SWE-1.5

CALIBRATED. Cognition's first numbered SWE model, released 2025-10-29 in Windsurf (now Devin Desktop). DIRECT PREDECESSOR of SWE-1.6 (in-graph anchor 37ee10dd-cb86-818b) — SWE-1.6 was post-trained from the same lineage. NARROW SOFTWARE-ENGINEERING AGENT, not a general chat model: frontier-size (hundreds of billions of params), trained end-to-end via RL on real coding tasks using the Cascade agent harness + Devin production traffic (millions of agent rollouts); first public production model trained on the NVIDIA GB200 generation. SIGNATURE TRAIT = SPEED: partnered with Cerebras for up to 950 tok/s (~6x Haiku 4.5, ~13-14x Sonnet 4.5, ~10x Opus 4.7, ~3x Composer 2) — designed to keep devs in the 'flow window' (<5s task completion). BENCHMARK (semi-independent, Scale AI's SWE-Bench Pro, 731 tasks / 41 repos): 40.08%, ranked #2 at launch behind Claude Sonnet 4.5's 43.6% — near-frontier coding at a fraction of the latency. DIMENSION RATIONALE vs SWE-1.6 (Code 15): Code Gen 14 (SWE-Bench Pro 40.1% vs 1.6's 50.4%, ~10pp lower); Planning/Exec 14 + Tool Use 15 + Tenacity 13 (SWE-agent strength, just below 1.6 which added length-penalty/anti-looping refinements); Speed 19 (SAME 950 t/s Cerebras tier — top of the entire DB, tied with 1.6); Cost 17 (free tier at launch window). DELIBERATELY LOW generalist dims (World 8, Sci 9, Creativity 8, Multimodal 6, Math 11) — coding agent with no general-knowledge/multimodal remit; scoring it as a generalist would misrepresent it (same honest-narrowness pattern as SWE-1.6). No Artificial Analysis Index (not a general API model). Lineage note: SWE family initiated by the original Windsurf team May 2025, before the OpenAI acquisition collapsed and Cognition acquired Windsurf. Sources: cognition.ai/blog/swe-1-5, devin.ai/blog/swe-1-5, Cerebras case study, Scale AI SWE-Bench Pro, Simon Willison testing notes. Anchored against SWE-1.6.