The official documentation explicitly lists gpt-6.2-sol as a model that supports the Responses API multi-agent beta. It demonstrates assigning three subagents to review a PR for correctness, security, and test coverage, then having the root agent synthesize their findings.
Good fit: PR diff reviews that can be divided into independent checks, such as reviewing logic for correctness, checking for security issues, and identifying missing tests.
Poor fit: Tasks that must be completed sequentially, are very small, require frequent writes to the same shared resource, or depend on a fixed, deterministic execution graph. Parallel agents increase token usage.
Supported model version: GPT-6.2 Sol (the model identifier in the documentation example is gpt-6.2-sol).
Supported client, agent, or API: Responses API beta. The documentation demonstrates the Python beta SDK. HTTP requests must include the beta identifier.
Recommended reasoning level and parameters: This page does not recommend a reasoning level. The example sets max_concurrent_subagents to 3. The documentation says the default is 3 and that this value suits most workloads.
The following is a simplified adaptation of the official Python example, not a verbatim prompt from the source. Replace diff with the changes to review:
from openai import OpenAI
client = OpenAI()
response = client.beta.responses.create(
model="gpt-6.2-sol",
input=(
"请并行审查以下 PR diff:分别检查正确性、安全性和缺失测试。"
"合并重复或冲突的问题,并按优先级列出发现,附文件和行号。\n\n"
f"<diff>\n{diff}\n</diff>"
),
multi_agent={
"enabled": True,
"max_concurrent_subagents": 3,
},
betas=["responses_multi_agent=v1"],
)The example only shows the request configuration. Production code should also extract the root agent's final response from the SDK return structure. The official example filters response.output for the /root final_answer message.
Confirm that the task can be divided into independent, clearly scoped checks, and provide the complete diff as input.
Create a Responses request with gpt-6.2-sol, set multi_agent.enabled=True, and declare responses_multi_agent=v1 in the beta list.
Set the number of concurrent subagents to 3. The root agent coordinates the work and synthesizes the subagents' findings, including priorities, files, and line numbers.
Manually verify the locations and issues in the report. The multi-agent configuration itself does not prove that a review has passed.
The documentation describes multi-agent as a beta capability in which a root agent delegates tasks in parallel and aggregates the subagents' results. It explicitly lists GPT-6.2 Sol.
The page's Python PR review example uses model="gpt-6.2-sol", enables multi-agent, sets the concurrency limit to 3, and includes the beta identifier responses_multi_agent=v1.
The documentation says the concurrency limit applies to the entire agent tree, excluding the root agent. The default is 3, which it recommends for most tasks. Adding agents also increases token usage.
The documentation's beta support also includes GPT-6.2 Astra, GPT-6.1 Sol, and all GPT-5.6 models. This is the API beta's general availability and does not mean these models are equivalent to the target of this article. This article covers only the visible GPT-6.2 Sol example.
The collected page did not show an author or publication date, and the latter part of the body was not fully visible. No publication date was inferred and no unseen parameters were added.
Multi-agent is available as a beta in the Responses API. The documentation warns that the item schema may change. Check the relevant model page before integrating and manage compatibility as a beta capability.
Subagents share the request's model and available tools. In the documentation's example, the application should not execute or submit the outputs of hosted collaboration actions itself.
The documentation's example is a PR review workflow, not an independent evaluation. The page provides no measured quality improvement, benchmark data, or specific reasoning level.
Use the official Python example to create a Responses request with a PR diff, request separate checks for correctness, security, and test coverage, and inspect the root agent's final response. The original page is accessible. This evidence comes from the content visible during collection, as the page extraction did not fully show the rest of the body.
GPT-6.2 Sol