GPT-6.1 Sol can spawn its own subagents, and OpenAI sets no limit on the number
OpenAI's GPT-6.1 Sol can now create its own subagents and run them in parallel, and the documented ceiling on how many is that there is not one. "Multi-agent runs also have no fixed limit on tree depth or the total number of subagents created during a run," OpenAI says in its Multi-agent guide, which describes the feature as letting a model "spin up and coordinate subagents in parallel, synthesizing their work to provide a final response".

OpenAI says Multi-agent is in beta with GPT-6.1 Sol and all GPT-5.6 models. Image: OpenAI
It is a beta, switched on per request. You set multi_agent.enabled on a Responses API call and pass the beta header OpenAI-Beta: responses_multi_agent=v1, and OpenAI says it works with GPT-6.1 Sol and all GPT-5.6 models. OpenAI announced the pairing in its API changelog on September 29, the day it released GPT-6.1 Sol.
There is one knob, and it is not a cap on the total. max_concurrent_subagents sets how many subagents can be active at the same time across the whole tree, counting grandchildren and deeper but excluding the root agent. The default is three, which OpenAI recommends "for most workloads", and the guide says the API "does not impose a fixed upper bound on this setting".
The model does the delegating, not your code
The root agent is called /root, and subagents take paths beneath it such as /root/researcher or /root/reviewer/tester. Six collaboration actions are hosted by OpenAI rather than your application: spawn_agent, send_message, followup_task, wait_agent, interrupt_agent and list_agents. "Your application should not execute these or submit outputs for them," OpenAI says.
Your code still has work to do. Every agent in the tree can reach the tools configured in the request, and any of them can emit a function call that your application has to execute and return.
One design choice is unusual enough to flag. When Multi-agent is on, OpenAI appends its own instructions to the root agent and every subagent as a new developer message, and "you cannot edit or remove these instructions". OpenAI publishes that injected text and tells developers to frame their own instructions as additive to it, so this is documented behaviour rather than something hidden.
Token usage is the part OpenAI leaves open
On spend, the guide offers one line: "Note that adding subagents can increase token usage", with a warning that parallelism helps least on tasks that depend on a single ordered chain of reasoning or are dominated by one slow external operation. No cost figures, no speed-up factors and no benchmarks are given.
The prices are published separately. GPT-6.1 Sol runs at $2 per million input tokens and $10 per million output, with cached input at $0.10, and prompts over 272K input tokens priced at double the input rate and 1.5 times the output rate for the whole request. Multiply that by a tree whose size the model decides and the arithmetic gets hard to forecast.
Separately, we reported in September that an OpenAI Codex developer called two parallel agents the practical limit after a $19,300 run. That was Codex, a different product with a different harness, and nothing in the Multi-agent guide refers to it.
"No fixed upper bound" also stops at the setting itself. Account rate limits still apply: GPT-6.1 Sol is capped at 500 requests and 500,000 tokens a minute on tier one, rising to 15,000 requests and 40 million tokens on tier five.
The beta has rough edges to match. The /responses/compact endpoint, reasoning.summary and max_tool_calls are all unsupported while Multi-agent is enabled, server-side compaction switches itself on and is applied separately to the root agent and each subagent, and OpenAI warns that "item schemas may change" before the feature leaves beta.