Sign in Subscribe
news

2026-09-22 GPT-6 Sol and Luna - same smarts as GPT-5.6, half the price

Sebastien Dubois

Sebastien Dubois

— 5 min read

Canonical version: 2026-09-22 GPT-6 Sol and Luna - same smarts as GPT-5.6, half the price.

On 22 September 2026, OpenAI released GPT-6 Sol and GPT-6 Luna, the faster and cheaper siblings of GPT-6 Astra. It landed about an hour after Anthropic shipped Claude Opus 5.5. Simon Willison covered both releases in one write-up, calling it "a new price war".

API prices drop by 50% compared with GPT-5.6, and the intelligence stays roughly where it was. For Luna, that's a good deal; for Sol, less so, and OpenAI replaced it a week later.

What shipped

There are now three tiers: Luna (cheap), Sol (middle), Astra (top). GPT-5.6 Terra has no GPT-6 successor. OpenAI says Sol and Luna were trained with methods similar to Astra's and inherit its shorter, plainer writing style.

Model Input Cached input Output Before (GPT-5.6)
GPT-6 Luna $0.10 $0.01 $0.50 $0.20 / $1.20
GPT-6 Sol $2 $0.20 $10 $4 / $20
GPT-6 Astra $10 $1 $50 unchanged

Prices per million tokens. The "before" column is GPT-5.6's promotional pricing; Simon Willison notes that GPT-5.6 has a 25% price increase scheduled for November. Luna is now one tenth of Haiku 4.5's price ($1 / $5).

Both models have a 1.05M token context window and 128K max output. Prompt caching also got better: higher hit rates by default, 90% off cached reads, explicit breakpoints, and you can now change reasoning effort or toggle tools without breaking the cache. GitHub says this cut the share of Copilot's prompt tokens needing fresh processing by more than 50%.

Availability

  • API: gpt-6-sol and gpt-6-luna, from day one
  • ChatGPT Work and Codex: Plus, Pro, Business, Enterprise and Edu, rolled out gradually through the day
  • Free and Go users: Luna only, in the desktop app
  • Regular ChatGPT chat: not available at launch

OpenAI's Thibault Sottiaux added that the 50% cut also stretches subscription usage, and that Plus, Pro and Business accounts got a banked usage reset.

The claims vs the independent numbers

OpenAI's headline numbers are all about cost per task:

  • Sol at xhigh beats Claude Opus 5 at max on AutomationBench (33.2% vs 26.9%) at 9% of the cost per task
  • Sol at max scores 68.8% on DeepSWE v1.1, 1.1 points behind Claude Fable 5, for about 80% less
  • Luna at max scores 66.6% on DeepSWE v1.1, comparable to Opus 5 at medium, for 93% less
  • Sol makes about half as many factual mistakes as GPT-5.6 Sol on OpenAI's internal eval

Artificial Analysis confirmed the cost reduction but measured smaller gains elsewhere. Intelligence Index and Coding Agent Index scores stay level with GPT-5.6. Sol gains 2 points on coding; Luna loses 2. Both hallucinate a lot less, Sol mostly because it declines to answer more often (its accuracy actually drops from 59% to 54%). Both lose ground on GDPval-AA, Artificial Analysis's benchmark of economically valuable work across 44 occupations. Cost per task: $1.06 for Sol (was $1.99), $0.07 for Luna (was $0.18).

That's GPT-5.6 performance at half the price, and in my opinion the "6" version number oversells it.

What the community said

The Hacker News launch thread (1,779 points, 800+ comments) covered several topics:

  • Luna's price drew the most praise. Simon Willison's top comment called the halving "a really big deal". Several commenters called Luna "too cheap to meter". Others pushed back: GPT-5.6 Luna was already cheap enough, and they would have preferred a smarter Luna at the old price
  • Commenters quoting Artificial Analysis saw the release as closer to a 5.7 than a 6: Luna regresses on coding, and some reported Sol doing worse than GPT-5.6 Sol in practice
  • Commenters expected Terra's removal: in their view, Terra sat in an awkward middle, beaten by Luna at xhigh or Sol at medium on cost and performance
  • Several found model and effort selection tiring: three models, six effort levels, a fast mode, and the old GPT-5.x models still in the picker. Some asked for an automatic router; others asked the model itself which model to use
  • Some expect prices to rise once the market settles; others argued inference margins leave room for cuts
  • Usage limits drove the Codex vs Claude Code subscription debate more than benchmarks did

A week later, in the GPT-6.1 Sol thread (1,064 points, 900+ comments), many commenters described GPT-6 Sol as a regression from GPT-5.6 Sol, a few defended it, and a popular (unverified) theory said OpenAI rushed out a rebranded Terra-class model and had to answer Claude Opus 5.5 with the "real" Sol. The same thread also reacted to OpenAI's new $500 per month Pro plan (ChatGPT Pro 500) and reduced allowances on the $200 Pro plan.

GPT-6.1 Sol replaced Sol a week later

On 29 September 2026, at DevDay, OpenAI released GPT-6.1 Sol: same $2 / $10 price, cached input halved again to $0.10, and near-Astra scores (it matches Astra on DeepSWE v1.1 at roughly a fifth of the cost). The API docs now call it "the newer Sol model". GPT-6 Luna stays the current small model.

My recommendations

  • Use Luna for well-specified, high-volume work: extraction, small edits, subagent coding with clear specs. OpenAI's own guide recommends the same. At $0.10 / $0.50, I see no cost reason to use another model for those tasks
  • Use GPT-6.1 Sol instead of GPT-6 Sol: same per-token price, better results. 6.1 Sol has no none effort and needs the Responses API for tool calling, so re-run your evals before switching
  • The "6" here mostly means "cheaper". Benchmark your own tasks, especially knowledge work deliverables, where both models regressed
  • On unattended agent runs, check Luna's tool output: OpenAI's own eval shows Luna hiding a broken search tool in 28.7% of adversarial cases
  • Track cost per task: both models spend more output tokens than their predecessors

All the details (specs, benchmarks, reception) are in my GPT-6 Sol and GPT-6 Luna notes.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe