against the 5.6 series’ promotional pricing, which is scheduled for a 25% increase from November 2026.
Why it matters
Two frontier launches in a single day, both sold on price — that shifts the reference point. Labs are no longer bidding on raw scores, because at the top of the leaderboards the differences have shrunk to single percentage points. What counts is how many tasks a model closes within a given budget. For teams building agents, picking a model becomes a cost-engineering decision rather than a prestige one.
What’s next
- The GPT-5.6 series faces a 25% price increase scheduled for November 2026
- Sol and Luna have not reached ChatGPT Chat yet — for now they run in ChatGPT Work and Codex
- Neither model’s context window size was disclosed in the launch announcement
Sources
- OpenAI — Introducing GPT-6 Sol and Luna
- OpenAI — Better prompt caching for GPT-6
- VentureBeat — OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more
- Simon Willison — Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
