Which GPT-6 Model Should You Use?

Tasks within the same project can differ in difficulty

Suppose you are developing a new feature. First you organize user feedback, then edit the code, and finally investigate a failure that appears only under certain conditions. Feedback may have fixed fields and classification rules. Code changes require checking callers, while investigating the failure may require hypotheses and evidence.

You can evaluate models separately for these tasks. Work with simple criteria and easily verified results is a useful place to try a lower-cost model. When the material is more interconnected and failures take longer to investigate, a larger model budget can be worthwhile.

OpenAI currently positions Astra for its highest capability, 6.1 Sol for a balance of capability, speed and cost, and Luna for focused, high-volume tasks. Sol was the workhorse before this upgrade and serves as a migration baseline for existing workflows.

OpenAI API — Using GPT-6

Rates and context: what the numbers mean

All four official specifications list a 1,050,000-token context window and a 128,000-token maximum output. The context window describes the token capacity available for a request. Once the material is included, use a concrete task to check whether the model finds relevant evidence and retains the requirements for an edit.

The table lists Standard text-token rates for input, cache reads and output. The last column suggests a starting workload so you can connect each model to the work at hand.

OpenAI API — GPT-6 Astra · OpenAI API — GPT-6.1 Sol · OpenAI API — GPT-6 Sol · OpenAI API — GPT-6 Luna

Compare the four models on concrete work

6.1 Sol can be a starting point for new development work. To add a field to an endpoint, for example, specify the new behavior, compatibility requirements for existing callers, and tests for success and error responses. Inspect the patch and test results to check whether the requirements are met.

Consider Astra for difficult investigations: interacting failures, for example, or work combining experimental results, code and theoretical assumptions. In its Terminal-Bench Science 0.1 evaluation, OpenAI still recommends Astra for the hardest scientific research tasks.

Start a Luna trial with a small set of data whose answers you already know. Assign feedback to fixed labels, extract specified document fields or screen logs for possible errors. Check omissions and classification accuracy before increasing the volume.

Sol can continue handling work that already runs reliably. Programs using older reasoning settings or API configurations can stay operational while you verify the migration to 6.1 Sol.

OpenAI — Introducing GPT-6.1 Sol · OpenAI API — GPT-6 Sol · OpenAI API — GPT-6 Luna

Calculate each model's bill for the same token volume

Start with a budget example: each request has at most 272K input tokens, and the workload totals 10M uncached input plus 2M billed output tokens. The listed rates give about $200 for Astra, $40 for each Sol version and $2 for Luna.

Those amounts compare prices for equal token volumes. In actual work, also record how many tokens each model uses and how many attempts it takes. Divide spending across all attempts by the number of accepted tasks to get an average task cost, and track manual repair time separately.

Long-context rates apply to each request. Above 272K input tokens, the whole request's input and cache-related rates double, and output rates rise to 1.5 times the baseline. Two workloads with equal total usage can cost different amounts because their individual requests differ in size.

The calculator below lets you adjust total volume and the cached-input share. If long and short requests are mixed together, estimate the two groups separately and add the totals.

OpenAI API — Pricing

When to hand a task to another model

Even with one default model, keep track of tasks that may need escalation. If tests repeatedly fail, sources conflict, or the model changes a system boundary that has not been explained, pause and collect the current task state.

For example, after two controlled attempts fail to fix a bug, give Astra the original requirements, failing tests, edited files and confirmed facts to continue the investigation. Set the attempt policy for your own project. Explaining the work already done helps avoid repeating the investigation.

Start batch tasks with Luna, then review flagged results and pass them to 6.1 Sol. After extracting fields, for instance, separate records with missing required fields. For factual judgments, also check the result against the original text.

In DSH or another agent, configure the model and tools together. MCP and computer-use workflows depend on the application providing tools and permissions. Verify that a small task runs from request to result before expanding the scope.

OpenAI API — Using GPT-6

Takeaway: choose a starting point, then adjust from results

Begin choosing between GPT-6 Astra, 6.1 Sol, Sol and Luna with three workload categories: everyday development, complex research and batch processing. Try 6.1 Sol for new development, evaluate Astra for difficult research and Luna for extraction and classification. Migrate stable Sol workflows gradually.

Record results, time and cost, then adjust how the work is assigned. If you already use Sol, continue with the upgrade guide to check reasoning settings and tool APIs when switching to 6.1 Sol.

Frequently asked questions

OpenAI Academy — Model launch resources

Official specifications and selection resources

OpenAI API — Using GPT-6 · OpenAI API — GPT-6 Astra · OpenAI API — GPT-6.1 Sol · OpenAI API — GPT-6 Sol · OpenAI API — GPT-6 Luna · OpenAI — Introducing GPT-6.1 Sol · OpenAI API — Pricing · OpenAI Academy — Model launch resources

Related reading