AI coding agents keep getting smarter — and more expensive. Cognition, the company behind Devin, thinks most of your daily coding work doesn't need the most expensive brain in the room. On November 4, 2026, it launched SWE-2 Flash, a smaller, cheaper, faster sibling of its flagship coding model SWE-2.
The headline numbers are hard to ignore: three times the speed at one-fifth the price — and it's free to use in Devin Desktop and the Devin CLI through November 30, 2026. Here's what that actually means for developers, indie hackers, and anyone paying an AI coding bill.
What Is SWE-2 Flash?
SWE-2 Flash is Cognition's new coding model, purpose-built for routine software-agent work. Cognition describes it as "a fully American coding model that sets a new bar for intelligence at low cost and high speed." It's available today — no waitlist — inside Devin Desktop and the Devin CLI.
Under the hood, Flash is a 276-billion-parameter mixture-of-experts model with only 12 billion parameters active per token — adapted from Inkling-Small, the open-weight base model released by Thinking Machines Lab on July 30, 2026. The MoE design is the whole trick: activating a small fraction of the network per token keeps inference fast and cheap while the total parameter count stays large enough to handle real coding tasks.
Cognition is not sharing the model's weights publicly for now. The company says it's building a secure sharing program for trusted industry partners and research labs instead. If you wanted to self-host or fine-tune it, you're out of luck — unlike the open-weight models we've covered recently, such as DeepSeek, Qwen, and Kimi's open releases, Flash runs exclusively inside Devin.
Flash vs SWE-2: What's Actually Different
SWE-2, launched in September 2026, is Cognition's closest-to-frontier coding model — post-trained from Moonshot AI's Kimi K3 and landing within a point of Claude Fable 5.1 on FrontierCode 1.1 Main at 64% lower cost, per Cognition's reported figures. Flash is explicitly not trying to be that model. It's the everyday-workhorse version.
Here's how Cognition frames the split:
| SWE-2 | SWE-2 Flash | |
|---|---|---|
| Role | Flagship — hardest coding tasks | Everyday work — speed and cost |
| Speed | Baseline | ~3x faster |
| Price | $3.00 / $15.00 per 1M tokens (input/output) | ~1/5th of SWE-2 |
| Base model | Post-trained from Kimi K3 | Adapted from Inkling-Small (Thinking Machines) |
| Weights | Not public | Not public |
| Available in | Devin (Desktop, CLI, Web, Fusion) | Devin Desktop + Devin CLI |
One honest note: Cognition's own speed illustration is a simulation — each model's Terminal-Bench run replayed at 200 vs. 60 tokens per second, with transcripts shortened to equal length. That's marketing visualization, not a measured benchmark. Independent tests will tell the fuller story, and they're worth waiting for before you bet a production workflow on the 3x figure.
Benchmarks: Where It Wins (and Doesn't)
Cognition reports that Flash delivers substantial gains over its Inkling-Small base on the evaluations it publishes — the post-training clearly added real capability. Against its bigger sibling, the picture is mixed, which is exactly what you'd expect from a cost-optimized model:
- FrontierCode 2.0: Flash edges out SWE-2 — a genuinely interesting result for a model that costs a fifth as much.
- Terminal-Bench 4.0: SWE-2 scores higher. The big model still wins on long, complex terminal tasks.
- DeepSWE 1.1: SWE-2 scores higher here too.
Cognition also claims Flash outperforms all American open-weight models on the evaluations it reports, including models roughly three times larger. The important framing — and Cognition deserves credit for being upfront about it — is that Flash is a cost-and-speed alternative, not an across-the-board replacement for SWE-2. It's aimed at automations, code review, feature development, and documentation updates: the high-volume, low-drama work that eats most engineering budgets.
As with every vendor-reported benchmark (a rule we apply to every model we cover — see our AI coding assistants roundup), treat these as the company's best-case numbers until third-party evaluations land.
Pricing: The One-Fifth Promise, in Numbers
Cognition hasn't published a separate per-token price table for Flash — it positions pricing strictly relative to SWE-2. SWE-2's Devin list price is $3.00 per million input tokens and $15.00 per million output tokens (with $0.30 per million cached-input tokens). One-fifth of that works out to roughly $0.60 input / $3.00 output per million tokens.
But here's the number that matters most right now: Flash is free in Devin Desktop and the Devin CLI through November 30, 2026. That puts it squarely in the same "try before the meter starts" pattern Cognition used for SWE-2 itself — and in the same spirit as the open-weight cost revolution we covered when Revolut cut its AI coding costs ~90% by self-hosting an open model.
The catch: Flash runs only inside Devin. There's no standalone API, no model router, no self-hosting. You're buying into Cognition's platform to get the discount — a very different proposition from open-weight models you can run anywhere, like DeepSeek's open-source Super-Agent.
Who Should Use It
Flash is for you if: you live in Devin Desktop or the CLI; your agent spend is dominated by routine work (code review, docs, small features, automations); you want the free window to test an agentic workflow before committing budget; or you're building demos where latency matters more than peak reasoning.
Stick with SWE-2 (or a frontier model) if: your work is dominated by long-horizon, complex terminal tasks — that's where SWE-2's Terminal-Bench and DeepSWE advantage shows; you need the Devin Web or Fusion surfaces; or you want to evaluate before the December billing reality sets in.
How to Try It Free
- Install or update Devin Desktop (or the Devin CLI).
- Select SWE-2 Flash in the model picker — it's live now, no waitlist.
- Point it at the boring work first: a code review pass, a docs refresh, a small feature branch. That's what it's built for.
- Watch your usage before November 30, 2026, when the free period ends.
Devin's entry plan starts around $20/month, and free models don't count against your usage allowance — so the free window is genuinely free for testing.
Verdict: Fast, Cheap, and Honestly Positioned
SWE-2 Flash is one of the more honest model launches of 2026. Cognition isn't claiming frontier dominance — it's claiming that most coding work doesn't need the frontier, and pricing accordingly. The 3x-speed, 1/5th-price framing, the free window through November 30, and the clear "routine work, not everything" positioning all point the same direction: this is the model you reach for when the task is ordinary.
The caveats are real: vendor-reported benchmarks, a simulated speed graphic, no public weights, and lock-in to Devin. But the thesis — that a 276B MoE with 12B active parameters can handle the daily grind — is the same efficiency story reshaping the whole industry, from GPT-6.1 Terra's cheap computer use to open-weight self-hosting. Try it free while it costs nothing, and decide with your own usage data.
FAQ
Is SWE-2 Flash really free?
Yes — Cognition offers free use of SWE-2 Flash in Devin Desktop and the Devin CLI through November 30, 2026. After that, it moves to paid pricing at roughly one-fifth of SWE-2's rate.
Can I use SWE-2 Flash outside Devin?
No. There's no standalone API and the weights aren't public. Flash runs exclusively in Devin Desktop and the Devin CLI.
Is SWE-2 Flash better than SWE-2?
Not overall. It's about 3x faster and much cheaper, and it edges SWE-2 on FrontierCode 2.0 — but SWE-2 scores higher on Terminal-Bench 4.0 and DeepSWE 1.1. Flash is the everyday-work model; SWE-2 is the flagship for hard tasks.
Is SWE-2 Flash open source?
No. The weights are not being shared publicly; Cognition cites a secure sharing program for trusted partners and research labs instead. The base model it was adapted from — Thinking Machines Lab's Inkling-Small — did have its weights released publicly in July 2026.
No comments:
Post a Comment