Cognition Says SWE-2 Flash Beats Every American Open-Weight Model. Its Own Table Shows It Losing Terminal-Bench 4.0 to DeepSeek and GLM by 20 Points, and You Cannot Download It.
Cognition launched SWE-2 Flash on November 4: a 276-billion-parameter mixture of experts with 12 billion active, mid-trained and reinforcement-trained from Thinking Machines' open-weight Inkling-Small. Cognition's claim is that it outperforms all American open-weight models. On Cognition's own table it scores 20.0 on Terminal-Bench 4.0 against 42.3 for DeepSeek V4.1 Flash 1017 and 41.8 for GLM 5.3 Flash, and its best result, 24.0 on FrontierCode 2.0, is on a benchmark Cognition built. Access is Devin Desktop and CLI only, with no API and no weights; one-fifth the price of SWE-2 is a ratio with no rate card behind it. We read Cognition's post and the Devin docs.
By T.W. Ghost
Cognition launched SWE-2 Flash today, November 4, 2026: a 276-billion-parameter mixture-of-experts model with 12 billion active, built by mid-training and then reinforcement-training Thinking Machines' open-weight Inkling-Small. The headline Cognition chose is that it "outperforms all American open-weight models." The qualifier is doing the work. On the table Cognition published, the models that beat it on the hardest agentic benchmark are Chinese, open, and downloadable, and SWE-2 Flash is none of those three.
The numbers, all Cognition's: Terminal-Bench 4.0 rose from 0.9% on the base model to 20.0% after training. On the same table, DeepSeek V4.1 Flash 1017 scores 42.3, GLM 5.3 Flash 41.8, Claude Haiku 5.5 39.2, and Cognition's own larger SWE-2 27.3. FrontierCode 2.0, the private benchmark Cognition built and which now makes up 15% of Artificial Analysis's index, rose from 3.2% to 24.0%, a hair above SWE-2 at 23.8 and under Haiku 5.5 at 27.7. Access is Devin Desktop and the Devin CLI, with no API and no public weights; a partner weight-sharing program is "in a few weeks." We read Cognition's post, the Devin model docs, and AA's changelog. Here is what held up.
What Cognition Shipped
| SWE-2 Flash (November 4, 2026) | |
|---|---|
| Architecture | 276B total, 12B active MoE (vendor) |
| Base | Thinking Machines' Inkling-Small, open weights |
| Training | Mid-training, then reinforcement learning (vendor) |
| Access | Devin Desktop and CLI; no API; no weights; partner weight program "in a few weeks" |
| Price | "One-fifth the price" of SWE-2; no per-token rate published |
| SWE-2 list price | $3.00 in / $15.00 out / $0.30 cached per 1M (Devin docs) |
| Free period | SWE-2 free for self-serve through November 30 (docs); SWE-2 Flash free through November 30 is press only |
| Speed | "3x the speed" of SWE-2; the demo replays transcripts at a simulated 200 vs 60 tokens per second |
| AA score | None as of November 5 |
Open base, closed result. Inkling-Small is open weights. SWE-2 Flash, trained on top of it, is not. That is permitted; it is also the opposite of the "open-weight" framing in the headline.
A ratio is not a price. One-fifth of $3.00 and $15.00 would be $0.60 and $3.00; that is arithmetic, not a rate card, and Cognition has not published one.
The speed demo is a replay. Two hundred tokens per second against sixty is a transcript played back at simulated rates, by Cognition's own description, not a measured endpoint.
The Benchmarks, All of Them Cognition's
| Benchmark (Cognition's table) | SWE-2 Flash | Comparators on the same table |
|---|---|---|
| Terminal-Bench 4.0 | 20.0 (base: 0.9) | DeepSeek V4.1 Flash 1017 42.3; GLM 5.3 Flash 41.8; Claude Haiku 5.5 39.2; SWE-2 27.3 |
| FrontierCode 2.0 (Cognition's own benchmark) | 24.0 (base: 3.2) | SWE-2 23.8; Haiku 5.5 27.7; DeepSeek V4.1 Flash 19.3 |
| DeepSWE 1.1 | 60.1 (base: 17.4) | |
| ExploitBench | 34.8% (base: 11.3%) |
Three honest observations:
- "American" is the load-bearing word. Drop it and the sentence is false on Cognition's own table, where DeepSeek and GLM roughly double SWE-2 Flash on Terminal-Bench 4.0.
- The best score is on the home field. FrontierCode 2.0 is Cognition's benchmark. A win there is a vendor result on a vendor test, and it is also the row where SWE-2 Flash edges its own bigger sibling.
- The base-to-trained jumps are real and interesting. From under one percent to twenty on Terminal-Bench is what the mid-training and RL bought. It is a good result for a 12B-active model. It is not a leaderboard position.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Announced Nov 4; 276B-A12B; Inkling-Small base; mid-training plus RL | Vendor (Cognition post) |
| Terminal-Bench 4.0, FrontierCode 2.0, DeepSWE, ExploitBench scores and comparators | Vendor (Cognition's table) |
| Devin Desktop and CLI only; no API; no weights; partner program in weeks | Vendor (Cognition post) |
| SWE-2 $3.00 / $15.00 / $0.30; SWE-2 free for self-serve through Nov 30 | Verified (Devin docs) |
| SWE-2 Flash per-token price | Unpublished |
| SWE-2 Flash free through Nov 30 | Press only |
| FrontierCode is Cognition's private benchmark, part of AA Index v5 | Verified (AA changelog) |
| Independent AA score | None as of Nov 5 |
Who Should Use It
Try SWE-2 Flash if you already live in Devin; the model is there and the comparison worth running is against your own repository, not Cognition's chart.
Skip the "best American open model" line, because the model is not open and the best open models on the chart are not American. If you need weights you can host, the two that win Cognition's own Terminal-Bench column are DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash.
The frontier is elsewhere: on AA's Coding Agent Index, where SWE-2 Flash does not yet appear, and where Cognition's own Devin Fusion scores 62. Our model comparison has the current standings side by side.
Sources
*Not sure whether a closed model trained on an open base belongs in your stack? Take the free 2-minute quiz and get matched. Then read about the base model's successor, announced the next day with its own weights still pending: Thinking Machines' Inkling 2 Small.*