Cosine

You will notice the difference on large tasks, because the deciding trait here is fan-out: a primary agent delegates to subagents that work in parallel, so a job that would have been a long single thread becomes several short ones. The same account follows you from the command line to a hosted workspace to a desktop window, and imported repositories come back as pull requests rather than patches you have to place.

Pick it if your tasks are big enough to divide. Pick Factory Droid when you want the work to live in the cloud by default, or a terminal agent if you prefer one thread you can watch.

The inconsistency is the problem. Work executed remotely runs in a container built from a Dockerfile; the identical agent run on your own machine has no such boundary, and the row says so plainly. That means the risk of a command depends on which surface you happened to start from, and nothing in the interface makes that distinction loud. Users generalise from the safe case to the unsafe one, every time.

What it does right: those remote environments are defined by a Dockerfile you write, so the execution context is reproducible and reviewable rather than a vendor image you must trust.

Take the citation apart. 1. The score is self-reported and appears on the marketing site rather than in a harness anyone can run. 2. The evaluation set is the vendor's own, so the figure is not comparable with any public leaderboard, and 53.9% carries no meaning without the subset, the scaffold and the attempt count. 3. No methodology page is linked from the claim.

This is not evidence of weakness; it is an absence of evidence, and the two are routinely confused in this category. A reader who wants a number should generate one on their own repository, since the vendor has supplied the tooling to do exactly that.

Most companies in this category resell somebody else's inference. This one trains its own family and offers a third-party catalogue beside it, which is the difference between a wrapper and a supplier, and it is the reason the ownership question here is worth asking at all. The metered layer is where power sits: additional credits priced between four and six dollars a million let the company hold a spread that a pure reseller cannot.

Training is expensive and the catalogue undercuts it, so the model bet must pay off soon. Likely acquirer: a platform vendor wanting a coding model team. Position: buy the product, not the roadmap.

The number I take to finance starts at $1,140 a month, sixty seats at the nineteen-dollar tier, and then becomes a variable I cannot forecast because the included credits are consumed by task size rather than headcount. Above that the plans climb steeply and the enterprise tier is a conversation, not a price. There is no headless mode, so nothing here runs unattended in our pipelines, and the row records no SSO, no SCIM and no audit log.

Onboarding is a day. Not yet. I need a per-seat spending cap and an identity story before this reaches a procurement meeting.

Proprietary, so there is nothing to read and no fork to keep. Grudgingly, two things work in my favour. Signing in with a subscription I already pay for routes billing to that provider instead of the vendor's credits, which is the rare escape hatch a closed tool actually documents. And the browser tooling is honest plumbing: I launch Chrome with a debugging port myself and point a config file at it, rather than being handed a black box.

MCP servers plug in as a client. The model never lives on my hardware, which is where this stops being mine.