On 4 November 2026, Cognition launched SWE-2 Flash, a smaller, cheaper coding model for its Devin agent. The pitch is simple: three times the speed of the flagship SWE-2 model at one-fifth of the price, free for paid Devin users through 30 November 2026. The catch is just as simple: you cannot download it. Cognition is keeping the trained weights private, citing cybersecurity risk, even though the base model it was built from is openly available. This guide explains what SWE-2 Flash is, how it was trained, what its benchmarks actually show, what it costs inside Devin, and who should — and should not — switch to it.

Table of Contents
- At a Glance
- What Is It?
- How It Was Trained
- Benchmarks: The Full Table
- Flash vs SWE-2
- Pricing and the Free Trial
- Why the Weights Stay Private
- Vs the Competition
- Who Should Use It
- How to Try It in Devin
- The Honest Limits
- What Happens Next
- FAQ
- The Bottom Line
SWE-2 Flash at a Glance
SWE-2 Flash is the most interesting kind of model launch: a cheaper, faster workhorse rather than a new frontier flagship. Cognition did not try to beat its own biggest model on every test. It built a smaller model that handles the large share of everyday engineering work — automations, code review, feature development, documentation updates — at a fraction of the cost and latency. Here is the short version, with every figure tied to what Cognition actually published.
- What happened: On 4 November 2026, Cognition released SWE-2 Flash inside Devin Desktop and the Devin command-line interface (CLI).
- What it is built on: Thinking Machines’ Inkling-Small, a mixture-of-experts model with 276 billion total parameters and 12 billion active parameters. The base model’s weights are public under Apache 2.0; the trained SWE-2 Flash weights are not.
- Speed and price: Cognition says SWE-2 Flash runs three times faster than SWE-2 at one-fifth of the price inside Devin.
- Free trial: Free for paid Devin users (Pro, Max and Teams plans) through 30 November 2026.
- Headline benchmarks (Cognition-reported): Terminal-Bench 4.0 rose from 0.9% for the base model to 20.0% for SWE-2 Flash; FrontierCode 2.0 rose from 3.2% to 24.0%, narrowly ahead of SWE-2’s 23.8%.
- The one benchmark where it trails: On Terminal-Bench 4.0, SWE-2 Flash scores 20.0% against SWE-2’s 27.3%. The cheaper model does not match its bigger sibling everywhere.
- Why you cannot download it: Targeted cybersecurity training lifted its ExploitBench score from 11.3% to 34.8%. Cognition says the same skills could help attackers, so it is withholding the weights and building a sharing programme for trusted partners and research labs instead.
- What it is not: SWE-2 Flash is not open-weight, not downloadable, and has no public API or published per-token price. The only way to use it today is inside Devin.
That list is the whole announcement, stripped of spin. The rest of this article is about what the numbers mean, where the claims need careful reading, and what the launch says about where coding agents are heading.
What Is SWE-2 Flash?
SWE-2 Flash is Cognition’s second model in the SWE-2 family, and it exists for a different reason than the first. SWE-2, which shipped on 10 September 2026, is the flagship: a large model post-trained from Moonshot’s Kimi K3, aimed at the hardest agentic coding work. Flash is the everyday model. It is built to handle the large volume of routine engineering tasks that do not need the flagship’s full capability — and to do them faster and cheaper, so that running an agent all day stops being an expensive habit.
Cognition’s own framing makes the point directly. In its launch post, the company calls SWE-2 Flash “a fully American coding model that sets a new bar for intelligence at low cost and high speed.” That is the company’s language, and it is worth reading carefully: the claim is about intelligence at low cost and high speed, not about intelligence in absolute terms. On the hardest tests, the bigger SWE-2 still wins. The argument for Flash is that most real work does not need the hardest test to be passed — it needs a capable model that answers quickly and costs little.
What “Flash” Means Inside Devin
Inside Devin, SWE-2 Flash is a model choice, not a separate product. Devin is Cognition’s autonomous software engineer: you assign it work through Slack, Linear, Jira, the web app, the desktop app or the CLI, and it plans, writes, runs and tests code in its own cloud environment before opening a pull request. Flash now appears as an option in Devin Desktop and the Devin CLI, alongside the larger SWE-2. You pick the model when you start a session, the same way you might pick a faster or slower engine for a job.
Cognition positions Flash for four kinds of work: automations, code review, feature development and updating documentation. Those four cover most of what a development team actually asks an agent to do on a normal day. The pattern is familiar from the rest of the industry — every major lab now ships a “flash”, “mini” or “lite” model next to its flagship — but the Devin angle is different. Devin is not a chatbot you prompt once; it is an agent that runs for minutes or hours, making dozens or hundreds of model calls per task. In that setting, a model that is three times faster and five times cheaper changes the economics of the whole product, not just the price of one answer.

Built on Inkling-Small, Not Built From Scratch
Flash was not trained from nothing. Cognition started from Inkling-Small, a model from Thinking Machines Lab that was released on 30 July 2026 with its full weights published under the Apache 2.0 licence. Inkling-Small is a mixture-of-experts model with 276 billion total parameters, of which only about 12 billion are active for any given token. Each token is routed through a small set of specialist components plus shared components, rather than activating the whole model — the design that lets a large model run at a fraction of the cost of a dense model of the same size.
The base model is capable in its own right. Thinking Machines reported 80.2% on SWE-Bench Verified for Inkling-Small, and a context window of up to 1 million tokens. What Cognition added was the post-training: a new mid-training stage, followed by large-scale reinforcement learning, tuned specifically for the work Devin does. The result is a model that is far better than its base at agentic coding, while keeping the base model’s efficiency. Our guide to Inkling 2 Small, Thinking Machines’ newer model, explains how the base-model family has developed since.
The distinction between the base model and the finished model is the most important fact in this whole story, and the easiest to get wrong. Inkling-Small is open-weight: anyone can download it. Flash is not. Cognition’s training turned a public base into a private product, and the company has decided, for now, that the private product stays private. That decision is the subject of its own section below, because it is where the launch gets genuinely controversial.
How SWE-2 Flash Was Trained
The most interesting part of the SWE-2 Flash launch is not the model itself but the training recipe behind it. Cognition says that for the first time, it added a mid-training stage before reinforcement learning. Mid-training sits between the base model’s original pre-training and the final reinforcement-learning phase: a large additional training run on carefully chosen data, designed to teach the model the knowledge and skills it will need before it is tuned with task feedback. Cognition describes the result as significant gains over the base model on every benchmark it measured — and the numbers support that description, at least on the tests the company chose to publish.
The mid-training dataset was assembled from a few trillion tokens of high-quality scientific articles, blogs, code and other sources. Choosing the mix was, by Cognition’s own account, the hard part. The team built a high-quality held-out coding dataset that reflects model intelligence in real-world software engineering, then used “token efficiency plots” to compare candidate data mixtures: for each mixture, they measured how many tokens it needed to reach the same validation loss as a baseline run. The mixtures that reached the target loss with the fewest tokens won. Training then proceeded in three stages, so the model saw the scarcest, highest-quality data last. Cognition reports that held-out validation loss improved at a stable log-linear rate through each stage — the steady, predictable improvement that training teams like to see.
After mid-training came large-scale reinforcement learning, where the model is trained through feedback on real tasks. Cognition says it made some tweaks to the reinforcement-learning recipe it used for SWE-2, but that most of the improvement came from the quality and diversity of the training tasks rather than from recipe changes. On quality, the team carefully audited the instructions, the environment setup and the verifiers for its tasks. On diversity, it curated several new task types covering a far wider range of capabilities than before, aiming to build the general reasoning and problem-solving skills that real work inside Devin demands. The company attributes the benchmark gains to that combination: better data before reinforcement learning, and better tasks during it.
There is also a security thread running through the training. Cognition targeted cybersecurity work specifically, and it says that training lifted the model’s ExploitBench score from 11.3% to 34.8%. ExploitBench measures a model’s ability to find and exploit software vulnerabilities — a skill that is plainly useful for defensive security work, and equally plainly useful for attackers. That dual-use character is the reason the weights are being withheld, and it is also the reason the training story matters beyond the benchmark table. Flash is not just a cheaper coding model; it is a model that was deliberately taught security skills, and then deliberately kept out of public hands.

SWE-2 Flash Benchmarks: The Full Table
Every benchmark number in this section comes from Cognition’s own launch post and benchmark table. None of it has been independently verified, and the tests were chosen by the company that built the model. That does not make the numbers useless — vendor benchmarks are still the main public evidence for a new model — but it does mean they should be read as Cognition’s best case, not as a neutral measurement. With that caveat stated, here is the full picture.
| Benchmark | Inkling-Small (base) | SWE-2 Flash | SWE-2 (flagship) |
|---|---|---|---|
| Terminal-Bench 4.0 | 0.9% | 20.0% | 27.3% |
| FrontierCode 2.0 | 3.2% | 24.0% | 23.8% |
| SWE-Bench Pro v2 Hard | — | 89.9% | 90.6% |
| Terminal-Bench 2.1 | — | 89.2 | — |
| ExploitBench | 11.3% | 34.8% | — |
Three things stand out in that table. First, the training worked: Terminal-Bench 4.0 rose from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%, which is a genuine transformation of the base model into an agentic coding model. Second, SWE-2 Flash is genuinely close to the flagship on several tests — it edges SWE-2 on FrontierCode 2.0 (24.0% against 23.8%) and sits within one point of it on SWE-Bench Pro v2 Hard (89.9% against 90.6%). Third, it is not close everywhere: on Terminal-Bench 4.0, the flagship’s 27.3% is more than a third higher than Flash’s 20.0%, and that gap is the honest price of the smaller, cheaper model.
Cognition also published comparisons against rival models on FrontierCode 2.0, where SWE-2 Flash at 24.0% sits ahead of OpenAI’s GPT-6 Luna at 21% and behind Anthropic’s Claude Haiku 5.5 at 27%, according to the figures reported by independent coverage of the launch. Those cross-vendor comparisons deserve extra caution: different labs run different harnesses, and a point or two either way may reflect the test setup as much as the model. The safer reading is the shape of the result rather than the exact ranking — Flash is competitive with the current generation of efficient coding models, and it is not the best model on any single test.
On Terminal-Bench 2.1, Cognition reports 89.2 for SWE-2 Flash. That number became a small story of its own on launch day, because a rival lab, Ricursive Intelligence, launched its Ripple-65B Preview model about ninety minutes earlier and reported 80.5 on the same test. Cognition’s launch post says Flash “outperforms all American open-weight models, even models that are 3x larger” — a claim aimed squarely at that comparison. The two scores come from two self-reported tables produced by different teams, so the 8.7-point gap is suggestive rather than decisive. What is not in dispute is the difference in access: Ripple’s weights are public on Hugging Face under Apache 2.0, while Flash’s are not.
Reading the Speed Claim Carefully
The headline speed figure — three times faster than SWE-2 — needs a careful reading, because the way Cognition demonstrated it is unusual. The company showed transcripts from each model’s run on the same Terminal-Bench 4.0 task, replayed at a simulated 200 tokens per second for Flash and 60 tokens per second for SWE-2, with the transcripts shortened to equal length. That is an illustration of the claimed output-speed difference, not a measurement of how long each complete task took. It shows what a three-times speed gap would look like; it does not prove the gap by timing real work.
That does not mean the speed claim is false. A smaller model with only 12 billion active parameters should, in principle, generate tokens faster than a much larger flagship, and the one-fifth price claim points in the same direction. But until independent users run SWE-2 Flash on their own tasks and time it, the “3x faster” figure should be treated as Cognition’s claim, not as a measured fact. The same applies to the price ratio: Cognition has published no per-token price for Flash, so the one-fifth figure cannot be checked against a price list. Inside Devin, where usage is metered through plan quotas rather than per-token billing, the practical meaning of “one-fifth the price” is that Flash consumes quota more slowly — which is real money for heavy users, even if the exact arithmetic is not public.
SWE-2 Flash vs SWE-2: Speed, Price, and Capability
The choice between Flash and SWE-2 is the choice the launch is really about. Cognition is not asking users to replace the flagship; it is asking them to stop using the flagship for work that does not need it. The table below puts the trade-off in one place.
| SWE-2 Flash | SWE-2 | |
|---|---|---|
| Launched | 4 November 2026 | 10 September 2026 |
| Built from | Thinking Machines Inkling-Small | Moonshot Kimi K3 |
| Size | 276B total / 12B active parameters | Not publicly disclosed |
| Speed (Cognition’s claim) | 3x faster than SWE-2 | Baseline |
| Price (Cognition’s claim) | One-fifth of SWE-2 | Baseline |
| Terminal-Bench 4.0 | 20.0% | 27.3% |
| FrontierCode 2.0 | 24.0% | 23.8% |
| SWE-Bench Pro v2 Hard | 89.9% | 90.6% |
| Weights | Not public | Not public |
| Available in | Devin Desktop and CLI | Devin (all surfaces) |
| Free trial | Free through 30 November 2026 for paid plans | — |
The pattern is clear: SWE-2 Flash is near-parity with the flagship on two of the three shared benchmarks and clearly behind on the third. For code review, documentation updates, routine feature work and automations, that gap is unlikely to matter — and the speed and price advantages will matter a great deal. For the hardest agentic tasks, where the model must plan across many files, recover from its own mistakes and keep a long task on track, the flagship’s Terminal-Bench 4.0 advantage is the number to watch. A sensible default is to run Flash first and escalate to SWE-2 when a task stalls, rather than paying flagship prices for every session.
There is a second comparison worth making, with the model SWE-2 Flash replaced rather than the one it sits beside. The previous generation of efficient coding models — DeepSeek’s V4.1 Flash, MiniMax’s M3.1 Flash and similar — made the same promise: most of the capability at a fraction of the cost. Flash follows that pattern inside a single product, Devin, rather than as a standalone API model. That is both its strength and its limit: the economics are excellent for Devin users, and irrelevant to everyone else, because there is no other way to reach the model.
SWE-2 Flash Pricing and the Free Trial
SWE-2 Flash has no price of its own. There is no per-token rate, no API listing and no standalone subscription. It is priced through Devin, and Devin’s pricing is what determines what the model actually costs you. That is an important difference from most model launches, where the headline is a dollars-per-million-tokens figure you can compare directly. Here the comparison is indirect: Flash costs one-fifth as much as SWE-2 inside Devin, and Devin itself is sold as a quota-based subscription.
Devin’s self-serve plans, as listed on its pricing page and verified by independent pricing trackers in October 2026, are straightforward. The Free plan costs $0 and gives a light quota with limited model availability. Pro costs $20 per month for one developer, with daily and weekly quotas, full frontier-model access and Devin Cloud sessions. Max costs $200 per month for heavy individual use, with a much larger weekly quota and no daily cap. Teams starts at $80 per month plus $40 per full developer seat, for up to 200 users, with shared credits, centralised billing and an admin dashboard. Enterprise pricing is custom-quoted and billed in Agent Compute Units under an order form. The old $500-per-month Team plan and the public per-ACU rate card were retired in April 2026, so pricing articles that still quote them are out of date.
| Devin plan | Price (USD) | Who it fits |
|---|---|---|
| Free | $0 | Trying Devin on small tasks |
| Pro | $20 / month | One developer using Devin regularly |
| Max | $200 / month | Heavy individual use, no daily cap |
| Teams | $80 / month + $40 per developer seat | Teams up to 200 users |
| Enterprise | Custom | VPC or dedicated deployment |
The free trial is the part that matters most this month. Cognition says SWE-2 Flash is free through 30 November 2026 for paid Devin users on Pro, Max and Teams plans. That is a genuine free window, not a discount: during November, Flash sessions do not consume the quota you would otherwise spend on SWE-2. For a Pro user paying $20 a month, that means a month of faster, cheaper agent sessions at no extra cost — a good reason to try the model on real work now, while the trial is live, and to measure for yourself whether the speed claim holds up on your codebase.
Two cautions belong here. First, the trial ends on 30 November, and Cognition has not said what Flash will cost afterwards beyond the one-fifth ratio. If your workflow becomes dependent on the model in November, check the price before December or you may find your quota disappearing faster than you expect. Second, because Devin meters usage through quotas that refresh daily and weekly, with overage billed at API list prices, the real cost of a task depends on how long Devin works, not just on which model you pick. A cheap model that runs for an hour can still cost more than an expensive model that finishes in ten minutes. Speed and price have to be judged together, on completed work, not on tokens alone.
Why the SWE-2 Flash Weights Stay Private
The most argued-about part of the SWE-2 Flash launch has nothing to do with speed or price. It is the decision to keep the weights private. Cognition is not sharing the trained model’s weights publicly, and it is unusually open about why. During training, the company targeted cybersecurity work, and the model’s ExploitBench score rose from 11.3% to 34.8%. ExploitBench tests the ability to find and exploit software vulnerabilities. Cognition says those capabilities should help defenders — the security teams who need to find flaws before attackers do — but acknowledges that the same skills could be used by malicious attackers, pointing to a recent surge in cyberattacks discussed in its October security post. The weights are therefore not being released publicly yet.
This is a real tension, not a marketing excuse. The base model, Inkling-Small, is openly available under Apache 2.0, and anyone can download it and fine-tune it themselves. What Cognition is withholding is the trained model: the version that went through its mid-training and reinforcement-learning stages, including the security-focused training. The distinction matters because it draws a line the industry has been arguing about all year. Open weights let researchers inspect a model, test it for flaws and run it on their own infrastructure. Closed weights let the developer control who gets a capability that could be misused. Cognition has chosen control, at least for now.
The company’s proposed compromise is a programme to share the weights securely with trusted industry partners and research labs, with more detail promised in a few weeks. The stated goal is to allow transparent access to the weights, including the ability to fine-tune the model and run it in sensitive environments such as on-premises infrastructure, without a public release. That is a narrower form of openness than the open-weight community wants, and a broader one than a purely closed model offers. Whether it satisfies anyone will depend on who gets in, what they are allowed to do with the weights, and how quickly the programme actually opens. Until then, the only confirmed way to use SWE-2 Flash is through Devin Desktop and the Devin CLI.
For readers, the practical consequence is simple: do not plan a workflow that depends on downloading or self-hosting Flash. If you need a downloadable coding model, the base Inkling-Small weights are public, as are rival open-weight models such as Ripple-65B Preview, which launched the same day with its weights on Hugging Face. What you cannot get anywhere else is Cognition’s trained version — and that is exactly the point of the company’s decision. If you want to evaluate the model, do it inside Devin during the free trial, and judge it on what it does, not on what you might one day be able to download.

SWE-2 Flash vs the Competition
SWE-2 Flash launched into a crowded week. The same day, Ricursive Intelligence released Ripple-65B Preview, an open-weight coding model that claims a new state of the art for United States open-weight models on agentic coding tasks. The two launches frame the central argument in coding models right now: is the future open weights that anyone can run, or capable models that only their maker can serve? Flash is the clearest example of the second answer, and Ripple is the clearest example of the first.
On the numbers the two labs chose to publish, SWE-2 Flash leads: 89.2 against 80.5 on Terminal-Bench 2.1, in Cognition’s and Ricursive’s own tables. But the comparison is not as clean as the headline suggests. The two scores come from different teams running different harnesses, and only Ripple’s weights are public, so only Ripple’s score can be checked by anyone else. Flash also has about four times Ripple’s 65 billion total parameters, which makes Cognition’s claim of beating “models 3x larger” read oddly — the size comparison runs the other way. The honest summary is that Flash looks stronger on the one shared test, and that the test itself is not enough to settle the question.
The wider competitive picture is a three-way race between efficient models. OpenAI’s GPT-6 Luna, Anthropic’s Claude Haiku 5.5 and SWE-2 Flash all chase the same job: a fast, cheap model that handles everyday coding work while a flagship handles the hardest tasks. On FrontierCode 2.0, the reported order is Haiku 5.5 at 27%, Flash at 24.0% and GPT-6 Luna at 21% — but again, these are vendor-reported figures from different harnesses, and the gap between first and third is small enough that the test setup could explain it. The more useful comparison is availability: GPT-6 Luna and Haiku 5.5 are available through APIs and products anyone can buy, while Flash is available only inside Devin. For a developer choosing a coding assistant, that difference matters more than a few points on a benchmark.
There is also the open-weight cluster that SWE-2 Flash deliberately stands apart from: DeepSeek’s V4.1 Flash, MiniMax’s M3.1 Flash, and the base Inkling-Small itself. Those models can be downloaded, self-hosted and fine-tuned, and their prices are public — DeepSeek, for example, cut its V4 Pro input price by 28% to $0.21 per million tokens in the same month Flash launched. Flash cannot compete on openness or on transparent pricing, because it offers neither. It competes on integration: it is the fast, cheap model inside the agent that many teams already use, available in one click, free for a month, with no setup at all. That is a genuine advantage, and it is also the whole of the advantage. If you are not a Devin user, Flash is not a product you can buy.
For readers weighing the alternatives, our best AI coding assistant 2026 roundup compares GitHub Copilot, Cursor, Claude Code, ChatGPT and the other major options on price, free tier and agent mode, so you can see where a Devin-centred choice fits. And if you are following the Codex side of the same market, today’s OpenAI Pierre acquisition story explains how Codex is moving from writing code toward owning the storage and review of the code its agents produce.
Who Should Use SWE-2 Flash
SWE-2 Flash is a good fit for some developers and a poor fit for others, and the line between the two is mostly about where you already work. The model is at its best when it is doing the routine work that fills a normal engineering day, inside a tool you already pay for. It is at its worst when you need it to be something it is not: a downloadable model, an API, or a general-purpose chatbot.
Use Flash if you:
- Already pay for Devin. If you are on Pro, Max or Teams, the model is free through 30 November 2026 and costs a fifth of SWE-2 afterwards. There is no reason not to try it on your next few tasks.
- Run agents all day. Teams that keep Devin sessions running for hours — triaging issues, reviewing pull requests, updating documentation — will feel the speed and price difference most, because the savings multiply across every session.
- Do mostly routine engineering work. Code review, small feature work, test updates, documentation and automations are exactly the jobs Cognition built SWE-2 Flash for, and the benchmark gap with the flagship is smallest on this kind of work.
- Care about response time. If you watch Devin work and wait for it to finish, a model that generates three times faster shortens every session, even if the final result is the same.
Skip SWE-2 Flash, or keep SWE-2 as your default, if you:
- Do not use Devin. There is no other way to reach the model. No API, no download, no standalone product. If you work in Cursor, Claude Code, Codex or GitHub Copilot, Flash is not an option for you at all.
- Work on the hardest agentic tasks. Large refactors, multi-service changes and long debugging sessions are where the flagship’s Terminal-Bench 4.0 advantage (27.3% against 20.0%) is most likely to show. Use SWE-2 for those, and Flash for the rest.
- Need to self-host or fine-tune. If your employer requires models to run on your own infrastructure, or you need to fine-tune a model on your own code, Flash cannot do it. The base Inkling-Small weights are public, and rival open-weight models exist, but Cognition’s trained version is not available for either purpose today.
- Need a published price. If you have to put a per-token cost in a budget or a procurement document, SWE-2 Flash has no number to quote. The one-fifth ratio is relative to SWE-2, and SWE-2’s own price inside Devin is expressed through quotas, not a price list.
A practical pattern that several early users have described is the two-model workflow: start every task on Flash, and escalate to SWE-2 only when the smaller model stalls, loops, or produces work that fails its own tests. That pattern uses the cheap model for the bulk of the work and reserves the expensive model for the moments when capability, not speed, is the constraint. It is also the pattern Cognition’s pricing encourages, whether the company intended it or not.
How to Try SWE-2 Flash in Devin
Trying SWE-2 Flash takes minutes, because there is nothing to install beyond Devin itself. The model is available today in Devin Desktop and the Devin CLI, and it is free for paid users through 30 November 2026. Here is the process, step by step.
- Open Devin Desktop or the Devin CLI. Flash is available in both. If you only use the Devin web app, switch to Desktop or CLI for this model — the launch announcement names those two surfaces specifically.
- Check your plan. The free trial covers paid plans: Pro at $20 a month, Max at $200 a month, and Teams at $80 a month plus $40 per developer seat. Free-plan users have a light quota and limited model availability, so the model may not appear for them.
- Select Flash when you start a session. Pick the model the way you would pick any other model in Devin, then assign the task as normal — through a prompt, or from Slack, Linear or Jira if your workspace is connected.
- Start with a routine task. A code review, a documentation update, a small feature or a test fix is the right first test. That is the work the model was built for, and it is where you will learn fastest whether the speed difference is real on your codebase.
- Compare like with like. Run the same kind of task on Flash and on SWE-2, and note three things: how long each took, how much quota each consumed, and whether the result passed its tests without extra fixes. Those three numbers are the real benchmark, and they are the only ones that matter for your budget.
- Decide before 30 November. The trial ends on that date. If SWE-2 Flash becomes part of your daily workflow, check what it will cost under your plan before December, so the first paid month does not surprise you.

If you are new to Devin entirely, the Free plan at $0 is the way to start, but be aware that the free tier has limited model availability and may not include SWE-2 Flash. The realistic entry point for the trial is Pro at $20 a month, which is also the cheapest plan that gives full frontier-model access and Devin Cloud sessions. That is a low price for a month of testing — and it is the price Cognition set when it rebuilt Devin’s pricing in April 2026, dropping the entry point from the old $500-a-month Team plan to $20.
The Honest Limits of SWE-2 Flash
Every launch post reads like a victory lap, and this article would be misleading if it stopped at Cognition’s claims. Flash has real limits, and several of them are visible in the company’s own material if you read it carefully.
The benchmarks are vendor-reported. Every score in this article comes from Cognition’s launch post or its benchmark table. No independent lab has re-run these tests, and the tests were chosen by the company that built the model. The Terminal-Bench 2.1 comparison with Ripple is a good example of the problem: two labs, two harnesses, two self-reported tables, and a headline gap that may or may not survive independent testing. Treat the numbers as a claim to be checked, not as a settled result.
The speed figure is a demonstration, not a measurement. The 200-versus-60 tokens-per-second replay is a simulation of a speed gap, shown on one task, with transcripts shortened to equal length. It is not a timed measurement of complete tasks. The model is probably faster than SWE-2 — a 12-billion-active-parameter model should be — but “three times faster” is Cognition’s number, and it has not been verified by timing real work.
There is no public price. “One-fifth the price” is a ratio, not a price. Without a per-token figure, you cannot compare SWE-2 Flash with DeepSeek V4.1 Flash, MiniMax M3.1 Flash or any other model on cost. Inside Devin the ratio has a practical meaning through quotas, but outside Devin it has none at all.
There is no API and no download. This is the biggest limit, and it is a deliberate choice rather than an oversight. Flash exists only inside Devin Desktop and the Devin CLI. Developers who want to call a model from their own code, run it on their own hardware, or fine-tune it on their own data cannot do any of those things with this model. The weights-sharing programme may change that for a small group of partners and researchers, but it is not a public release, and its details are still weeks away.
The security training cuts both ways. Cognition is open about the reason for withholding the weights: the same ExploitBench improvement that helps defenders could help attackers. That is an honest statement, and it is also an admission that the model is more capable at finding vulnerabilities than its base model was. Users who adopt SWE-2 Flash for defensive security work should know that the capability exists, and that it exists because Cognition deliberately taught it.
The trial ends soon. Free through 30 November 2026 sounds generous, and it is — but it is also less than four weeks from launch. Teams that build a workflow around Flash in November need a plan for December, and Cognition has not published what the model will cost them then.
What Happens Next for SWE-2 Flash
The next few weeks will answer the two open questions about SWE-2 Flash. The first is the weights-sharing programme. Cognition has promised more detail “in a few weeks” about how trusted industry partners and research labs will get access to the weights for inspection, fine-tuning and on-premises deployment. Watch for who is included, what they are allowed to do, and whether the programme is a genuine path to transparency or a narrow list of favoured partners. The second is the price after 30 November, when the free trial ends and Flash starts consuming quota at its one-fifth rate. Watch for whether Cognition publishes a per-token price at that point, or keeps the cost expressed only through Devin quotas.
Beyond those two questions, the launch points to a broader shift. Cognition’s own post notes that smaller models are increasingly able to handle a significant share of the work software agents do, and that the company remains focused on lowering the cost of everyday engineering work alongside pushing the capability frontier. That is the same shift the whole industry is making: flagship models for the hardest problems, efficient models for everything else, and the economics of agent work decided by the efficient models rather than the flagships. Flash is Cognition’s entry in that race, and its success will be measured less by its benchmark table than by how many Devin users make it their default.
There is also the question of what Thinking Machines does next. Inkling-Small was released in July 2026, and its successor, Inkling 2 Small, arrived in November — covered in our Inkling 2 Small guide. If Cognition’s post-training recipe works as well on the newer base as it did on Inkling-Small, a second-generation Flash model is a plausible next step. For now, though, SWE-2 Flash is the model that exists, it is free for paid Devin users until the end of November, and the best way to judge it is to run it on your own work.
SWE-2 Flash FAQ
What is SWE-2 Flash?
Flash is Cognition’s cheaper, faster coding model for Devin, launched on 4 November 2026. It is built on Thinking Machines’ Inkling-Small and runs at three times the speed of SWE-2 at one-fifth of the price, according to Cognition. It is available in Devin Desktop and the Devin CLI.
Is SWE-2 Flash free?
SWE-2 Flash is free through 30 November 2026 for paid Devin users on Pro ($20 a month), Max ($200 a month) and Teams ($80 a month plus $40 per seat). After that date, it consumes Devin quota at one-fifth the rate of SWE-2, according to Cognition. There is no standalone free version of the model itself.
Is SWE-2 Flash open source?
No. SWE-2 Flash is not open-weight and cannot be downloaded. Only the base model, Thinking Machines’ Inkling-Small, is public under Apache 2.0. Cognition is withholding the trained Flash weights over cybersecurity concerns and is building a sharing programme for trusted partners and research labs instead.
How fast is SWE-2 Flash compared to SWE-2?
Cognition claims SWE-2 Flash is three times faster than SWE-2. The demonstration was a replay of one task at simulated rates of 200 versus 60 tokens per second, so the figure is the company’s claim rather than an independently measured result. The model has 12 billion active parameters, against a much larger flagship, so a real speed advantage is plausible.
Can I use SWE-2 Flash through an API?
No. There is no public API for Flash, no published per-token price and no model endpoint to call. The only way to use it today is inside Devin Desktop or the Devin CLI.
How good is SWE-2 Flash on benchmarks?
On Cognition’s own reported figures, SWE-2 Flash scores 20.0% on Terminal-Bench 4.0 (against 27.3% for SWE-2), 24.0% on FrontierCode 2.0 (against 23.8% for SWE-2) and 89.9% on SWE-Bench Pro v2 Hard (against 90.6% for SWE-2). All figures are vendor-reported and have not been independently verified.
Why are the SWE-2 Flash weights not public?
Cognition says targeted cybersecurity training raised the model’s ExploitBench score from 11.3% to 34.8%, and that the same vulnerability-finding skills could be used by attackers. The company is building a programme to share the weights securely with trusted industry partners and research labs, with more detail promised in a few weeks.
Should I switch from SWE-2 to SWE-2 Flash?
For routine work — code review, documentation, small features, automations — yes, try SWE-2 Flash first, especially while it is free through 30 November. For the hardest agentic tasks, keep SWE-2 as the default, since it leads on Terminal-Bench 4.0 by a clear margin. A two-model workflow, starting on Flash and escalating when a task stalls, is the sensible default.
The Bottom Line
SWE-2 Flash is a useful model with an honest trade-off at its centre. It is faster and cheaper than the flagship, close to it on several benchmarks, and free for paid Devin users until 30 November 2026. It is also closed: no download, no API, no published price, and a speed claim that has been demonstrated rather than measured. For Devin users, the decision is easy — try it now, on real work, while it costs nothing, and judge it by whether your tasks finish faster and your quota lasts longer. For everyone else, Flash is a reminder of the direction the coding-model market is moving: the future belongs to efficient models that do everyday work cheaply, and the argument is no longer about which model is smartest, but about who gets to run it.
If you use Devin, open Devin Desktop or the CLI, pick SWE-2 Flash for your next routine task, and see for yourself. The trial window is short, the model is genuinely different from the base it was built on, and the only benchmark that finally matters is the one you run on your own code.
