SWE-1.5
Cognition's SWE-1.5 is a frontier-size coding agent model served by Cerebras at up to 950 tokens/s, claimed near-SOTA on SWE-Bench Pro.
Model, inference stack and agent harness were co-designed, and Cognition says it runs 6x faster than Haiku 4.5 and 13x faster than Sonnet 4.5. Tasks that took 20+ s finish in under 5 s. Same day as Cursor's Composer.
- Date
- Wednesday, 29 October 2025
- Lab
- Cognition (Windsurf)
- Kind
- model
- Access
- app only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Inference speed | up to 950 tok/s 6x Haiku 4.5, 13x Sonnet 4.5 (Cognition) | company |
'Near-SOTA' on SWE-Bench Pro is Cognition's own measurement.
Sources
This record was checked against its sources on 6 October 2026. How we check