Cognition

SWE-2

Name: SWE-2
Author: Cognition

SWE-2 is a Cognition model with Not verified context, Not verified in / Not verified out per million tokens, last verified October 2, 2026. Open weights: not verified. Public-board scores are linked to the publisher; vendor-only figures stay in the watch-out, not on the board.

Cognition's software-engineering model is offered as a research preview within Devin, with Medium, High and Max reasoning settings.

Recorded profile verification October 2, 2026Individual provider fields may changeAccess: preview — Research preview inside Devin Desktop, CLI and sessions; enterprise/admin enablement and rollout apply. Generic public inference API access is unverified.

AI model specification details
Specification	SWE-2
Provider	Cognition
Tier	Frontier
Context window	Not verifiedUnverified
Max output	Not verifiedUnverified
Input / 1M tokens	Not verifiedUnverified
Output / 1M tokens	Not verifiedUnverified
Weights	Not verifiedUnverified
Parameters	Multi-trillion-parameter regime (publisher; exact count undisclosed)
Reasoning levels	Medium, High, Max
Modalities	text
Released	September 10, 2026

Pricing tiers: Devin model docs list free self-serve usage through October 15, 2026, then $3 input/$15 output/$0.30 cache input per million tokens. The pricing page says free through October 16, so the exact promotional cutoff conflicts. Legacy credit-based enterprise has 75% off list through December 31; ACU billing differs. These are in-product billing terms, not verified generic direct-API rates.

Best for

Evaluating software-engineering tasks in Devin
Comparing supported reasoning settings in an enabled account

Watch out

A stable generic API model ID/endpoint, exact context/output limits, numeric parameter count and weights/license are unverified. The September 10 announcement establishes in-product Desktop/CLI availability; Web/Fusion were rolling out. Cognition reports a multi-trillion-parameter model trained from Kimi K3, which does not establish inherited limits, licences or scores. Availability and behavior can change during research preview. Do not inherit Devin or older SWE model scores; no model-specific comparable numeric benchmark was established. Voice calls use a separate model and do not receive SWE-2 promotions. Inspect the enabled account's selector and billing terms before relying on promotional access.

Source receipts

Catalog figures for this model were checked against the following sources.

September 10 in-product model introduction (accessed 2026-10-02)
Model variants and route-specific billing (accessed 2026-10-02)
Research preview release notes (accessed 2026-10-02)
Pricing and promotional cutoff (accessed 2026-10-02)
Desktop changelog (accessed 2026-10-02)

Common questions

SWE-2

Answered from the verified figures on this page rather than general guidance.

What is SWE-2 best for?

SWE-2 is a frontier tier from Cognition. It suits evaluating software-engineering tasks in devin, comparing supported reasoning settings in an enabled account. A stable generic API model ID/endpoint, exact context/output limits, numeric parameter count and weights/license are unverified. The September 10 announcement establishes in-product Desktop/CLI availability; Web/Fusion were rolling out. Cognition reports a multi-trillion-parameter model trained from Kimi K3, which does not establish inherited limits, licences or scores. Availability and behavior can change during research preview. Do not inherit Devin or older SWE model scores; no model-specific comparable numeric benchmark was established. Voice calls use a separate model and do not receive SWE-2 promotions. Inspect the enabled account's selector and billing terms before relying on promotional access.

Model comparison catalog Benchmark library