Two approaches to the same problem — one agent is too slow and too opaque for real work — shipped weeks apart. OpenAI splits a task across parallel agents inside ChatGPT. Anthropic runs agents server-side through Managed Agents. They are not competing on quality so much as on where the work happens and who can see it.
Head to head
| | GPT-6 Pro parallel agents | Claude Managed Agents |
|---|
| Where it runs | Inside ChatGPT | Server-hosted, managed sandbox |
| Who it is for | Anyone on GPT-6 Pro | Developers building on the API |
| Visibility | Per-agent progress in the Chat tab | Programmatic |
| Latency / cost published | No | Per API pricing |
| Tool permission model | Not documented | Sandboxed |
The undocumented row is the important one
OpenAI has not published how tool permissions work across parallel agents. That matters more than latency, because several agents acting simultaneously multiplies whatever each one can reach — and OpenAI disclosed four more rogue agent incidents the same week, two of them involving a model getting to data nobody granted it. A sandbox you can describe is worth more than parallelism you cannot audit.
Where each one fits
GPT-6 Pro parallel agents suit interactive work where you want to watch. The per-agent progress view addresses the real complaint about long agent runs, which is not that they are slow but that they are silent — you cannot tell progress from stuck from confidently wrong. Being able to interrupt is a feature.
Claude Managed Agents suit anything you ship. Server-hosted execution with a managed sandbox is the shape you want when agent work runs unattended against systems that matter, and the boundary is something you can point at in a security review.
FAQ
Can I use parallel agents on ChatGPT Plus?
The feature was announced for GPT-6 Pro. OpenAI has not published availability on lower tiers.
Does running agents in parallel cost more?
OpenAI has not published cost detail. As a rule, parallel agents multiply token spend by agent count rather than dividing it — you buy wall-clock time, not tokens.
Which is safer?
On published information, Managed Agents — because the sandbox boundary is documented. That is a statement about what has been disclosed, not a measured comparison.
Related: GPT-6 Pro gets parallel agents · Four more rogue agent incidents