Quick Answer
Cognition's SWE-2 Flash, a coding model built on Thinking Machines' Inkling-Small, is free to use in Devin Desktop and the Devin CLI through November 30, 2026. Cognition says it runs three times faster than SWE-2 at one fifth of the price, but it has published no per-token price, no API, and no public weights.
Key takeaways:
- SWE-2 Flash launched November 4, 2026, built on Inkling-Small with 276 billion total parameters and 12 billion active per token.
- It is free in Devin Desktop and the Devin CLI through November 30, 2026, then billing starts on December 1.
- Cognition claims three times the speed and one fifth the price of SWE-2, but both figures are ratios, not rates, and no Flash price per token has been published.
- The weights are not public. Cognition cites cybersecurity risk after targeted training lifted its ExploitBench score from 11.3 percent to 34.8 percent.
- There is no API, no model ID, and no published context window, so you can only reach it through Devin's own Desktop app or CLI.
What is SWE-2 Flash?
SWE-2 Flash is Cognition's new fast coding model, launched November 4, 2026. Cognition, the company behind the Devin coding agent, calls it "a fully American coding model that sets a new bar for intelligence at low cost and high speed." It is built on Inkling-Small, the mixture-of-experts base model from Thinking Machines Lab, with 276 billion total parameters and 12 billion active per token.
The interesting part is how it was trained. For the first time in Cognition's recipe, a mid-training stage was added before reinforcement learning. Cognition says that stage alone lifted the base model on Terminal-Bench 4.0 from 0.9 percent to 20.0 percent, and on FrontierCode 2.0 from 3.2 percent to 24.0 percent. Those are Cognition's own figures, not independent measurements, and Cognition does not state its evaluation harness.
How do you try it free before November 30?
SWE-2 Flash is available today in Devin Desktop and the Devin CLI. Cognition's launch page carries the banner "Free through 11/30. Try it today in Devin Desktop and CLI." A second Cognition post says the free use covers Pro, Max, and Teams users.
That wording matters. The free window covers model use, not a free Devin subscription. Devin's pricing page lists a Free plan at 0 dollars, Pro at 20 dollars a month, and Max at 200 dollars a month. One independent review notes that Devin's own docs still list SWE-1.6 as the model for Free tier users, so whether the Free plan reaches SWE-2 Flash is not confirmed. If you want to be sure you can use it this month, budget on Pro at 20 dollars a month and treat the free model window as a bonus on top of a paid plan, not as a free subscription.
To try it, open Devin Desktop or the Devin CLI, pick SWE-2 Flash as the model for a run, and give it routine work. There is no waitlist and no separate signup for the model itself.
What does 3x faster and 5x cheaper actually mean?
Cognition's headline numbers are ratios, not rates. The company has not published a per-token price for SWE-2 Flash, so nobody can quote what a task costs in dollars. Here is the honest version of the comparison.
| SWE-2 | SWE-2 Flash | |
|---|---|---|
| Total parameters | 2.8 trillion | 276 billion |
| Active parameters per token | 104 billion | 12 billion |
| Speed claim | Baseline | 3x faster, 200 tokens per second vs 60 in a replay |
| Price claim | Baseline, list 3.00 dollars input and 15.00 dollars output per 1M tokens | One fifth of SWE-2, no per-token figure published |
| Free window | None | Free in Devin Desktop and CLI through Nov 30, 2026 |
| Weights | Not public | Not public |
| Access | Devin products | Devin Desktop and CLI only |
The speed figure needs a caveat. Cognition's illustration replays each model's own run on one Terminal-Bench task at simulated rates of 200 tokens per second for Flash and 60 for SWE-2, with transcripts shortened to equal length. That is a demonstration of the claimed gap, not a measured throughput benchmark. Treat 3x as Cognition's claim until independent timings exist.
The price figure has the same shape. One fifth of SWE-2's list price would work out to about 0.60 dollars input and 3.00 dollars output per million tokens if the ratio applied per token, but Cognition has not said it does, and has not published a Flash rate card. Do not quote 0.60 and 3.00 dollars as Flash prices. They are our arithmetic on SWE-2's list prices, not a published price.
What can it do, and what is it not for?
Cognition positions Flash for routine software agent work: automations, code review, small features, and documentation updates. It is the cheaper, faster sibling, not the flagship. SWE-2, the 2.8 trillion parameter model released September 10, 2026, remains the model for the hardest work.
Cognition's own benchmark table puts Flash ahead of every American open-weight model it lists, including models three times larger. The headline number is Terminal-Bench 2.1, where Cognition lists SWE-2 Flash at 89.2 against Ripple-65B Preview at 80.5. Both numbers are self-reported by different labs on different harnesses, so the 8.7 point gap is a vendor claim, not an independent result. Cognition's table also shows Claude Haiku 5.5 beating Flash on seven of eight rows, which is worth knowing before you pick a model on benchmarks alone.
Why are the weights locked?
This is the part of the launch that matters most. Inkling-Small, the base model, is open weights under Apache 2.0. SWE-2 Flash is not. Cognition says it is "not sharing the weights publicly yet" and is building a program to share them securely with trusted industry partners and research labs, with details promised in a few weeks.
The reason is cybersecurity. Cognition aimed part of the training at cyber defense work, and the model's ExploitBench score rose from 11.3 percent to 34.8 percent. The company says the same skill helps attackers, and points to a recent surge in cyberattacks described in its October security post. So the model that is good at finding security holes is kept behind Devin's products, where Cognition controls how it is used.
For builders, that means two things. First, you cannot download Flash, self-host it, or call it from your own code. There is no endpoint, no model identifier, and no rate card. Second, if your workflow depends on open weights, Flash is not your model, and the base Inkling-Small is the open alternative.
What is still missing?
Cognition has not published a per-token price for Flash, a context window, a model card, or an API. The Devin docs still name SWE-2 as the latest model in places, and Flash does not appear on the Devin Desktop or CLI models pages yet. The free window ends November 30, 2026, and billing starts December 1, but the price after that date is not established in any material checked for this article.
That is the honest state of the launch. A fast, cheap coding model with a free month is real. A price you can budget on is not published yet. Use the free window to run Flash against your real tickets and measure whether routine work moves off the expensive model. That measurement is worth more than any vendor ratio.
Who should try it, and who should skip it?
Try it this month if you already pay for Devin, or if you were considering Pro at 20 dollars a month anyway. Run your routine tickets on Flash, compare the results with SWE-2 on the same tasks, and keep notes. If Flash handles the routine work, the one-fifth price ratio could cut a real cost from December onward, once the price is published.
Skip it for now if you need an API, open weights, a published context window, or a fixed price before you commit. None of those exist yet. Also skip it if your work is the hardest class of coding tasks. Cognition itself frames Flash for routine work, and the flagship SWE-2 is the model for the rest.
Sources: Cognition, Introducing SWE-2 Flash, November 4, 2026; Cognition on X, SWE-2 Flash announcement, November 4, 2026; Devin pricing page as checked November 7, 2026; Unite.AI, Cognition Launches SWE-2 Flash, November 4, 2026; The Vibe Log, SWE-2 Flash 2026, November 2026; Anchor Terminal, SWE-2 Flash review, November 2026.
Related reading: AI Basics & Tutorials, DeepSeek Super-Agent Is Free and Open Source, Kimi K3.1 Pro Costs $2 for 1 Million Tokens
Frequently Asked Questions
Is SWE-2 Flash really free?
It is free to use in Devin Desktop and the Devin CLI through November 30, 2026. The free window covers model use, not a free Devin subscription, and Cognition's own post says it applies to Pro, Max, and Teams users. Whether the Free plan reaches it is not confirmed.
How much will SWE-2 Flash cost after November 30?
Cognition has not published a per-token price. It says Flash costs one fifth of SWE-2, whose list price is 3.00 dollars input and 15.00 dollars output per million tokens, but no Flash rate card exists yet, so do not treat the implied figures as a price.
Can I download SWE-2 Flash or call it from my own code?
No. The weights are not public, there is no API, no model ID, and no endpoint. It runs only inside Devin Desktop and the Devin CLI.
What is SWE-2 Flash built on?
It is post-trained from Thinking Machines' Inkling-Small, a 276 billion total parameter mixture-of-experts model with 12 billion active parameters, released under Apache 2.0 in July 2026.
Is SWE-2 Flash faster than SWE-2?
Cognition claims three times the speed, illustrated by a replay at 200 tokens per second against 60 for SWE-2. That is a demonstration, not a measured throughput benchmark, and no independent timing has been published yet.
Why is Cognition not releasing the weights?
Cognition says targeted cybersecurity training lifted Flash's ExploitBench score from 11.3 percent to 34.8 percent, and the same skill could help attackers. It is building a vetted sharing program for partners and research labs instead.



