OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, cutting API prices roughly in half compared to the GPT-5.6 models they replace. Sol is the capable mid-tier option for demanding work; Luna is the high-volume workhorse that now costs almost nothing when you account for prompt caching. If you are still calling gpt-5.6-sol in production, you are paying double for the same quality.
The Numbers #
Here is what actually changed in the pricing catalog:
| Model | Input ($/1M) | Output ($/1M) | Cached Input | Context |
|---|---|---|---|---|
| GPT-6 Astra | ~$10.00 | ~$50.00 | ~$1.00 | 1.05M |
| GPT-6 Sol | $2.00 | $10.00 | $0.20 | 1.05M |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | 1.05M |
| GPT-5.6 Sol (old) | ~$4.00 | ~$20.00 | — | 1.05M |
Both new models share the same 1,050,000-token context window and 128,000-token max output as Astra. You get the full context at a fraction of the price.
Sol vs. Luna: Which One to Use #
Sol targets demanding everyday work — coding agents, multi-step reasoning, complex document analysis. OpenAI says Sol reaches Astra-level reliability, making roughly half as many factual mistakes as GPT-5.6 Sol at around 20% of Astra’s per-task cost. On the DeepSWE v1.1 coding benchmark, Sol at max reasoning effort scores 68.8%, sitting within 1.1 points of Claude Fable 5 at approximately 80% lower cost per task.
Luna is for scale. Classification pipelines, structured data extraction, high-volume document processing — anything where you need millions of calls per day and the work is predictable. At $0.10 per million input tokens and $0.01 for cached reads, Luna’s economics are genuinely different from anything that has existed before in this capability class.
The practical split: use Sol where you previously would have reached for a premium model and quality matters. Use Luna everywhere else.
Prompt Caching Is the Quiet Win #
Both models apply a 90% discount to cached input reads automatically. Sol cached input drops to $0.20 per million tokens. Luna drops to $0.01 per million tokens.
To put that in concrete terms: if your agent reads 50,000 tokens of codebase context on every turn and you are hitting those cached reads on Luna, the per-turn input cost is $0.0005. That is not a rounding error — it is a rearchitecting moment. Complex RAG pipelines and context-chunking strategies built to manage token costs become engineering overhead you no longer need for Luna-class workloads.
There is also a fix that has been a persistent developer irritation: changing the reasoning.effort parameter between API calls used to invalidate the cached prefix, forcing a full re-read of the context. GPT-6 models now preserve the prompt cache across effort adjustments. You can tune effort per request without blowing up your caching strategy.
GitHub Copilot and Codex #
Both models are live in GitHub Copilot’s model picker through the Agent HQ system. Copilot subscribers can assign Sol or Luna to issues and pull requests alongside Claude and other agents in the roster. They are also available in Codex for asynchronous coding tasks. For enterprise teams already running Copilot, this is effectively immediate access to a capable mid-tier model without touching your API budget.
Migration: One Line of Code #
GPT-5.6 models are not deprecated. OpenAI confirmed they will continue working and you are not being forced to migrate. But there is no rational reason to stay.
The migration for most integrations is a single string change:
model="gpt-5.6-sol"
model="gpt-6-sol"
Same for Luna: gpt-5.6-luna becomes gpt-6-luna. No other API changes are required for standard Chat Completions usage. If you rely heavily on built-in tools and function calling, OpenAI recommends the Responses API — the Chat Completions path for function calling requires setting reasoning.effort to none. The full parameter range is: none, low, medium (default), high, xhigh, and max.
The Bottom Line #
GPT-5.6 is not officially dead, but it is practically obsolete. Sol and Luna deliver the same or better quality at half the price, with a larger tooling ecosystem and improved caching semantics. Luna’s $0.01 cached input rate means entire classes of cost-optimization work simply stop being worth the engineering effort. Full model specs are in the OpenAI API docs.
Update your model IDs. The only downside is the two minutes it takes to deploy.