GPT-6.2 Sol and Astra API: Free Trials and Discounted Access
Explore GPT-6.2 Sol and Astra on Felo API: compare capabilities, try both models, and check discounted provider pricing.
OpenAI's GPT-6.2 Sol and GPT-6.2 Astra are now available on Felo API Platform. Both model pages currently display a limited-time free trial, and selected Astra provider routes offer discounted token rates for ongoing use.
The two models address different needs. Sol is built for everyday coding, writing, and agent workflows where efficiency matters. Astra is OpenAI's frontier option for harder projects that demand sustained reasoning, complex software changes, or polished professional deliverables. Here's what each model offers, how to choose, and how to call them through Felo.
GPT-6.2 Sol vs. Astra at a glance
| GPT-6.2 Sol | GPT-6.2 Astra | |
|---|---|---|
| Best starting point | Everyday coding, automation, writing, and agents | Complex engineering, research, analysis, and professional documents |
| Input and output | Text and image input; text output | Text and image input; text output |
| Context window | 1.05 million tokens | 1.05 million tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| Reasoning effort on Felo's model page | Low, medium, high, xhigh, max | Low, medium, high, xhigh, max |
| OpenAI standard input / output price | $2 / $10 per million tokens | $8 / $40 per million tokens |
| Felo trial status at publication | Limited-time free trial displayed | Limited-time free trial displayed |
Context is the combined space available for a request and its response, not a promise that every task will benefit from filling it. Output limits, supported tools, and settings can depend on the route and protocol. See the current Sol and Astra pages for the latest details.
What makes GPT-6.2 Sol useful?
Sol is the model to consider when an application needs capable assistance throughout the day. OpenAI describes it as more capable and efficient than GPT-6.1 Sol, with improvements in multistep agent work and natural writing. That combination is useful when a single request involves reading context, making a plan, using tools, and returning a result in a prescribed format.
Coding and everyday development
Sol can help inspect existing code, implement a scoped change, write or adjust tests, and explain tradeoffs. Give it relevant files, the desired behavior, and a way to verify the result. For a maintenance task, that might mean providing the failing test, the surrounding module, and the command used by your project.
OpenAI reports that on Cognition's FrontierCode 2.0, which evaluates changes to real codebases against maintainers' standards, Sol came within 0.5 points of Astra at roughly one-fifth the cost per task, with both tested at maximum reasoning. This is one benchmark under a particular configuration; your own repository, tools, and review standards may produce different results. OpenAI's launch announcement explains the evaluation.
Reports, plans, and repeatable workflows
Sol is also a practical starting point for turning meeting notes into a project plan, summarizing a set of documents, drafting a customer response, or categorizing incoming requests. Ask for a specific audience, output format, and treatment of missing information. For example, request a table of decisions, owners, deadlines, and unresolved questions rather than a generic meeting summary.
For agent applications, the model's tool capabilities include function calling, web search, file search, and computer use as described on its model page. The actual tools available to an application depend on the integration and selected route: a model's ability to use a tool does not mean your request automatically has that tool attached.
Vision and long context
Sol accepts images as well as text, so an application can ask it to interpret a screenshot, read a chart, or use visual context while discussing a coding task. Its 1.05-million-token context window also makes it possible to work with substantial amounts of material in one request. Large context still needs careful organization: identify the important sources, label them clearly, and ask the model to cite or identify the material behind its conclusions.
What makes GPT-6.2 Astra different?
Astra is designed for tasks where the quality of the finished work and the ability to reason across many steps justify a higher per-token price. OpenAI positions it for complex professional work, including codebases, research, data analysis, and the creation of slides, spreadsheets, and documents.
Complex software engineering and frontends
Astra can be useful when a change spans multiple parts of a codebase or requires architectural judgment. A good request describes the existing system, constraints, acceptance criteria, and tests. It can then help investigate dependencies, propose an implementation, make changes within the tools provided, and evaluate the result.
OpenAI says Astra improved at implementing and testing complicated changes in existing repositories and at building more polished frontends, including websites, games, and 3D designs. Treat these as capabilities to evaluate in your environment: run tests, review generated code, and check the interface yourself before shipping.
Spreadsheets, documents, and slides
Astra's document work is particularly relevant when the deliverable has to follow an established format. Given source materials and a template, it can help construct a financial model, prepare an investment memo, or draft a presentation that follows the supplied structure and style. OpenAI reports improvements in template matching, including layout, colors, typography, and writing style.
A useful prompt supplies the source files, intended reader, template, assumptions, and any checks the output must pass. For a workbook, ask for formulas rather than unexplained hard-coded numbers and specify reconciliation checks. Review formulas and source data independently before relying on a financial or business decision.
Research and long-horizon reasoning
Astra is suited to workflows that require several connected steps: interpreting data, developing an approach, testing possibilities, and revising the result. OpenAI reports a nine-percentage-point improvement over GPT-6 Astra on Terminal-Bench-Science 0.1, an evaluation of scientific work in a terminal, at less than one-third of the cost per task in its test. These are provider-reported benchmark results, not guarantees for a particular experiment. Human review remains essential, particularly where results affect research or operations.
Both models have stronger cybersecurity capabilities, and OpenAI says it has tightened safeguards for higher-risk cybersecurity requests. That can affect some legitimate security testing workflows. See the GPT-6.2 system card for details.
Free trials and API pricing on Felo
As of November 18, 2026, both Felo model pages show a limited-time free trial. Check the model page and your account for current availability, eligibility, and any limits before testing. Trial availability can end; after it ends, calls are billed at the applicable route's published rate. Felo's pricing page explains how model pricing works.
The provider you choose matters. At publication, the pages list the following base rates per million tokens:
| Model and route | Input | Output | Cached input |
|---|---|---|---|
| Sol: OpenAI standard | $2 | $10 | $0.10 |
| Astra: OpenAI standard | $8 | $40 | $0.40 |
| Astra: openai-mix, listed at 80% off | $1.60 | $8 | $0.08 |
| Astra: openai-official-cheap, listed at 40% off | $4.80 | $24 | $0.24 |
These are published base rates, not a promise that every provider or request receives a discount. Sol's listed standard route was at the same base token rates as OpenAI's standard API at publication. The Astra discounts apply to the named routes, relative to the prices displayed on the model page. Availability, conditions, and rates can change.
Long inputs and additional services can change the bill. For example, the pages list higher token rates when input context exceeds 272,000 tokens, and separately list charges for cache writes and web search. Inspect the provider's pricing details before production use.
For a simple illustration, one million uncached input tokens plus one million output tokens at Astra's listed standard base rates would cost $48; the same token quantities at the listed openai-mix base rates would cost $9.60, before other charges and assuming the base pricing conditions apply. Real requests usually use different input and output amounts, and trial usage or route conditions can change the result.
How to use GPT-6.2 through Felo API
- Open the Sol model page or Astra model page, and check the current trial badge and provider pricing.
- Create a Felo API key from your account. Check your balance and add funds if needed for usage outside an available trial.
- Use the Playground to compare the models with a representative prompt, or send a request to the Responses-compatible endpoint.
- Measure output quality, latency, token usage, and total cost on your own tasks before choosing a production route.
Set your key as an environment variable, then call Sol:
export FELO_API_KEY="YOUR_API_KEY"
curl https://openapi.felo.ai/api/v1/responses \
-H "Authorization: Bearer $FELO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6.2-sol",
"input": "Turn these project notes into a concise plan with milestones and open questions."
}'
To try Astra, replace the model value with gpt-6.2-astra and use the same endpoint. Keep your API key on the server or in a secure environment; do not embed it in browser code or a public repository. For streaming and compatible SDK examples, see the quick-start section on each model page. Provider selection, reasoning controls, and tool support depend on the route and request configuration; check the current documentation before assuming a particular setting is available.
Which model should you choose?
Start with Sol when you are building a high-volume assistant, automating routine development, or drafting and organizing everyday work. It offers a lower standard token price and is a useful baseline for measuring quality and cost.
Test Astra when the problem involves a complicated codebase, a difficult research workflow, or a deliverable that must follow detailed professional conventions. It may earn its higher standard rate if it reduces retries or review work, but measure the complete task rather than comparing token prices alone.
Some teams may use both: send routine work to Sol and route selected difficult cases to Astra. Evaluate that approach with a representative test set, including failure cases, and compare completed-task quality, latency, human review, and full billing.
Frequently asked questions
Can I try GPT-6.2 Sol and Astra for free?
Both model pages displayed a limited-time free-trial badge at publication. Availability and limits can change, so check each page and your account before sending requests. A trial is not a recurring free allowance or a guarantee that all routes are free.
Is Astra always more expensive than Sol?
OpenAI's standard base token rates for Astra are higher than Sol's. Felo lists discounted Astra routes, so compare the actual provider rates and conditions for your workload. Total cost also depends on tokens used, cache behavior, input length, and any tool charges.
Can both models analyze images?
Yes. Both model pages list text and image input and text output. Check your selected API route and request format for implementation details.
What should I do if an answer looks incomplete?
Ask the model to identify missing information and verify important claims against the source material. OpenAI notes that both models were more likely in two tests to omit certain obstacles that blocked their work. For agent workflows, explicitly require reporting of missing files, failed tools, and incomplete steps.
Ready to compare them? Explore the GPT-6.2 Sol API and GPT-6.2 Astra API on Felo, check the current trial and provider rates, and run your own test.