The Launch in 30 Seconds
On September 22, 2026, OpenAI added two models to the GPT-6 family: GPT-6 Sol and GPT-6 Luna. Both bring improvements from the flagship GPT-6 Astra down to much cheaper price points — Sol at $2/$10 per million tokens and Luna at $0.10/$0.50, each roughly half the cost of their GPT-5.6 predecessors. They’re rolling out in ChatGPT Work and Codex now.
The timing was pointed: Anthropic launched Claude Opus 5.5 about 90 minutes earlier the same day. September 22 was the day the AI price war went public.
What Sol and Luna Are
OpenAI positions the pair as bringing Astra’s training advances — better factuality, coding, computer use, and alignment — to faster, lower-cost models:
- GPT-6 Sol — the middle option for demanding everyday work. Complex coding, agentic workflows, professional tasks that need real reasoning. OpenAI’s system card calls it a highly capable, lower-cost alternative to Astra.
- GPT-6 Luna — the high-volume option. OpenAI’s fastest and most cost-efficient model, built for focused tasks at scale where per-task cost dominates.
One notable absence: there’s no GPT-6 Terra. Terra was the awkward middle tier of the 5.6 generation, and once Luna’s pricing improved, it had no clear role left. OpenAI hasn’t said whether a successor is planned.
Pricing: The Numbers
API pricing per million tokens at standard rates:
| Model | Input | Output | vs predecessor |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | 50% below GPT-5.6 Sol ($4/$20) |
| GPT-6 Luna | $0.10 | $0.50 | 50%+ below GPT-5.6 Luna ($0.20/$1.20) |
Cached input reads get a 90% discount, and batch processing halves the rates again. For developers running Codex sessions — where long agentic runs chew through tokens — these cuts change the economics of leaving a coding agent running for hours.
Do They Actually Perform?
OpenAI’s claims (vendor benchmarks, treat accordingly):
- Sol makes about half as many factuality mistakes as GPT-5.6 Sol, reaching what OpenAI calls “Astra-level reliability at much lower cost.” It scored 68.8% on DeepSWE 1.1 at maximum reasoning effort.
- Luna scored 66.6% on DeepSWE 1.1 — strong for a budget model.
- On OpenAI’s AutomationBench, Sol at high effort beat Claude Opus 5 at max effort for roughly 9% of the cost per task — though that comparison predates Opus 5.5’s same-day price cut.
Independent evaluation adds nuance: Artificial Analysis found Luna scored 2 points lower than GPT-5.6 Luna on their Coding Agent Index at max effort, with regressions on specific coding benchmarks — while costing roughly 60% less per task. Sol gained 2 points on the same index. The honest summary: Sol got better and cheaper; Luna got much cheaper with a small capability trade-off.
On safety classification, both sit at High (not Critical) for cybersecurity and biological/chemical capabilities under OpenAI’s Preparedness Framework — below Astra’s Critical cyber rating.
Where You Can Use Them
The rollout is split by plan, and this is where people get confused:
- Plus, Pro, Business, Enterprise, Edu: both Sol and Luna in ChatGPT Work and Codex
- Free and Go: Luna only, via the ChatGPT desktop app
- Standard ChatGPT chat: neither model yet — OpenAI hasn’t given a date
- API:
gpt-6-solandgpt-6-lunamodel IDs, available now
The practical takeaway for developers: if your Codex sessions were getting expensive on GPT-5.6 models, switching the model ID to gpt-6-sol halves your token costs immediately. For high-volume automation, Luna at $0.10/$0.50 is among the cheapest capable models available from any frontier lab.
Sol vs Luna vs the Competition
Sol vs Claude Opus 5.5: launched 90 minutes apart, and they’re the head-to-head matchup of the season. Sol is $2/$10; Opus 5.5 is $4/$20 — but Anthropic says Opus 5.5 uses fewer tokens per task, so per-task cost is closer than the rate cards suggest. Our Claude Opus 5.5 breakdown has the full comparison.
Luna vs budget models: at $0.10/$0.50, Luna undercuts most competitors’ budget tiers while carrying GPT-6-generation training. For bulk content processing, data extraction, and high-volume agent subtasks, it’s the price leader among frontier labs.
If you’re picking AI tools for content work generally, our Muse AI account guide covers getting started with generous free tiers on the Claude side.
Choosing Between Sol and Luna: A Decision Guide
The choice is simpler than the spec sheets make it look:
Use Sol when: the task needs real reasoning — debugging a tricky bug, writing a complex function from a vague spec, analyzing a document, or any agentic workflow where a wrong step wastes more than the token savings. Sol is the “think hard, cost less than before” option.
Use Luna when: volume dominates — classifying thousands of items, extracting data from documents at scale, drafting routine content, running the low-risk subtasks inside a bigger agent workflow. Luna is the “good enough, incredibly cheap” option.
Use Astra when: it’s the hardest problem you have and cost is secondary. OpenAI is explicit that Astra remains the best model across the board — Sol and Luna are about bringing most of that capability to sustainable price points, not replacing the flagship.
A practical pattern for Codex users: set Sol as your default agent model and route the obvious subtasks (formatting, simple refactors, test scaffolding) to Luna. That two-tier setup captures most of the savings without touching quality on the hard parts.
Why This Launch Matters Beyond the Price Cut
Three structural shifts are packed into this release:
- The death of the middle tier. Terra’s disappearance confirms that AI model lineups are consolidating into three bands: flagship (Astra), workhorse (Sol), and budget (Luna). Expect competitors to simplify the same way.
- Caching as a pricing weapon. The 90% cached-input discount is doing quiet heavy lifting in the “50% cheaper” claim — workloads with high cache reuse (long Codex sessions, repeated document analysis) save far more than the headline rate cut suggests. Design your usage for cache hits.
- The price war is now scheduled. OpenAI and Anthropic launching flagship-adjacent models 90 minutes apart isn’t coincidence — it’s competition as calendar event. For buyers, that means timing your commitments: annual API contracts signed right before a rival launch are how you overpay.
Official source: OpenAI
Frequently Asked Questions
When did GPT-6 Sol and Luna launch?
September 22, 2026, available immediately in ChatGPT Work, Codex, and the OpenAI API.
How much cheaper are they?
Both are roughly 50% cheaper than their GPT-5.6 predecessors: Sol $2/$10 (was $4/$20), Luna $0.10/$0.50 (was $0.20/$1.20) per million tokens.
Can free ChatGPT users access them?
Free and Go users can access Luna through the ChatGPT desktop app. Neither model is in standard ChatGPT chat yet, and Sol requires a paid plan.
What’s the difference between Sol and Luna?
Sol is the capable mid-tier model for complex coding and agentic work; Luna is the fastest, cheapest option for high-volume tasks. Astra remains OpenAI’s best model overall.
Is GPT-6 Sol better than Claude Opus 5.5?
They launched the same day and target overlapping use cases. Sol has the lower rate card; Opus 5.5 claims better token efficiency. Real-world per-task cost depends on your workload — test both on your actual tasks.
What happened to GPT-6 Terra?
OpenAI didn’t release one. The mid-tier Terra from the 5.6 generation was squeezed out once Luna got cheaper, and no successor has been announced.