SWE-2 Flash Review: Cognition's 3x Faster Coding Model Is Free Until Nov 30

Cognition has a new coding model, and the headline writes itself: three times faster than SWE-2 at one-fifth of the price, free to use through November 30. SWE-2 Flash, announced November 4, 2026, is the company's smaller, cheaper sibling to its flagship coding model, built for the routine work that fills most of a developer's day — automations, code review, feature work, and documentation updates.

The catch is in the price itself. Cognition published a ratio, not a rate card. "One-fifth of the price" means very little when the price of SWE-2 was never published either, and SWE-2 Flash has no public API, no per-token price, and no downloadable weights. What Cognition did publish is a benchmark table — and read closely, that table is more honest than the headline. It shows a capable model that is a clear step down from its bigger sibling on long terminal tasks, and behind Claude Haiku 5.5 on seven of eight tests.

That does not make SWE-2 Flash a bad model. It makes it a model whose real value this month is simple: it is free inside a product you may already pay for. Here is what it is, what it costs, and where its own numbers say you should not use it.

What Is SWE-2 Flash?

SWE-2 Flash is a coding model Cognition post-trained from Inkling-Small, a 276-billion-parameter mixture-of-experts model with 12 billion active parameters from Thinking Machines Lab. The base model's weights are public; the Flash version's are not. Cognition added a mid-training stage before reinforcement learning for the first time in its recipe, training on a few trillion tokens of scientific articles, blogs, and code in three stages, then running large-scale reinforcement learning on a wider set of task types than before.

The training gains over the base model are large. Cognition reports Terminal-Bench 4.0 rising from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%. That is a real jump, and it is the strongest evidence in the launch post that Cognition's recipe works. The model is available now in Devin Desktop and the Devin CLI, with Devin Web and Fusion rolling out, according to launch coverage. There is no model ID for third-party tools and no listing on model routers, so the only way to reach it is through Devin's own apps.

For scale, SWE-2 itself launched September 10, 2026, as a 2.8-trillion-parameter model with 104 billion active parameters, post-trained from Moonshot's Kimi K3. Flash is roughly a tenth of its size. Cognition is building the two-tier lineup every lab now builds: the big model for hard tickets, the small one for cheap, frequent ones.

How Much Does SWE-2 Flash Cost?

There is no per-token price for SWE-2 Flash. Cognition has published no input or output rate, no context-window length, and no model card. The free period is the only firm number: the model is free to use in Devin Desktop and the CLI through November 30, 2026, for customers on paid Devin plans, according to launch reporting. Cognition's own post says only "Free through 11/30."

What you can price is the Devin plan around it. Devin's current pricing page lists a Free plan at $0 with a light quota and limited model availability, Pro at $20 a month, Max at $200 a month, and Teams at $80 a month plus $40 a month per full developer seat. Enterprise is quote-only. Each paid plan includes a usage allowance that refreshes daily and weekly, and extra usage is billed at API pricing. At November 6, 2026 market rates (1 USD = £0.75 and C$1.41), Pro works out to about £15 or C$28 a month, Max to about £150 or C$283, and Teams to about £60 or C$113 plus roughly £30 or C$57 per seat. These are approximations; Devin bills in US dollars.

The honest way to read "one-fifth the price" is that it is one-fifth of an unpublished number. Cognition has not published a price for SWE-2 either; its September post reported average spend per rollout instead of a rate card. Until Cognition publishes a real per-token price after the free period ends, SWE-2 Flash is a good free trial inside a paid product, not a model you can plan a budget around. That is the decision to watch in December.

The Benchmarks: What Cognition's Own Table Actually Shows

Every number below is vendor-reported by Cognition in its launch post, with no independent evaluation yet. The post does not name the harness or say whether scores are pass@1, a gap other analysts have flagged. Independent cross-checks by The Vibelog found the rival figures consistent with what those vendors published themselves.

Count the rows and the story changes. Claude Haiku 5.5 beats SWE-2 Flash on seven of eight benchmarks in Cognition's own table, including the four above and the four remaining rows: FrontierCode 1.1 Main (46.4 to 45.5), Terminal-Bench 2.1 (94.4 to 89.2), SWE-Atlas QA (56.5 to 50.0), and ExploitBench (38.0 to 34.8). Flash's single win is DeepSWE 1.1, by 0.8 points.

Terminal-Bench 4.0 is the row to weigh most for an agent that runs shell commands all day, and it is Flash's weakest result. At 20.0% it is roughly half of Haiku 5.5's 39.2% and below SWE-2's 27.3%. DeepSeek V4.1 Flash 1017 leads the row at 42.3%, with GLM 5.3 Flash at 41.8%. On long terminal tasks, SWE-2 Flash is a clear step down from the models it is being compared with. It does edge its bigger sibling on FrontierCode 2.0, 24.0% to 23.8%, and it stays close on SWE-Bench Pro v2 Hard, 89.9% to SWE-2's 90.6%. The gap opens on DeepSWE 1.1, where SWE-2 scores 73.0% against Flash's 60.1%.

The 3x Speed Claim, Read Carefully

Cognition says Flash runs at 200 tokens per second against 60 for SWE-2. The figure comes from a replay, not a live timing. Cognition replayed each model's own run of one Terminal-Bench 4.0 task, gsea-proteomics, at simulated speeds of 200 and 60 tokens per second, with the transcripts shortened to equal length.

That shows how the speed difference feels. It does not prove the production endpoint holds 200 tokens per second under load, and Cognition did not publish time-to-first-token. Treat "3x faster" as a design target until someone outside Cognition times it. The speed is real in the sense that a smaller model with 12 billion active parameters will answer faster than a 104-billion-active-parameter sibling; the exact multiple is the company's own staging.

The "Beats American Open-Weight Models" Claim

Cognition's launch post calls SWE-2 Flash "a fully American coding model that sets a new bar for intelligence at low cost and high speed," and says it "outperforms all American open-weight models, even models that are 3x larger." The claim is narrower than it sounds.

The American open-weight models in Cognition's table are Inkling-Small, Reflection Beam 1.0, and Meta's Muse Spark Open. Flash beats Beam on every row and beats Muse Spark Open on most rows, including Terminal-Bench 4.0 by 12 points. It loses to Muse Spark Open on SWE-Atlas QA, 50.0 to 54.0, and wins DeepSWE by 0.1 points, 60.1 to 60.0.

Two things make the headline less useful than it looks. First, SWE-2 Flash is not itself an open-weight model, so it is being compared against models you can download while you cannot download it. Second, the strongest cheap models in the same table are DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash, which are open-weight but not American. Both beat SWE-2 Flash on Terminal-Bench 4.0 by more than 20 points. If you care where the weights come from for compliance reasons, the claim matters. If you care which cheap model clears the most tickets, it mostly does not.

Why the Weights Stay Private

Cognition is not releasing SWE-2 Flash weights publicly, and it gives a specific reason: the model was deliberately trained on cybersecurity work, raising its ExploitBench score from 11.3% to 34.8%. The same capabilities can help attackers, the company says, citing a recent surge in cyber attacks. Instead of a public release, Cognition says it is building a program to securely share the weights with trusted industry partners and research labs, with more details promised in a few weeks. No terms, licence, or application process has been published yet.

The reasoning is understandable, and it is also convenient: a model you can only run inside Devin is a reason to buy Devin. Enterprises and governments that need to host a model on their own infrastructure, including on-premises, are the audience for the partner program. Until it exists, SWE-2 Flash is a Devin feature, not a model in the open ecosystem. The base model, Inkling-Small, remains downloadable — but that is a different, far weaker model, scoring 0.9% on Terminal-Bench 4.0 against Flash's 20.0%.

How to Try SWE-2 Flash

Open Devin Desktop or the Devin CLI on a paid Devin plan and select SWE-2 Flash in the model picker. Cognition positions it for routine agent work: automations, code review, feature development, and documentation updates. It is free through November 30, 2026, after which Cognition has not said what it will cost.

Two practical notes. First, the Free plan's limited model availability means the free tier is not the way in; the trial sits on top of a paid plan, starting at $20 a month for Pro. Second, keep the model on the work it is tuned for. Short, well-scoped tickets with clear success criteria are where a cheap, fast model earns its keep. Long chains of shell commands, migrations, and anything that needs a dozen debugging loops are where the Terminal-Bench gap bites, and that is where SWE-2 or a frontier model belongs.

SWE-2 Flash vs Claude Haiku 5.5: Which Should You Use?

If you already pay for Devin, use SWE-2 Flash for routine tickets this month. It is free through November 30, it is the model Cognition tuned for its own agent harness, and the target list of automations, code review, feature work, and documentation updates is exactly the work a cheap model should take.

If you do not use Devin, nothing here changes your stack. Claude Haiku 5.5 scores higher in Cognition's own table, runs in Claude Code as the haiku alias, and costs $0.10 per million input tokens and $0.50 per million output tokens on prompts up to 100K. DeepSeek V4.1 Flash 1017 is cheaper still on paper at $0.15 input and $0.60 output per million tokens off-peak, doubling at peak hours, with a public API. Both are models you can call directly and budget for. SWE-2 Flash is neither, yet.

Common Mistakes

Budgeting around "one-fifth." A ratio without a base price is not a price. Do not forecast December spend on SWE-2 Flash until Cognition publishes a per-token rate. The free period ends November 30, and the price after that is unpublished.

Using it for long terminal chains. The 20.0% Terminal-Bench 4.0 score is the honest signal in the launch post. Multi-step migrations and debugging loops belong on SWE-2 or a frontier model.

Expecting the weights soon. The partner program is "a few weeks" away with no terms published. If your workflow needs a downloadable model, the open-weight options in Cognition's own table — DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash — are the ones you can actually host.

Confusing Flash with its base model. Inkling-Small is open and downloadable, but it scores 0.9% on Terminal-Bench 4.0. The capability in SWE-2 Flash comes from Cognition's mid-training and reinforcement learning, and that part stays inside Devin.

Frequently Asked Questions

What is SWE-2 Flash?

SWE-2 Flash is a coding model from Cognition, the company behind Devin, released on November 4, 2026. It is post-trained from Thinking Machines Lab's Inkling-Small, a 276-billion-parameter mixture-of-experts model with 12 billion active parameters. Cognition says it runs about three times faster than SWE-2 at one-fifth of the price. It is available only inside Devin Desktop and the Devin CLI.

How much does SWE-2 Flash cost?

SWE-2 Flash is free for paid Devin users through November 30, 2026. Cognition has not published a per-token price or a price after the trial, only that it costs one-fifth as much as SWE-2, whose price is also unpublished. Devin itself costs $20 a month for Pro, $200 for Max, and $80 a month plus $40 per developer seat for Teams.

Is SWE-2 Flash better than Claude Haiku 5.5?

Not on Cognition's own numbers. In the SWE-2 Flash launch table, Claude Haiku 5.5 scores higher on seven of eight benchmarks, including Terminal-Bench 4.0 at 39.2% versus 20.0%. SWE-2 Flash only leads on DeepSWE 1.1, 60.1% to 59.3%. Haiku 5.5 also has a public API at $0.10 per million input tokens, while SWE-2 Flash is locked inside Devin.

Can I download the SWE-2 Flash weights?

No. Cognition is not releasing SWE-2 Flash weights publicly, citing a recent surge in cyber attacks and the model's cybersecurity training, which raised its ExploitBench score from 11.3% to 34.8%. The company says it is building a program to share weights with trusted industry partners and research labs, with more details promised in a few weeks. The base model, Inkling-Small, has public weights.

How do I use SWE-2 Flash?

Open Devin Desktop or the Devin CLI on a paid Devin plan and select SWE-2 Flash in the model picker. Cognition positions it for routine agent work such as automations, code review, feature work, and documentation updates. There is no model ID for third-party tools, no API endpoint, and it is not available on model routers.

What is SWE-2 Flash's context window?

Cognition has not published one. The launch post gives no context length, no output limit, and no model card for SWE-2 Flash. The base model, Inkling-Small, is listed with a context window of up to 1 million tokens by Thinking Machines Lab, but that figure describes the base model, not the version Cognition ships inside Devin.

SWE-2 Flash is worth trying this month if you already pay for Devin, and worth waiting on if you do not. The free period makes the trial free; the missing price makes the commitment impossible to judge. When Cognition publishes a real rate card in December, the comparison with Haiku 5.5 and DeepSeek V4.1 Flash becomes possible — and that is the number to watch. If you want a cheap coding model you can call and budget for today, read our Claude Haiku 5.5 breakdown and our DeepSeek V4.1 Flash 1017 review first.