Breaking

Cognition Unveils SWE-1.7 Coding Model, Served at 1,000 Tokens per Second

July 8, 2026 at 11:47 EDT

Software Dev & Coding
Foundation Models
AI Agents

Cognition AI on July 8, 2026 announced SWE-1.7, its latest coding model optimized for Devin, and made it immediately available across Devin (Web, Desktop and CLI). The company says the model reaches benchmark scores within a few points of frontier models while running at far lower cost and generating 1,000 tokens per second.

July 8, 2026 · Cognition

SWE-1.7: near-frontier coding, at a fraction of the cost

Cognition's new coding model for Devin scores within a few points of the leading frontier models while running at 1,000 tokens per second — for under $2 a task.

42.3%

FrontierCode 1.1 Main

1,000

tokens / second

$1.97

cost per task

FrontierCode 1.1 — Main set (% code worth merging)

46.5%

Opus 4.8

43.0%

GPT-5.5

42.3%

SWE-1.7

38.5%

Opus 4.7

30.1%

Kimi K2.7

9.4%

SWE-1.6

SWE-1.7 (orange) sits just below GPT-5.5 and Opus 4.8 — and leaps far past its own predecessor SWE-1.6 and its Kimi K2.7 base.

"RL is not hitting its limit"

A refined reinforcement-learning recipe keeps yielding gains as Cognition scales — trained with distributed RL across four data centers on multiple continents.

Compressed weight deltas Entropy maintenance Self-compaction Muon optimizer

The upside

More time spent on investigation and research means greater care — and its fast 1,000 tok/s generation helps preserve daily workflow flow.

The watch-out

The deeper research behavior can also touch a wider range of edited files. Real-world reactions are still accumulating.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.