# Cognition's SWE-2 Flash: One-Fifth the Price of SWE-2 in Devin, and the Benchmarks Most Coverage Left Out

Cognition calls it the best American coding model in its class. Its own full table shows SWE-2 Flash ahead of other US open-weight models but behind DeepSeek's and Zhipu's smaller Flash models on Terminal-Bench 4.0, and its own weights aren't public yet.

By Priya Nair (AI Industry Reporter) — published 2026-11-05, updated 2026-11-05
Source: https://aiscoutdaily.com/news/cognition-swe-2-flash-devin-full-benchmarks

## TL;DR
Cognition released SWE-2 Flash on 4 November 2026, a 276-billion-parameter mixture-of-experts coding model with 12 billion active parameters, mid-trained and then reinforcement-trained on top of Thinking Machines' Inkling-Small. It is available now in Devin Desktop and the Devin CLI at one-fifth the price and three times the speed of SWE-2, and is free through 30 November. Cognition's full benchmark table shows it beating every US open-weight model it tested, but trailing GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 on Terminal-Bench 4.0 (20.0 vs 41.8 and 42.3) and SWE-2 itself on DeepSWE 1.1 (60.1 vs 73.0). Cognition is not publishing the weights, citing cyber risk, and plans a program to share them with vetted partners.

Most reports on SWE-2 Flash reprinted three rows of Cognition's results: Terminal-Bench 4.0, FrontierCode 2.0 and DeepSWE 1.1 against SWE-2 and the base model. Cognition's post has a much longer table that includes non-US models and small closed models. That table is the better guide to where Flash fits.

## What shipped

- Model: 276B total and 12B active parameters, the same shape as its base, Inkling-Small from Thinking Machines. Cognition added a mid-training stage of a few trillion tokens before reinforcement learning, the first time its recipe has done so.
- Where: Devin Desktop and the Devin CLI. There is no standalone API.
- Price: one-fifth the price of SWE-2, with no absolute per-token or per-session figure given. It is free through 30 November. RuntimeWire says the free offer is for paid Devin users; Cognition's post just says "free through 11/30".
- Speed: Cognition shows 200 tokens a second for Flash against 60 for SWE-2, using simulated replays of a Terminal-Bench task rather than a live timing test.
- Weights: not public. Cognition says it raised the model's ExploitBench score from 11.3% to 34.8% to help defenders, and is holding the weights back because the same skills help attackers. A partner and researcher access program is promised "in a few weeks".

## Cognition's full benchmark table

| Benchmark | SWE-2 Flash | Beam 1.0 | Muse Spark Open | GLM 5.3 Flash | DeepSeek V4.1 Flash 1017 | GPT-6 Luna | Claude Haiku 5.5 | SWE-2 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Terminal-Bench 4.0 | 20.0 | 4.0 | 8.0 | 41.8 | 42.3 | 16.4 | 39.2 | 27.3 |
| Terminal-Bench 2.1 | 89.2 | 80.1 | 76.4 | 84.3 | 91.0 | 82.0 | 94.4 | 92.8 |
| DeepSWE 1.1 | 60.1 | 43.3 | 60.0 | 63.4 | 77.0 | 66.6 | 59.3 | 73.0 |
| FrontierCode 2.0 | 24.0 | 2.9 | 6.2 | 14.2 | 19.3 | 21.9 | 27.7 | 23.8 |
| SWE-Bench Pro v2 Hard | 89.9 | 76.5 | 82.4 | 88.2 | 86.3 | 80.4 | 90.2 | 90.6 |
| SWE-Atlas QA | 50.0 | 33.9 | 54.0 | 56.5 | 71.0 | 45.2 | 56.5 | 56.3 |
| Model size | 276B-A12B | 501B-A23B | 816B-A42B | 320B-A18B | 763B-A16B | Not given | Not given | 2.8T-A104B |

*All figures from Cognition's announcement, run by Cognition. Not independently verified.*

## What the table shows

- Against US open-weight models (Beam 1.0, Muse Spark Open), Flash leads on almost every row, which supports Cognition's headline claim.
- Against open-weight models from China, the picture changes. GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 score about twice as high on Terminal-Bench 4.0, and DeepSeek's model also leads on DeepSWE and SWE-Atlas QA.
- Against small closed models, Claude Haiku 5.5 beats Flash on Terminal-Bench 4.0 and 2.1, FrontierCode 2.0 and SWE-Bench Pro, while Flash leads on DeepSWE.
- Against SWE-2, Flash is close on FrontierCode and SWE-Bench Pro but well behind on Terminal-Bench 4.0 and DeepSWE, the two longest agentic tests.
- Ricursive's Ripple-65B Preview, an open-weight US model released the same day, isn't in Cognition's table. On Terminal-Bench 2.1, where both report scores, Ricursive reports 80.5 for Ripple, below Flash's 89.2.

## What this means if you pick coding tools

- If you already pay for Devin, try Flash on routine work like code review, documentation and small features through 30 November, while it's free.
- Keep SWE-2, or another frontier model, for long multi-step terminal work, where Cognition's own numbers show a large gap.
- If you need weights you can host yourself, Flash isn't an option yet. DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash score higher on the hardest agentic test here and their weights are available now.

## Sources
- [Cognition: Introducing SWE-2 Flash (official announcement, Nov 4, 2026)](https://cognition.com/blog/swe-2-flash)
- [Unite.AI: Cognition Launches SWE-2 Flash for Faster, Lower-Cost Coding Agents (Nov 4, 2026)](https://www.unite.ai/cognition-launches-swe-2-flash-for-faster-lower-cost-coding-agents/)
- [Superpower Daily: Cognition adds a lower-priced coding model to Devin, keeping its weights private (Nov 4, 2026)](https://superpowerdaily.com/posts/cognition-adds-a-lower-priced-coding-model-to-devin-keeping-its-weights-private)
- [RuntimeWire: Cognition launches SWE-2 Flash in Devin at one-fifth the price of SWE-2](https://runtimewire.com/article/cognition-swe-2-flash-devin-launch)