GPT-6 Sol and Luna vs GPT-5.6: better models at half the price
GPT-6 Sol and Luna improve coding, factuality and agentic work while cutting API prices roughly in half. What that changes for developers and Codex users.
OpenAI's latest model release is unusual for a simple reason: GPT-6 Sol and GPT-6 Luna are not only more capable than their GPT-5.6 counterparts, they are also dramatically cheaper.
That combination matters more to me than another benchmark win. The limiting factor with coding agents is increasingly not whether they can solve a task, but how much sustained iteration you can afford before usage limits or cost start shaping how you work.
I have already started using GPT-6 Sol at medium reasoning effort in Codex. My first impression is that the model can take on fairly large changes without immediately destroying the usage budget. It is still too early for me to call that a benchmark, and Codex quotas are not the same thing as API token billing, but the direction is noticeable.
The obvious change: roughly half the API price
OpenAI says improvements in inference and caching allow it to cut Sol and Luna pricing by 50% compared with GPT-5.6 promotional pricing.
The headline is simple: GPT-6 is not only more capable — the API price has moved sharply in the opposite direction.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| GPT-5.6 Sol | $4.00 | $20.00 |
| GPT-6 Sol | $2.00↓ 50% | $10.00↓ 50% |
| GPT-5.6 Luna | $0.20 | $1.20 |
| GPT-6 Luna | $0.10↓ 50% | $0.50↓ 58% |
For Sol, pricing drops by exactly 50% on both input and output.
For Luna, input also drops by 50%, while output falls from $1.20 to $0.50 —
roughly 58% lower.
This is the part that changes architecture decisions. A cheaper model is useful. A better model is useful. A model that improves while becoming substantially cheaper can move workloads that were previously hard to justify into the default path.
GPT-6 Sol vs GPT-5.6 Sol
OpenAI positions GPT-6 Sol as the model for complex coding and agentic workflows, with a 1.05 million token context window and up to 128,000 output tokens. It supports reasoning from none through max.
The benchmark gains are not limited to one category. OpenAI reports improvements in professional workflows, factual reliability, coding and computer use. On its internal factuality evaluation, GPT-6 Sol makes roughly half as many mistakes as GPT-5.6 Sol. On FrontierCode and DeepSWE, the new Sol also improves substantially on real-codebase software engineering tasks.
The interesting part is not that every score is suddenly perfect. It is that the model appears to push the capability curve forward while the token price moves in the opposite direction.
GPT-6 Luna may be the more interesting model
Sol is the obvious daily driver for difficult work, but Luna may be the more disruptive release.
GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, while keeping the same 1.05M context window and 128K maximum output as Sol. OpenAI describes it as its most efficient model for focused, high-volume tasks.
And the capability gap is getting surprisingly small in some workloads. OpenAI reports that Luna at high reasoning effort improves on GPT-5.6 Luna by 5.4 percentage points on AutomationBench while costing 58% less per task. On OSWorld, GPT-6 Luna at max effort can exceed GPT-5.6 Sol at medium effort at around one tenth of the cost.
That makes Luna interesting for things I would previously have sent to a larger model almost by reflex: classification with context, repetitive code changes, structured extraction, tool-heavy agents, tests, migration work and background automation.
Token price is not the same as cost per task
This is probably the most important caveat.
A model that costs 50% less per token does not automatically make every Codex session last twice as long. Reasoning effort changes token usage. Tool calls add work. Longer tasks may consume more context. Codex subscription quotas and rolling usage limits are also not a direct mirror of public API token pricing.
That matches what I am seeing in practice. I can run a substantial task with GPT-6 Sol Medium and see a relatively small hit to my available usage, but not necessarily exactly half what I would have expected from GPT-5.6. The model may also be doing more useful work inside that consumption.
So the metric I care about is increasingly cost per completed task, not cost per token.
Caching matters more for agents now
GPT-6 also improves prompt caching. OpenAI says cached input reads receive a 90% discount and that the new caching system is designed to preserve more reusable context even when reasoning effort or available tools change.
For traditional one-shot prompts that is nice. For agents it is much more important. Coding agents repeatedly send large parts of the same repository context, instructions, tool definitions and conversation history. Better cache hit rates can reduce the amount of context that needs fresh processing on every turn.
This is one of those improvements that is less visible in a benchmark screenshot but can matter enormously when an agent runs for hours.
So which one would I use?
My current split is fairly simple:
- GPT-6 Sol: the default for serious development, architecture, debugging and agentic work where mistakes or weak reasoning become expensive.
- GPT-6 Luna: high-volume tasks, repetitive implementation, extraction, classification, background agents and workloads where cost scales faster than complexity.
- GPT-6 Astra: the tasks where I genuinely want the strongest model available and the extra cost is justified.
The part I want to test over the next few weeks is how often Luna can replace Sol without becoming the bottleneck. If that boundary has moved significantly, the practical cost reduction could be much larger than the headline API price cut.
The bigger story is efficiency, not GPT-6
For a while, model releases mostly meant more capability in exchange for more compute. This release is more interesting because capability and cost are moving in opposite directions.
GPT-6 Sol is cheaper than GPT-5.6 Sol. GPT-6 Luna is cheaper than GPT-5.6 Luna. Both are more capable. Caching is better. And the models are being explicitly designed around sustained agentic workloads rather than isolated chat responses.
For developers using Codex every day, that is probably more meaningful than another leaderboard position.
I would rather have a model that lets me run an agent for longer, iterate more, and complete more real work for the same budget than one that wins a benchmark by two points and costs twice as much.
Sources: OpenAI — Introducing GPT-6 Sol and Luna, OpenAI API pricing, and OpenAI model catalog. Prices and availability checked September 24, 2026.