Launch the AI feature, not the model: harness, free window, effort levels

You are deciding how much of a new AI model to expose at launch. Cognition made SWE-2 available through Devin, rather than offering standalone weights, a distinct API, or a public token price list. The product surface became the first trust test: the user needs to see what the feature does, what it costs, and where the trial ends before paying for a surprise. A model can be excellent and still launch badly if the interface hides the cost, the effort, and the trial boundary.

The launch checklist is a trust test

• Model access: Can someone use the feature without access to weights, a standalone API, or a public per-token rate card? If yes, the launch is harness-first.
• Free window: Does the trial show a start date, an end date, and a clear upgrade path? If not, the window leaks.
• Effort levels: Can a user pick medium, high, or maximum effort and see what each changes? If not, cost stays hidden.
• Benchmark disclosure: Does the page name the benchmark, the comparison set, and the cost or speed claim? If not, the claim is decorative.

The harness is the first trust surface

A harness-first launch keeps the model behind a product surface the user can understand. Cognition released SWE-2 to Devin Desktop and Devin CLI on September 10, 2026; Devin Web and Devin Fusion were still in rollout. That choice makes the first trust question practical: can the user see what the feature does before they see how the model is sold?

When a model is exposed as an API before the product is ready, the user starts comparing tokens, latency, and raw capability. A fair comparison is the wrong first conversation for a feature. The harness answers the task question first: what will this do for my work, what does it touch, and what does it cost when it works? Keep the model behind a named surface. A desktop app or a command-line tool is enough until the product can explain the result in the user's language. The first screen should name the surface, the task, and the boundary of the trial.

A free window is a contract, not a coupon

A free window works when it behaves like a contract. Say you give users a window to run the agent on real work. The trial screen shows the start date, the end date, and what happens when the window closes. If the user cannot see the boundary, the trial becomes a leak: they either stop before they trust it or continue after they should have decided. Put the deadline on the same screen as the action, and make the upgrade path a button, not a mystery. The boundary is part of the product, not a billing footnote.

The window also needs an exit. When it ends, the user should know whether their work remains accessible, whether the agent can resume, and whether the price changes. A trial that deletes context or locks files at the boundary teaches distrust. A trial that preserves the work and shows the next price teaches that the product is ready for a decision. Make the end state part of the onboarding, not a surprise in the billing email.

Effort levels and benchmark claims need visible receipts

Effort levels are the most honest cost control in an agent UI. Cognition trained the medium, high, and maximum reasoning-effort modes in a single reinforcement-learning run. That lets the product present the dial as a real choice instead of a hidden knob. On Cognition's FrontierCode Main 100-task set, SWE-2 medium produced its first meaningful code edit after a median of 18 steps; SWE-1.7 needed 48. On FrontierCode 1.1 Main, SWE-2 medium beat SWE-1.7 on score, used 58% fewer turns, and averaged 81% lower cost.

On FrontierCode Main, SWE-2 scored 50.0%, behind GPT-6 Astra at 53.3% and Fable 5.1 at 50.9%. SWE-2's mean rollout cost ran 64% below Fable 5.1 and about one-quarter of GPT-6 Astra. The named comparison set matters as much as the score does. A score without the set invites the user to guess the conditions. Put the benchmark name, the comparison set, and the cost claim next to the effort dial, so the user sees why the cost changes. A visible receipt turns a benchmark from marketing into a decision aid.

A visible label says what medium does: more reasoning, more steps, more cost, and usually a better chance on a hard task. The default matches the user's first real task, not the setting that looks impressive in a demo. If the product cannot show a rough cost difference, show the direction: this setting works harder, and it takes longer. The user can then choose the setting that fits the task and the budget.