GPT-6 Sol and Luna: Agent Cost, Coding and Caching

Quick answer

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. Both are available through the API, and OpenAI says they are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access Luna in the desktop app. OpenAI says the models are not yet available in Chat.

The practical story is bigger than a benchmark refresh. API prices are lower, coding performance improved, and GPT-6 prompt caching now gives persistent agents more ways to reuse expensive context. For teams running long coding or knowledge-work agents, model choice is increasingly an economics decision about cost per reviewed outcome, not just intelligence per prompt.

For related context, see the Codex tool profile, ChatGPT tool profile, and AI agent best practices.

What did OpenAI release?

GPT-6 Sol and Luna extend the GPT-6 family below GPT-6 Astra. OpenAI positions Astra as the highest-capability option, Sol as a strong model for difficult professional and coding work at lower cost, and Luna as the more cost-sensitive option for high-volume work.

For standard API prompts up to 272K input tokens, OpenAI's API changelog lists GPT-6 Sol at $2 per million input tokens, $0.20 per million cached input tokens and $10 per million output tokens. GPT-6 Luna is $0.10 input, $0.01 cached input and $0.50 output per million tokens.

OpenAI describes this as a 50% API price reduction versus GPT-5.6 promotional pricing for the corresponding Sol and Luna tiers. That comparison is OpenAI's own pricing baseline, not a claim that every real workload will cost exactly half as much.

Why does this matter for coding agents?

Long-running coding agents consume far more than one prompt and one answer. They repeatedly carry repository instructions, tool definitions, code context, test output and prior decisions across requests.

That means a model can be more useful even when the headline benchmark improvement is modest if it reduces the cost of keeping a productive agent running for hours.

OpenAI reports that GPT-6 Sol improves substantially over GPT-5.6 Sol on FrontierCode, an evaluation that grades real-codebase changes for qualities including correctness, test quality, scope discipline, style and adherence to repository standards. On DeepSWE 1.1, OpenAI reports GPT-6 Sol at max effort scoring 68.8% and GPT-6 Luna at max effort scoring 66.6%.

Those are benchmark results, not guarantees for your repository. Teams should validate on their own task mix before changing defaults.

Is GPT-6 Luna now enough for serious coding work?

Potentially for many bounded tasks, but the right answer depends on the cost of failure.

Luna's lower API price makes it attractive for repetitive implementation, test generation, repository exploration and other high-volume work where failures are easy to detect and retry. Sol may be a better default when a task is more ambiguous, crosses architectural boundaries or requires stronger reasoning before editing code.

A useful routing policy is not simply cheap model for easy work, expensive model for hard work. It should also account for verification cost.

If a cheaper run produces more failed tests, repair loops or reviewer effort, its lower token bill can become a false economy.

TowCue take: optimize cost per reviewed outcome

The most useful metric is not cost per million tokens.

For an agent workflow, measure the full path:

task → model → execution → tests → review → repair → accepted outcome

Then include:

model and cached-context cost;
elapsed time to a review-ready result;
failed tests and repair turns;
human review time;
rework after review;
defects that escape after merge.

A model that costs more per request can still be cheaper per accepted change. A cheaper model can win when the task is well specified and verification is strong.

Why is prompt caching part of the launch story?

OpenAI separately announced improved prompt caching for GPT-6 on September 22. Eligible shared prefixes reused within a 30-minute window can receive cached-input discounts, and OpenAI says cached input reads can receive discounts of up to 90%.

This matters because persistent agents often repeat large stable prefixes: system instructions, repository policy, tool schemas and reference material.

GPT-6 also adds controls intended to preserve cache reuse while an agent changes reasoning effort or which tools are callable. Developers can use explicit cache breakpoints, a Prompt Caching Dashboard and diagnostics for cache misses.

Caching does not make every token cheap. It rewards workflows that separate stable context from frequently changing context.

What does GitHub Copilot's caching result show?

In OpenAI's announcement, GitHub says prompt-caching improvements reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models over several months, relative to its previous baseline.

That is a GitHub-reported production result quoted by OpenAI, not a universal expected saving. Your hit rate will depend on prompt structure, tool schemas, request timing and how often shared prefixes change.

The lesson is still useful: context architecture can become a material infrastructure cost once an agent operates at scale.

How should teams route work between Astra, Sol and Luna?

Start from consequence and uncertainty rather than model prestige.

Use a lower-cost tier for tasks that are narrow, reversible and covered by deterministic checks. Increase capability when requirements are ambiguous, changes span several systems, failures are expensive or the reviewer cannot cheaply reconstruct the reasoning.

Keep a fallback path. If Luna repeatedly fails a bounded task, escalate to Sol. If Sol still cannot resolve a high-stakes or unusually complex problem, a stronger model such as Astra may justify its higher cost.

The routing decision should be observable. Record which model handled the task and compare acceptance, retries, latency and reviewer effort over time.

What should developers test before switching defaults?

Build a small evaluation set from work your team actually does rather than relying only on public benchmarks.

Include a mechanical refactor, a bug with a known root cause, a feature that touches several files, a test-writing task and one ambiguous issue that normally requires architectural judgment.

Run the same tasks with the candidate models under the same repository instructions and validation commands. Compare not only whether the code works, but how much repair and review it needs.

If your agent carries large repeated context, inspect cache hit rates separately. A model migration and a cache-architecture change should not be treated as the same experiment.

Who benefits most from this release?

Teams with sustained Codex usage, API-based coding agents or other persistent workflows have the clearest reason to evaluate Sol and Luna. Lower unit prices and better context reuse can compound when agents make many requests per task.

Occasional ChatGPT users should be more cautious about reading API economics as a direct personal benefit. Availability also differs by product: OpenAI says Sol and Luna are available in ChatGPT Work and Codex for eligible paid and organizational plans, while the models are not yet available in Chat itself.

TowCue take: the model is becoming one layer of the agent stack

The GPT-6 Sol and Luna launch reinforces a broader shift in AI workflow design.

A production agent is increasingly:

task policy → model routing → reusable context → tools → execution → verification → human review

Model quality still matters. But so do routing, cache design, tool stability and the cost of proving the result is good enough to accept.

As model prices fall, teams may be tempted to run agents everywhere. The better question is whether each additional autonomous run produces an outcome worth its full verification cost.

The winning workflow will not necessarily use the strongest model on every step. It will use enough intelligence for the task, preserve reusable context where it helps, and spend human attention where mistakes are expensive.

GPT-6 Sol and Luna launch: why agent economics now matter as much as model quality