For most of its life, Manus has been a closed loop: you signed up, Manus chose the model, and the bill arrived from Manus. That arrangement made the product easy to trust and impossible to argue with, and anyone who already had a negotiated inference contract or a preferred open-weight model had nowhere to put it. Manus Flex is the answer to that, announced this week as a new module inside the Manus workspace: bring your own API key, and let a supported inference provider supply the model while Manus keeps doing the part it is actually good at.

The framing matters here, because "bring your own key" usually means you are also bringing your own agent. Flex is not that. Manus is opening up the model layer, not the harness.

The agent stays, the model becomes yours

Since launch, Manus has run a managed agent product. It handled model selection, the agent harness, and the infrastructure underneath: web app hosting, databases, execution environments. Flex changes the first of those three and leaves the other two alone.

Connect a supported provider, and its inference gets wired into the same projects, tools, execution environments, and agent workflows you already use. Manus still owns the part that does the work: planning the task, calling tools, operating across sandboxes, and turning raw model output into something finished. What you replace is the model underneath, not the machinery around it.

Configuration is deliberately thin. A Flex setup is a model plus a reasoning-effort level for the provider you connected. There is no new orchestration layer to learn and no second workspace to manage, which means the learning curve is roughly: paste a key, pick a model, pick an effort setting, run the task you already had queued.

The bill is now split in two

Here is the part teams should read twice. Model inference is billed directly by the provider you connect, at that provider's rates, on your account. Everything else a Manus task touches still consumes Manus credits. Tool calls, hosted apps, databases, sandbox execution — all of it stays on the Manus meter.

That is not a footnote; it is the design. Flex is a way to move the largest variable line item — inference, which is usually the bulk of an agent workload — onto infrastructure you already pay for, without giving up the sandboxed environment the agent needs to finish the job. If your provider contract is cheaper than Manus's blended rate, you keep the difference. If you are sensitive to a model provider's data handling, you can now choose that provider explicitly.

It also means Flex is not a free tier in disguise. A cheap key does not make the surrounding infrastructure free, and anyone evaluating this should model both halves of the bill before switching a production workflow over.

Three inference partners, and a program behind them

Flex launched with a partner roster rather than a single option. OpenRouter, Fireworks, and Modal are the initial inference partners, and Manus framed the arrangement as the first cohort of a Flex Inference Partner Program — a standing path for more providers to plug into Flex.

The partners are describing it in complementary terms. Alex Atallah, co-founder and CEO at OpenRouter, framed the partnership around model choice. Pranav Jain, product lead at Fireworks, pointed at control over which model powers a given product. Adam Azzam at Modal framed it as connecting an open model ecosystem to Manus. Three different pitches, one shared theme: the agent is the product, and the model is now a decision the customer gets to make.

To use it, connect a supported inference provider in the Manus workspace, set the primary model and effort level, then run it inside the projects, automations, and agent workflows you are already building. Nothing about the surrounding workflow changes shape.

What I find most interesting is the direction of travel. Manus spent years as the layer that hid model choice from its users, which was the right call when the product was new and the harness was the hard part. Now the harness is apparently good enough to sell without the model attached. For teams that is a genuine unlock: your agent runtime and your inference bill finally come from two vendors you chose separately.