Cognition announced SWE-2 Flash on 4 November 2026, a smaller, faster coding model for Devin that the company says runs three times faster and costs one-fifth as much as its larger sibling, SWE-2. It is available now in Devin Desktop and Devin CLI, and it is free for Devin users through 30 November 2026. Built for routine engineering work — automations, code review, feature development, and documentation updates — it is a workhorse model, not a flagship. For a business, that is the point: most of a backlog is routine work.

Flash is built on Thinking Machines Lab's Inkling-Small, with a mid-training stage added before reinforcement learning. The trained weights are not publicly shared, and no standalone token prices are published. This post explains what it can do, what the benchmarks mean, and how to trial it.

Cognition SWE-2 Flash: Faster, Cheaper Coding Agents for Everyday Dev Work

What SWE-2 Flash actually is

SWE-2 Flash is the cheaper, faster sibling of SWE-2, Cognition's most advanced coding model. Cognition says it runs at one-fifth the price and three times the speed, streaming at roughly 200 tokens per second against about 60 for the flagship in Cognition's own Terminal-Bench replay. Those are Cognition's claims, not independently verified figures.

The base model underneath

Flash is built on Inkling-Small, a sparse mixture-of-experts model from Thinking Machines Lab with 276 billion total parameters and 12 billion active per token. The base model's weights are released under Apache 2.0 — but that openness does not carry over to SWE-2 Flash. Cognition's trained weights are not publicly shared. Only 12 billion parameters fire per token, which is what makes the model fast and cheap to run — see the Inkling-Small model card for the architecture detail.

What the benchmarks say — and what they don't

These benchmarks are Cognition's own, from its announcement. Read them as the vendor's evidence, not independent verification.

• Terminal-Bench 4.0: Flash scores 20.0%, against 0.9% for the base model and 27.3% for SWE-2.
• FrontierCode 2.0: Flash scores 24.0%, against 3.2% for the base model and 23.8% for SWE-2.
• ExploitBench (cybersecurity): targeted security training lifted Flash from 11.3% to 34.8%.

Cognition also claims Flash beats all American open-weight models on leading evaluations, even models roughly three times larger. See our guide to the GDPx AI agent benchmark — even the best models clear only about 12% of long-horizon tasks.

Do not buy a model on a benchmark — buy it on a trial against your own tickets.

The business case: cost per task, not cost per token

Cognition has published no standalone token prices — only the relative claim of one-fifth the price of SWE-2 inside Devin. That still matters, because the cost on your bill is the cost per finished task — and a model three times faster at a fifth of the price changes the maths on routine work, provided rework does not eat the saving.

Treat it as a routing problem, not a replacement problem. The expensive model is for genuinely hard work: unfamiliar code, ambiguous requirements, multi-file refactors. The cheap model is for work that is well-specified and repeatable.

A simple routing rule

• Send to Flash: dependency bumps, lint and test fixes, documentation updates, small features, code review drafts, and boilerplate.
• Send to the flagship: architecture decisions, debugging you do not understand yet, and anything touching payments or customer data.
• Send to a human: anything where the requirement itself is unclear — no model fixes a vague ticket.

This is the same discipline behind good AI agent development and AI automation: the model is the engine, but routing, permissions, and review are what make it safe and economical. See our guides to custom software development cost in the UK and SaaS MVP development cost in 2026.

The free window to 30 November — use it as a trial

SWE-2 Flash is free for Devin users through 30 November 2026 — a short window, best used as a structured trial. Pick one project and one queue of tickets for two to three weeks, and measure tasks completed without rework, second attempts, and time to merged pull request.

If Flash clears the routine queue at a fifth of the cost with the same review burden, you have a routing decision. If review time balloons, the saving is an illusion. Ask Cognition for your expected rates before the window ends — after 30 November, Flash is priced at one-fifth of SWE-2 inside Devin, but without published token rates you cannot calculate an exact monthly bill. The same logic applied to Gemini 4 Argon's pricing deadline.

The closed-weights caveat: what you give up

Here is the trade-off. Inkling-Small's base weights are open under Apache 2.0, but SWE-2 Flash's trained weights are not publicly shared. Cognition says it is building a secure sharing programme for trusted partners and research labs — not a public release. You cannot self-host or fine-tune Flash.

For most businesses, that is an acceptable trade: you are buying a service inside Devin, not a model to run. But your Flash-based workflows live inside Cognition's platform, priced on Cognition's terms — vendor lock-in in the ordinary sense. It is the same question we raise in our open-weights guide for business.

Keep your prompts, playbooks, and review checklists in your own repository, not only inside Devin. Do not build a business-critical workflow that works with only one model. If you need a model you can run yourself, our guides to Kimi K3.1 Pro coding agents and the DeepSeek open-source coding agent cover open-weight alternatives.

The security question: a model trained for exploit work

The most unusual part of the announcement is the cybersecurity training. Cognition says it raised Flash's ExploitBench score from 11.3% to 34.8%, and it is withholding the weights because the same skills could help attackers.

Defensively, a model that understands exploits can spot vulnerable patterns in code review. But it gets access to your repositories, so apply the usual controls: least-privilege access, no production credentials in its environment, human review before merge, and logging.

Do not let the security training become a reason to skip review: a model better at finding vulnerabilities is also better at introducing subtle ones. See our guide to AI agents for small business automation and AI chatbot vs AI agent.

A practical adoption plan

Start small, measure honestly, and decide on evidence. A four-step plan fits in the free window.

1. Pick one queue. Choose a real backlog of routine tickets in one repository, with one owner.
2. Set the routing rule in writing. Flash handles the routine list; the flagship or a human handles the rest.
3. Keep review mandatory. Every Flash output gets human review before merge. Track review time, not just model time.
4. Decide on 30 November, on numbers. Compare tasks completed, rework rate, review time, and post-trial price. Keep Flash only if the numbers support it.

If the trial works, fold the routing into your normal delivery process — the same way a custom software team or a SaaS product team would standardise any tool. To scope a pilot, see our contact page and about page.

Frequently asked questions

What is Cognition SWE-2 Flash?

SWE-2 Flash is a fast, lower-cost coding model from Cognition, announced on 4 November 2026 and available in Devin Desktop and Devin CLI. Cognition says it runs three times faster at one-fifth the price of the larger SWE-2. See Cognition's announcement, Introducing SWE-2 Flash.

Is SWE-2 Flash free?

Free for Devin users through 30 November 2026. After that, Cognition says it costs one-fifth of SWE-2 inside Devin, but no standalone token prices have been published — confirm your expected rates before the window ends.

Can I self-host or download SWE-2 Flash?

No. The Inkling-Small base model's weights are open under Apache 2.0, but Cognition is not sharing SWE-2 Flash's trained weights publicly, citing cybersecurity risk. For now, access is through Devin Desktop and the Devin CLI only.

Is SWE-2 Flash better than SWE-2?

Not overall. On Cognition's own benchmarks, Flash scores 20.0% on Terminal-Bench 4.0 against 27.3% for SWE-2, but edges it on FrontierCode 2.0, 24.0% against 23.8%.

Should my business try SWE-2 Flash now?

If you already use Devin, yes — as a structured trial before 30 November. Route a defined queue of routine tickets to Flash, keep human review mandatory, measure rework and review time, and decide on the numbers. Keep your playbooks in your own repository.