How-to
Migrating from GPT-5.6 to GPT-6 Sol and Luna — Prices, Model IDs, Caching, and What Breaks
OpenAI's GPT-6 Sol and Luna halve the GPT-5.6 API prices, and GPT-5.6 Sol's promo price ends in November. Here is what changes when you switch model IDs, what to re-test, and how to cut over safely.
Prices and dates checked on 26 September 2026 against OpenAI's announcement and the other sources at the end.
On 22 September 2026 OpenAI released GPT-6 Sol and GPT-6 Luna, trained with similar methods to GPT-6 Astra. For API customers, the news is in one table: both cost half their GPT-5.6 predecessors. GPT-5.6 Sol's current price is itself a promotion that ends on 21 November 2026. So most teams on GPT-5.6 Sol or Luna should plan a migration now rather than after the cliff.
This is the checklist. It covers what changes, what to re-test, and how to cut over without breaking production.
What changes: IDs and prices
| Old ID | New ID | Old price (in / out per 1M) | New price | Change |
|---|---|---|---|---|
gpt-5.6-sol | gpt-6-sol | $4 / $20 (promo until 21 Nov; list $5 / $30) | $2 / $10 | −50% vs promo, −60% input and −67% output vs list |
gpt-5.6-luna | gpt-6-luna | $0.20 / $1.20 | $0.10 / $0.50 | −50% input, −58% output |
gpt-5.6-terra | — (no GPT-6 Terra announced) | $2 / $12 | — | Stay, or test GPT-6 Sol, which now lists at the same input price and a lower output price |
What the savings look like at a steady monthly volume of 100M input and 20M output tokens:
| Workload | GPT-5.6 | GPT-6 | After 21 Nov if you stay on GPT-5.6 Sol |
|---|---|---|---|
| Sol | $800 | $400 | $1,100 (at the $5/$30 list) |
| Luna | $44 | $20 | — |
Those figures ignore caching, which makes the new models cheaper still.
What changes: caching
OpenAI shipped caching changes with GPT-6 that matter for agents and long conversations:
- Higher cache hit rates by default, with a 90% discount on cached input-token reads.
- Changing reasoning effort or tool availability no longer breaks the cache. You can raise effort for a hard step, or switch tools on and off mid-conversation, and earlier context stays reusable.
- Explicit breakpoints let you choose where a cached prefix ends.
- A Prompt Caching Dashboard and a diagnostics tool show what is cached and explain missed opportunities.
Migration action: if your application reorders the system prompt, injects timestamps at the top, or rebuilds tool lists per request, fix that first. Stable prefixes are where the 90% discount comes from. Check the dashboard in week one; a low hit rate usually means a prompt-assembly bug, not a model problem.
What can break
Price is the easy part. These are the things to re-test:
- Output shape. OpenAI says GPT-6 Sol and Luna answer with "more clarity, less jargon, fewer odd turns of phrase… and slightly shorter answers overall." That is good for users and risky for code that parses responses. Check length-based heuristics, regex extractors, and any eval that rewards verbosity.
- Your quality bar. OpenAI reports that GPT-6 Sol makes about half as many factual mistakes as GPT-5.6 Sol on its internal evaluation. It also says GPT-6 Luna at higher effort matches GPT-5.6 Sol at about a hundredth of the cost. These are vendor claims. Your acceptance tests decide whether Luna can take over work that Sol used to do.
- Effort settings. If you tuned a reasoning-effort level on GPT-5.6, re-tune it. The same label on a new model does not guarantee the same latency or token use.
- Long-context pricing and limits. The limits carry over: GPT-6 Sol and Luna keep the 1.05M-token context and 128K max output of their GPT-5.6 predecessors. The long-context step also stays at 272K input tokens. Above it, the whole request bills $4/$15 on Sol ($0.40 cached) and $0.20/$0.75 on Luna. Cache writes cost 1.25× input: $2.50 per million on Sol and $0.125 on Luna. Re-run your cost model for any workload that regularly crosses 272K.
- Pinned aliases and dashboards. Search config, infrastructure-as-code, and billing dashboards for
gpt-5.6-. Mixed IDs make a cost spike hard to read. - ChatGPT is not the API. In ChatGPT, GPT-6 Sol and Luna are in Work and Codex for Plus, Pro, Business, Enterprise and Edu. They are not yet in Chat. On Enterprise plans an admin must enable them. Do not use the Chat window to judge the API model.
A two-week migration plan
Days 1–2: inventory. List every call site, its current model ID, monthly tokens, and owner. Mark which ones have automated validators and which rely on humans spotting problems.
Days 3–5: shadow traffic. Send a copy of real requests to the GPT-6 ID behind the same adapter, without serving the result. Compare validator pass rates, output length, latency, and cost per request. Freeze 20–30 hard prompts per workload as a regression set.
Days 6–8: caching check. Turn on caching diagnostics for the GPT-6 route. Fix prompt-assembly issues until the hit rate matches what the workload's structure should allow.
Days 9–12: staged cutover. Move 10%, then 50%, then 100% of traffic per workload, holding each step for at least a day. Keep the GPT-5.6 route warm as a one-line rollback.
Days 13–14: tier review. Now that Sol costs no more per token than GPT-5.6 Terra, check whether Terra workloads should move to Sol. Check whether some Sol workloads pass on GPT-6 Luna. The OpenAI tier guide covers the routing logic.
Finish before 21 November, when GPT-5.6 Sol's promotional price is scheduled to end.
When not to migrate yet
- Regulated outputs with sign-off. If a workflow's outputs were validated by compliance or legal on GPT-5.6, budget for that review again. The price saving does not skip it.
- Workloads that run in ChatGPT, not the API. Migrating there depends on OpenAI's rollout, not your code.
- A frontier-model decision is pending. If you are also weighing Anthropic, test both at once. GPT-6 Sol vs Claude Opus 5.5 has the cost-per-task framing.
For a shortlist scored against your constraints, use the AI Model Picker. If the model sits inside a coding tool rather than your own code, the AI Coding Assistant Finder covers the harness side.
Sources (accessed 2026-09-26 unless noted)
- OpenAI — Introducing GPT-6 Sol and Luna (prices, 50% reduction,
gpt-6-sol/gpt-6-luna, caching changes, communication style, availability): https://openai.com/index/introducing-gpt-6-sol-and-luna/ - OpenAI — GPT-5.6 (Sol $4/$20 promo through 2026-11-21 from $5/$30; Terra $2/$12; Luna $0.20/$1.20), accessed 2026-09-05: https://openai.com/index/gpt-5-6/
- Notebookcheck — "ChatGPT Free gets GPT-6 Luna, but only in the desktop app" (24 Sep 2026; Enterprise admin enablement): https://www.notebookcheck.net/ChatGPT-Free-gets-GPT-6-Luna-but-only-in-the-desktop-app.1407282.0.html
- OpenAI — model docs for gpt-6-sol and gpt-6-luna (1.05M context, 128K max output): https://developers.openai.com/api/docs/models/gpt-6-sol
- OpenAI — API pricing (>272K input: Sol $4/$15, Luna $0.20/$0.75; cache writes 1.25× input): https://developers.openai.com/api/docs/pricing
- TechCrunch — OpenAI launches GPT-6 Sol and Luna (22 Sep 2026): https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/
Editorial note
AI Choice Engine publishes editorial guides to help readers understand fit, trade-offs, and next steps before choosing a tool or provider.
Newsletter
Get the buyer checklist that goes with this guide
Subscribe to download the matching checklist. Automated email delivery is still rolling out — this is not an inbox confirmation.
Updates only — checklists stay free from the resource library, with or without joining. Automated email delivery is still rolling out.