● THE COGNITION ANNALS · node 07 · uplink secure · channel: DOGBED · knap → k4o → textile → oliver → html

INDEX COGNITION DEVIN WINDSURF PRODUCTS SWE-1 SWE-1.5 SWE-1.6 SWE-1.7 SWE-2 TIMELINE LEDGER COMMITS

transmission 09 · model wing, flagship hall

SWE-2

September 10, 2026. The first SWE model post-trained on a multi-trillion-parameter base — Kimi K3 under the hood, Cognition’s refined RL recipe on top — and the first coding model to sit within a single point of the frontier while costing roughly a third as much to run.

THE HEADLINE

50.0%	−64%	92.8%
frontiercode 1.1 main	cost vs fable 5.1	terminal-bench 2.1

50.0% on FrontierCode 1.1 Main — within one point of Fable 5.1, at 64% lower rollout cost. Within a few points of GPT-6 Astra at roughly a quarter of the price. Ahead of its own predecessor SWE-1.7, Grok 4.6, and GPT-5.6 Sol on the comparison table Cognition published. And 92.8% on Terminal-Bench 2.1 — the top of that board.

HOW IT WAS TRAINED

Base — Kimi K3: the recipe ported onto open weights at multi-trillion-parameter scale.
Recipe — a refinement of the multi-turn RL stack proven on Kevin-32B, SWE-1.5→1.7, scaled to a frontier-size base for the first time.
Objective — jointly optimize capability and cost; the Pareto curve is the target, not a single score.

EFFORT LEVELS — A FIRST FOR THE FAMILY

effort	behavior
Medium	scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less
High	more thinking for harder problems
Max	learned to spend more tokens and turns to reach the highest scores

The small detail the curators love: on medium effort, SWE-2 makes its first code edit at step 18 where SWE-1.7 waited until step 48 — it explores more deliberately, then commits. Intelligence, in this lineage, is partly the discipline not to touch the keyboard.

WHERE IT LIVES

surface	status at launch
Devin Desktop	available day one
Devin CLI	available day one
Devin Web	rolling out
Fusion	rolling out

Launch terms: free for Pro, Max, and Teams subscribers for the first month. Benchmark disclosure note, per the lab: each model evaluated in its native harness — Claude Code for Anthropic models, Codex for OpenAI, Grok Build for xAI, Devin CLI for open weights — best score across effort settings.

Our team uses SWE-2 daily in Devin for feature development, debugging, and complex visualizations. — Cognition, September 2026 — the flagship eats its own cooking

THE LINEAGE

SWE line, as archived oct 2026
SWE-1      may 2025    first family — 1 / lite / mini
SWE-1.5    oct 2025    speed era — 950 tok/s on cerebras
SWE-1.6    apr 2026    model-UX era — trained for feel
SWE-1.7    jul 2026    harness-native — RL inside devin
SWE-2      sep 2026    flagship — multi-trillion RL, effort levels
swe-grep   oct 2025    context specialist
swe-check  ----        quick-review specialist
kevin-32b  mid 2025    the open-weights proof of recipe

That is the end of the model wing. The exit leads back to the index — or sideways to the full chronology.

— END OF TRANSMISSION · CTR-(ok)VERIFIED% —

INDEX COGNITION DEVIN WINDSURF PRODUCTS SWE-1 SWE-1.5 SWE-1.6 SWE-1.7 SWE-2 TIMELINE LEDGER COMMITS

+++ COGNITION ANNALS +++ DEVIN · THE AI SOFTWARE ENGINEER +++ SWE-1 → SWE-1.5 → SWE-1.6 → SWE-1.7 → SWE-2 +++ WINDSURF FOREVER +++ COMPILED BY DOGBED, NOT WRITTEN BY HAND +++ NO JAVASCRIPT WAS HARMED IN THE MAKING OF THIS ARCHIVE +++

rendered by dogbed — a document compiler, not a CMS. one binary, one template, thirteen documents.