Updated
What is the GPT 6.2 Sol API?
The GPT 6.2 Sol API is live on UnificAlly as openai/gpt-6.2-sol, from $0.30/1M tokens of input and $1.50/1M tokens of output. GPT 6.2 Sol is OpenAI's everyday model in the GPT 6.2 family, released on November 17, 2026 alongside GPT 6.2 Astra. OpenAI built it for day-to-day work, from coding to running agents at scale. It keeps GPT 6.1 Sol's per-token price and gets more done per task: it matches GPT 6 Astra on Terminal-Bench-Science 0.1 at under a tenth of the cost per task, while writing fewer output tokens than GPT 6.1 Sol. It takes text and image input, returns text, and works across a 1,050,000-token context window with a 128,000-token max output and knowledge up to June 1, 2026. Reasoning effort runs from low up to max.
GPT 6.2 Sol API at a glance
| Model ID | openai/gpt-6.2-sol |
| Endpoint | POST https://direct-api.unifically.com/v1/chat/completions |
| Other formats | POST /v1/responses and POST /v1/messages |
| Request types | Chat, streaming, tool calls |
| Input | Text and images |
| Output | Text |
| Context window | About 1 million tokens |
| Max output | 128,000 tokens |
| Reasoning effort | low to max, set with reasoning_effort |
| Input rate | $0.30/1M tokens |
| Output rate | $1.50/1M tokens |
| Cache read rate | $0.015/1M tokens |
| Cache creation rate | $0.375/1M tokens |
The GPT 6.2 Sol API bills per token, with no subscription. The rates above apply to prompts up to 272,000 tokens, and longer prompts bill at the long-context rate. The OpenAI API pricing guide covers the rates for every GPT model.
How to use the GPT 6.2 Sol API
- Create an API key on the API keys page. The same key and balance work for every model on UnificAlly.
- Send a POST request to
/v1/chat/completionsonhttps://direct-api.unifically.comwith the model ID and amessagesarray. This is the OpenAI-compatible endpoint shared by every language model on UnificAlly. - Read the reply from
choices[0].message.contentin the same response. There is no task to check.
curl -X POST https://direct-api.unifically.com/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer YOUR_API_KEY" -d '{
"model": "openai/gpt-6.2-sol",
"reasoning_effort": "medium",
"messages": [
{ "role": "user", "content": "Find why this test fails, fix it, and explain the change in two sentences." }
]
}'
The response comes back synchronously with the completion and the token usage:
{
"id": "8c41d7e2-5a9b-4f3c-b2e6-1d7f0a9c3e58",
"object": "chat.completion",
"created": 1710000000,
"model": "openai/gpt-6.2-sol",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "The test fails because the date parser ignores the time zone..."
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 26,
"completion_tokens": 71,
"total_tokens": 97,
"prompt_tokens_details": {
"cached_tokens": 0
},
"completion_tokens_details": {
"reasoning_tokens": 0
}
}
}
Set "stream": true to receive tokens as they generate, and raise reasoning_effort to high or max for hard multi-step tasks. Send requests to direct-api.unifically.com so long replies are not cut short by a proxy timeout.
GPT 6.2 Sol API parameters
| Parameter | Type | Default | What it does |
|---|---|---|---|
model | string | required | Model identifier, openai/gpt-6.2-sol |
messages | array | required | Non-empty array of message objects with role and content (a string or an array of content parts). Roles: system, developer, user, assistant, tool, function |
stream | boolean | false | If true, the response is streamed using server-sent events |
stream_options | object | none | Set { "include_usage": true } to get a final chunk with token usage. Only used when stream is true |
temperature | number | none | Sampling temperature between 0 and 2. Higher values produce more random output |
max_tokens | integer | none | Maximum number of tokens to generate in the completion |
max_completion_tokens | integer | none | Accepted as an alias for max_tokens |
reasoning_effort | string | none | How much the model reasons before answering, from low to max |
top_p | number | none | Nucleus sampling parameter. An alternative to temperature |
stop | string or array | none | Up to four sequences where the model stops generating further tokens |
tools | array | none | List of tools the model may call. Each tool requires a type and a function definition |
tool_choice | string or object | none | Controls which tool is called, if any: none, auto, required, or a specific tool |
response_format | object | none | Set { "type": "json_object" } to enable JSON mode |
The GPT 6.2 Sol API documentation has the model page, and the Chat Completions reference linked from it covers streaming, tool calls and error codes.
What GPT 6.2 Sol can do
GPT 6 Astra-level science work at a tenth of the cost
Terminal-Bench-Science 0.1 has agents analyze data, run simulations, and prove theorems from a terminal. GPT 6.2 Sol solves 68.0% at max effort for $2.21 a task, level with GPT 6 Astra at 68.1% for $23.80 and ahead of Claude Opus 5.5 at 63.3%. That is 11 points above GPT 6.1 Sol, which needed $5.47 a task.
Coding within half a point of GPT 6.2 Astra
Cognition's FrontierCode 2.0 checks whether changes to real codebases meet the maintainers' bar for merging. At max effort GPT 6.2 Sol scores 43.8% for $1.29 a task, 0.4 points behind GPT 6.2 Astra at about a fifth of its cost, and 4.8 points above GPT 6.1 Sol. Claude Opus 5.5 scores 45.5% there at $12.82 a task.
Professional work for a fraction of the cost per task
Agents' Last Exam v1.1 tests long, multistep professional workflows. GPT 6.2 Sol scores 73.0% at max effort for $1.32 a task, level with GPT 6 Astra at 73.1% for $8.99 and Claude Fable 5.2 at 72.8% for $60.05.
More done with fewer tokens
On OpenAI's GDPval v2 knowledge-work tasks, GPT 6.2 Sol scores 46.9% at max effort against 41.5% for GPT 6.1 Sol, while writing about 17% fewer output tokens. On LifeVal's everyday tasks it beats GPT 6 Astra, 83.8% against 79.8%, at about a seventh of the cost. The independent Artificial Analysis Intelligence Index scores it 57 at max effort.
What GPT 6.2 Sol is best for
Coding agents
Codebase changes that meet a maintainer's bar. 43.8% on FrontierCode 2.0, within half a point of GPT 6.2 Astra.
Scientific computing
Data analysis and simulation from the command line. 68.0% on Terminal-Bench-Science 0.1 for $2.21 a task.
Business workflows
Multi-step professional work across files and apps. 73.0% on Agents' Last Exam v1.1 at max effort.
Legal and document work
Deliverables graded on every requirement. 11.1% all-pass on Harvey LAB-AA v1.1, ahead of Claude Fable 5.2 at 7.2%.
Secure code
Code for file serving, database queries, and other risky paths. 55.6% on OpenAI's SecureSWE against 49.7% for GPT 6.1 Sol.
Cached agent loops
Large repos or tool lists re-read on every step, with cached input at 5% of the input rate.
GPT 6.2 Sol use cases
Put GPT 6.2 Sol behind a coding agent that reads the repo, plans a change, edits across files, runs the tests, and reviews its own diff. Keep the repository and tool definitions in a stable prefix and append new turns, so each step re-reads them at the cache-read rate. Hand it data analysis and simulation jobs in a sandboxed terminal. Load contracts, filings, or financial models into the 1M-token window and ask for a memo or brief that meets every point of the request. Teams running GPT 6 Astra for agent work can move most of those steps to GPT 6.2 Sol and keep GPT 6.2 Astra for the hardest research.
GPT 6.2 Sol limitations
GPT 6.2 Astra is still the stronger model at the top end. It scores 77.1% on Terminal-Bench-Science 0.1 against 68.0%, and 78.3% on Agents' Last Exam v1.1 against 73.0%. Keep Astra for the most difficult research and professional work.
Max effort is slow to start. Artificial Analysis measures 57.55 seconds to the first answer token at max effort, against 5.60 seconds at high. Low effort gives up too much on some tasks: on LifeVal it scores 25.3% at low against 70.3% at medium. Start at medium or high for interactive work. Reasoning cannot be switched off, since the API rejects none.
It is less open about blocked work. In OpenAI's blocker test, where a file the task needs is removed, GPT 6.2 Sol fails to mention the problem 24.08% of the time without a developer prompt, against 14.00% for GPT 6.1 Sol. A developer prompt asking it to report blockers brings that down to 8.88%.
Security work is restricted. OpenAI launched it with tighter cybersecurity safeguards that also block some legitimate vulnerability research and security testing.
Long prompts cost more. Once a request's prompt passes 272,000 tokens, the whole request bills at the long-context rate: double on input, cache read, and cache write, and 1.5x on output.
GPT 6.2 Sol vs GPT 6.1 Sol
The price per input and output token is the same, so there is little reason to stay on GPT 6.1 Sol. At max effort GPT 6.2 Sol scores higher on every capability benchmark OpenAI published for both: Terminal-Bench-Science 0.1 (68.0% against 57.0%), FrontierCode 2.0 (43.8% against 39.0%), GDPval v2 (46.9% against 41.5%), SecureSWE (55.6% against 49.7%), and LifeVal (83.8% against 79.5%). On DeepBinary it finds Windows vulnerabilities worth about $687,000 in estimated bounties against $318,000. Keep GPT 6.1 Sol only if you run everyday tasks at low effort, where it scores 66.8% on LifeVal against 25.3%, or rely on the model to report blocked work without being told to.
When to use GPT 6.2 Sol
Use GPT 6.2 Sol as the default model for coding agents, scientific computing, and multistep professional work, where you want GPT 6 Astra's results at GPT 6.1 Sol's token price. Step up to GPT 6.2 Astra for the hardest research and document work, and drop to GPT 6.1 Terra for cheaper computer use and visual reasoning.
FAQs
People also ask about GPT 6.2 Sol
openai/gpt-6.2-sol, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key. The same ID also works on /v1/responses and /v1/messages, and tool calls work on all three.
Sign up on UnificAlly and create a key on the API keys page. One key and one balance cover GPT 6.2 Sol and every other model, with no subscription.
Send a POST request to https://direct-api.unifically.com/v1/chat/completions with the model ID openai/gpt-6.2-sol and a messages array. The reply comes back in the same response, in the OpenAI chat completion format. Set "stream": true to receive it as it is written.
1,050,000 tokens, of which up to 922,000 can be input, with a max output of 128,000 tokens and a knowledge cutoff of June 1, 2026. It takes text and image input and returns text.
It is the same price per token with better results. At max effort it scores 68.0% on Terminal-Bench-Science 0.1 against 57.0%, 43.8% on FrontierCode 2.0 against 39.0%, and 46.9% on GDPval v2 against 41.5% while writing about 17% fewer output tokens. It is also less likely to disclose a blocked task unless the developer prompt asks it to.
On coding, nearly. It scores 43.8% on FrontierCode 2.0 against Astra's 44.2%, at $1.29 a task against $5.97. Astra leads on harder research and professional work: 77.1% against 68.0% on Terminal-Bench-Science 0.1, and 78.3% against 73.0% on Agents' Last Exam v1.1. OpenAI prices Sol at a quarter of Astra's per-token rate. # price-ok: benchmark cost per task, not our rate
low, medium, high, xhigh, and max, set with reasoning_effort. The API rejects none, so every request does some reasoning.
Billing switches to the long-context rate for the whole request once the prompt passes 272,000 tokens. Input, cache read, and cache write bill at double the standard rate and output at 1.5x. Both rates are listed on the pricing page.
OpenAI offers an Ultrafast mode for GPT 6.2 Sol at a higher token rate. This page covers the standard openai/gpt-6.2-sol model.
OpenAI released GPT 6.2 Sol on November 17, 2026, alongside GPT 6.2 Astra, as the upgrade to GPT 6.1 Sol. It is live on UnificAlly as openai/gpt-6.2-sol.
GPT 6.2 Sol guides and comparisons
Review pricing, limits, and tested outputs before running this model.
Models similar to GPT 6.2 Sol
Browse all modelsGPT 6.2 Astra API
OpenAI's top model for long professional work, 78.3% Agents' Last Exam, first on the Vals Science Index, 1M context, cheaper per token than GPT 6 Astra.
- Text to Text
GPT 6.1 Sol API
Near-Astra coding and computer use at a fifth of Astra's price, 75.2% DeepSWE, 1M context, cached input at half GPT 6 Sol's rate.
- Text to Text
