When a company builds an AI product and also sells you the AI inference that powers it, they have a conflict of interest that is structural rather than incidental. Their revenue is a function of how many tokens you consume. Your cost is also a function of how many tokens you consume. These incentives do not point in the same direction.
This conflict shows up in product decisions in ways that are sometimes subtle and sometimes not. Models are selected for verbosity because longer responses consume more tokens. Context windows are managed to maximise re-injection of previously seen content. Features that would reduce token usage are deprioritised. None of this is conscious malice; it is the gravitational pull of an incentive structure.
Dart does not have this problem because Dart does not sell tokens. We never have and we will not. This is not a launch-era positioning claim; it is a commitment we are making public so you can hold us to it.
How BYOM works in practice
When you configure Dart to use an LLM, you point it at a provider directly. You might paste an API key from Anthropic, OpenAI, or another provider. You might point it at a local Ollama instance running on your laptop. You might configure an enterprise gateway your organisation maintains.
In every case, the inference request goes directly from Dart to your configured endpoint. Dart is not in the network path. We do not see the request. We do not see the response. We do not log what model you used, what prompt you sent, or how many tokens the response consumed. There is no telemetry pipeline between the Dart Agent and our servers that carries any of this information.
Your API key is stored in your OS keychain — macOS Keychain, Windows DPAPI, or libsecret on Linux. It is not stored in Dart’s configuration database. It is not transmitted during licence validation. We do not have it. If you asked us what your OpenAI key is, we could not tell you.
What we charge for
We charge for the browser. Specifically, we charge for the features that we build and maintain: the local agent runtime, the hardware identity layer, the semantic indexing system, the BYOM integration layer, the Chromium fork and its security patch cadence, cross-device sync, and the enterprise features (SSO, SCIM, policy management, audit logs).
These are things we build. They have a cost structure that is not token-based. They are the appropriate unit for a subscription fee. A flat monthly price for a browser is something a customer can budget for and understand. A token-based price for AI inference that is secretly marked up is neither of those things.
The cloud inference option
Some users will want to use cloud inference — Anthropic, OpenAI, Google — rather than running a local model. This is fine. Dart supports it fully. The mechanism is the same: you provide your API key, it goes in your keychain, requests go directly from your browser to the provider. You pay the provider at their published rates. We are not in the transaction.
We do not have a referral arrangement with any model provider. We do not receive a cut of your inference spend. We do not have preferred provider status with anyone. The model picker in Dart is sorted alphabetically and we do not accept placement fees. If we ever changed this, it would require a privacy policy update and a public announcement; we are not planning to change it.
The case for local inference
The strongest version of BYOM is local inference — a model running on your own hardware, on your own network, with zero traffic to any external server. This is the architecture we are most enthusiastic about and the one we are building toward.
Local inference at usable speeds has only recently become practical on consumer hardware (see our benchmark post for current numbers). It will continue to improve rapidly. Our prediction is that by 2027, local inference will be the default choice for most Dart users, not a specialist option for people with high-end hardware and technical patience.
When that happens, the BYOM model does not change for us. We still charge for the browser. The inference is still free at the point of use (funded by the hardware you already bought). The privacy guarantee is even stronger: not only is your inference not proxied by us, it does not leave your machine at all.
Why this matters beyond pricing
The pricing point is the most legible way to explain BYOM, but the deeper reason is about alignment. A browser that sells you inference is a browser whose interests are not identical to yours. It wants you to use more AI, in more contexts, for more things. A browser that charges a flat fee for software wants you to find the software useful enough to keep paying for it. These are different motivations, and they produce different products over time.
The browser is the most powerful software most people run daily. It is the front door to every service, every communication, every piece of information that matters to them. The entity that controls the browser has enormous power over the user’s experience and, increasingly, over the decisions the AI makes on their behalf. That entity should have its incentives aligned with the user as cleanly as possible.
Flat software pricing is not perfect alignment — we still want you to renew your subscription, which means we want you to find the product useful. But it is significantly better alignment than “we want you to use more tokens” when we are also the ones deciding how the AI behaves.
BYOM is a design choice that removes a conflict before it can develop. We made it on day one, and we are making it permanent now by writing it down publicly.