QWE AI Academy End state: you pick Sol or Luna for the job, open Work/Codex or the API, set effort once, and ship a real task today – without burning tokens on the wrong tier.
GPT-6 Sol and Luna shipped September 22, 2026. Half the old API rates. Astra-class training methods. Threads already yelling about the model picker. This is the day-1 how-to, not a scoreboard recap.
Launch snapshot (90 seconds)
Two cheaper GPT-6 siblings under Astra. Sol = complex coding and agent loops. Luna = high-volume extract/summarize/classify/quick Q&A. That split comes straight from the OpenAI announcement. Both keep a 1.05M context window and 128K max output (as of launch docs, Sept 2026).
| Model | API ID | Input / 1M | Output / 1M | Best fit |
|---|---|---|---|---|
| GPT-6 Sol | gpt-6-sol |
$2 | $10 | Multi-step coding, agents, judgment calls |
| GPT-6 Luna | gpt-6-luna |
$0.10 | $0.50 | Volume: extract, summarize, classify, route |
| GPT-6 Astra | (flagship) | Higher tier – check live pricing | Hardest end-to-end work only | |
Short-context standard rates above, as of the Sept 22, 2026 launch tables. Cached reads land near 10% of input (Sol ~$0.20, Luna ~$0.01 per 1M). Batch/Flex: half. Fast/priority: about 2×. Cross 272K input and the math flips – details in Edge cases. Versus GPT-5.6 promo rates ($4/$20 Sol, $0.20/$1.20 Luna), these list prices are 50% lower per OpenAI’s pricing docs.
Funny gut check: if your agent resends the same system block every turn and cache still misses, the “cheap model” wasn’t the leak. The prefix was.
Method A vs Method B: Sol or Luna?
Don’t default to Sol because the name sounds tougher. Route by failure cost.
- Method A – Luna first: Narrow, repeatable, easy to check (field extraction, ticket tags, thread summaries, format rewrites). Start at
noneorloweffort. Escalate only after a spot-check fails. - Method B – Sol when wrong is expensive: Multi-file refactors, tool-using agents, migrations, research that must hold up. Start at default
medium. Touchhigh/xhigh/maxonly after medium misses.
There’s no GPT-6 Terra. Lived on 5.6 Terra? Split it: Luna for glue, Sol for the hard middle, Astra only when Sol clearly fails. That’s the ladder.
One honest open question: for borderline coding tasks, some early Codex runs favor Luna at high effort on cost-per-pass – your repo may disagree. Measure one real job before you rewrite defaults.
How to use GPT-6 Sol and Luna today
Most beginners win on ChatGPT Work or Codex (paid plan), then API if you automate. These models are not in regular Chat at launch – OpenAI’s availability section says Work and Codex only (Free/Go: Luna in the desktop app).
ChatGPT Work / Codex (Plus, Pro, Business, Enterprise, Edu)
- Open ChatGPT Work or Codex (web, desktop, CLI, or IDE) – not the ordinary Chat composer.
- Under the composer, open the model + reasoning control.
- Pick GPT-6 Sol or GPT-6 Luna when it shows. Gradual rollout; missing? Retry later the same day.
- First real task: medium effort. Raise only if the answer is shallow.
- Enterprise/Edu: admin must enable the models in workspace settings.
Free and Go → desktop app → Luna. Plus/Pro/Business also got a banked usage reset at launch (OpenAI Developer Community note).
API
IDs: gpt-6-sol, gpt-6-luna. Need tools plus reasoning? Responses API – not Chat Completions with thinking on. Minimal Luna call:
curl https://api.openai.com/v1/responses
-H "Authorization: Bearer $OPENAI_API_KEY"
-H "Content-Type: application/json"
-d '{
"model": "gpt-6-luna",
"reasoning": { "effort": "low" },
"input": "Extract company, amount, and due date as JSON from this invoice text: ..."
}'Swap to gpt-6-sol and medium/high for multi-step coding. Parameter gotchas (function calling only with reasoning_effort: none on Chat Completions; drop temperature/top_p when reasoning isn’t none) sit in the latest-model guide and the Luna model page.
Pro tip: Stable instructions and big reference docs go at the top so prompt cache can hit. Cached input reads are priced around 90% off list – agents that reuse the same prefix feel it first.
Edge cases that waste a day
Wrong surface. Chat picker still on 5.6-era names? You’re in Chat. Sol/Luna live in Work and Codex. Launch-day “Plus doesn’t have it” threads were mostly surface mistakes.
Tools + reasoning on Chat Completions. Function calling on gpt-6-sol / gpt-6-luna via Chat Completions works only with reasoning_effort: "none". Tools and real thinking → Responses API. Send temperature / top_p while reasoning is on and the API rejects the call.
The 272K cliff. Input past 272K tokens and the whole request bills at long-context rates (2× input/cache, 1.5× output) – not just the overflow. A 273K prompt can nearly double cost for a sliver more context. Compact before you cross. Numbers as of the Luna/pricing pages above; recheck if you bill weeks later.
Cutoff quirk. Luna’s knowledge cutoff (May 18, 2026) is later than Sol’s (April 20, 2026). “What changed last month?” at volume: Luna isn’t only cheaper – it’s newer. Fine-tuning isn’t supported on either at launch; 5.6 fine-tunes don’t carry over.
FAQ
Can I use GPT-6 Sol in normal ChatGPT Chat on Plus?
No – not at launch. Use Work, Codex, or the API.
Should I replace every GPT-5.6 Sol call with gpt-6-sol overnight?
On one production job, same effort, side-by-side. If quality holds, flip the model ID – token prices are lower, and OpenAI’s launch note says Sol makes about half as many mistakes as GPT-5.6 Sol on their factuality eval. Effort still defaults to medium. Tool paths may need a Responses migration. Keep a fallback flag for a week.
Is Luna “good enough” for coding?
Scoped edits and codegen glue? Often yes, especially at higher effort – unit cost is tiny. Ambiguous architecture, security-sensitive paths, or long agents that must stay coherent? Sol is the intended tier. Luna max is a budget experiment, not a blanket Sol swap. Escalate when tests fail or it invents APIs.
Next action: Open Work or Codex now. One real task on Luna low, same task on Sol medium. Compare quality and (if API) billed tokens. Lock your router to that result – not the launch noise.