The cheap model finishes its turn, then someone spends ten minutes repairing the answer. That hypothetical workflow captures the catch in any lower token price: paying less for a call and paying less for finished work are different achievements. OpenAI’s September 22 launch of GPT-6 Sol and Luna gives teams new prices and caching controls to test against that everyday distinction.
The official launch announcement placed both models below GPT-6 Astra. Its listed rates per million tokens were $2 for Sol input and $10 for output, and $0.10 for Luna input and $0.50 for output. OpenAI described the schedules as 50 percent below its predecessors’ promotional pricing. These are the launch’s listed rates, which give this dated first look a starting point for comparing workflows.
Both are large language models . Their different positions in the family let a team ask a practical question: which parts of the job need stronger reasoning, and which can be checked and completed with a smaller model?
Consider what an agent sends back on each turn. The task may change by a sentence while the repository brief, policies, tool definitions and earlier conversation remain much the same. Processing that repeated context can become a substantial part of the bill. Caching offers a way to reuse eligible earlier work, making the shape of the conversation part of the cost calculation.
OpenAI says GPT-6 cached-input reads receive a 90 percent discount. The launch also describes ways to change reasoning effort or tool availability while preserving earlier context. That gives a team more control within a supported model’s workflow. Switching to another model is a separate routing decision, and should not be treated as an automatic transfer of the same cache.
Give each model a job you can check
Sol is the general work model in the pair. OpenAI positions it for professional tasks, coding, and computer use where a team wants stronger capability than a small model but cannot justify Astra for every turn. Luna is the high-volume option, priced for classification, extraction, repetitive tool use, short transformations, and other workloads where the marginal cost of each call matters.
Imagine an agent first sorting a queue, then investigating the few records that need more work. Luna could handle a well-defined classification step whose output is easy to check. Sol could take a harder coding or analysis task, with Astra reserved for cases that justify it. Those are candidate routes to evaluate against real tasks, rather than guarantees that a particular model will succeed.
The bill follows every attempt. A low-priced call that produces a subtle error may trigger another call, a fallback and a human correction. Include all of those when comparing routes. Otherwise, a spreadsheet can show a saving while the team spends longer getting the work into an acceptable state.
Repeated context has its own cost
The OpenAI prompt-caching guidance explains how a matching prompt prefix can reuse earlier processing. It also describes monitoring, explicit breakpoints and supported ways to change reasoning or tools without rewriting that earlier prefix. The match still has to exist and remain available. A long conversation by itself does not guarantee a cache hit.
For a coding agent, a stable repository brief can sit before the changing task history. Later turns may reuse an eligible matching prefix. A supported configuration update can change reasoning effort while preserving earlier context, helping the agent spend more effort where the task becomes difficult. Keep that question separate from choosing a different model, then measure the actual cached-input usage for each route.
The launch reports GitHub’s observation of more than a 50 percent reduction in the share of prompt tokens needing fresh processing across billions of requests over several months. That is a useful report about its workload. A team with shorter prompts, less repetition or different traffic may see another result. Its own cache-hit rate belongs in the estimate.
Read the test conditions beside the score
The launch post reports gains across professional work, factuality, coding, computer use, and alignment. It says Sol at high or maximum reasoning levels can approach or exceed more expensive competitors on selected evaluations, while Luna improves materially over GPT-5.6 Luna at lower cost.
Read the benchmark conditions beside the scores. OpenAI used its research environment or API, and notes that production ChatGPT can behave differently because prompts and tools differ. Reasoning levels also vary across comparisons. The useful follow-up is to find a task that resembles your own and ask which settings produced the result and cost shown.
A sound first test should therefore use the team's own tasks. Measure completion quality, human correction time, retry rate, latency, total input and output tokens, cache-hit share, and final cost per accepted result. A model that is slightly less accurate but much faster may be preferable in a supervised queue. The same model may be unacceptable for an unsupervised workflow with costly errors.
The September launch had a staged rollout
The September 22 announcement said Sol and Luna were rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with Luna in the desktop app for Free and Go users. It said the models were not yet in Chat. That describes the launch rollout, with the announcement and product documentation providing routes to later availability information.
For developers, the announced model identifiers are gpt-6-sol and gpt-6-luna. Availability does not remove migration work. Existing prompts may behave differently, tool calls may shift, and cached-prefix design may need restructuring. Teams should keep a rollback path, pin model identifiers where supported, and compare results on representative production cases before changing a default route.
Keep the safety checks close to the task
OpenAI says Sol and Luna extend the alignment work introduced with Astra and show lower rates of misleading claims in selected internal tests. Its deployment-safety materials provide the detailed first-party evaluation context. These tests deliberately construct difficult situations and do not measure ordinary failure rates.
Independent Associated Press reporting on OpenAI's misalignment framework adds useful context: the company is trying to measure behaviors that can emerge when advanced systems face conflicting goals or opportunities to mislead. That reporting does not independently validate the new models, but it clarifies why coding deception and related tests now appear beside capability results.
Keep the application’s permissions, review boundaries, logs and recovery path tied to the consequences of a mistake. A task-level safety check and a model evaluation answer different questions. An agent working on a draft can tolerate a different level of supervision from one acting on a live account.
A small routing record makes the comparison usable
| Record for the same task | Why it belongs in the comparison |
|---|---|
| Every model call and its token subtotal | A fallback does not erase the failed first attempt |
| Observed cache-hit input | Repeated context is not automatically a discounted hit |
| Correction time and final acceptance | A lower invoice can conceal extra work for the person checking it |
| Model, settings and tool permissions | A result belongs to the route actually tested |
For the later model's current billing examples, use the separate GPT-6.1 Sol pricing guide. The launch comparison here remains tied to September 22.
What to test first
Start with one bounded job whose current cost and acceptance standard you understand. Run representative tasks on the current model, Luna and Sol. Record which results were accepted, how long they took, how much human intervention they needed and what all attempts cost. Then compare prompt-cache usage within each route. This produces a result your team can use, even if it disagrees with the headline benchmark.
A high-volume queue with clear checks may suit Luna. A difficult task with repeated context may favor Sol. Escalating to Astra can still be economical when it prevents costly failed attempts. Our analysis of GPT-6 Astra and critical-cyber capability follows the different controls and deployment profile around the top tier.
Give the trial one final column: total cost divided by accepted results. Put correction time beside it. If the token bill falls while those measures improve, the lower-priced route has earned its place. If the work keeps coming back for repair, you have found the part of the system to change next.
