Cognition SWE-2 Flash: The Discount Is Real, the Base Model Is the News

SWE-2 Flash shipped on 4 November 2026 with two numbers stapled to it: one-fifth the price and three times the speed of SWE-2, the 2.8-trillion-parameter model Cognition released eight weeks earlier. Both are ratios rather than rates, and both are measured inside Devin. The sentence that actually changes something is easier to skim past: this model is a post-train of Inkling-Small, the Apache-2.0 base from Mira Murati's Thinking Machines Lab. That is the first time one of Cognition's coding models has not descended from a Chinese open-weight base, and it is the reason the other two numbers exist at all.

For anyone arriving from a search box rather than a newsletter: SWE-2 Flash is a coding model, it runs inside Cognition's agent products, and it is not something you can call from your own code. There is no endpoint, no model identifier, and no rate card. The launch post gives you a date, a base model, a training story, a table of numbers and a free window that ends on 30 November. Everything below is read off that post, Cognition's own table, and the publicly callable alternatives — with the sourcing attached to each figure.

What shipped, and the two words missing from the post

Cognition is explicit about the surface. SWE-2 Flash is available today in Devin Desktop and Devin CLI, and the post's closing banner reads "Free through 11/30 — Try it today in Devin Desktop and CLI." That is the whole distribution story. There is no API launch, no waitlist for an endpoint, and no Hugging Face repository.

The two words missing from the pricing claim are "per token". "At one-fifth the price and 3x the speed of its much larger sibling" is a statement about what a task costs inside Devin, which folds the model, the harness, the scaffolding and Cognition's serving stack into one invoice. It is a legitimate way to price an agent, and it is not a rate you can compare against anybody else's per-million-token card. A reader who tries to convert it into one will fail, because the post never says what the fifth is a fifth of.

The speed figure needs the same reading. Cognition's speed section says 200 tok/s against 60 tok/s for SWE-2, and then explains what it measured: "Each model's own run on Terminal-Bench 4.0's gsea-proteomics task, replayed at a simulated 200 vs. 60 tokens per second. Transcripts are shortened to equal length for comparison." That is a replay at stated rates, not a live timing of two endpoints under load. It tells you the intended serving profile. It does not tell you what you will see at 2pm on a Tuesday.

The base-model swap is the actual release

Seven weeks ago, SWE-2 was post-trained from Kimi K3 — Moonshot's 2.8-trillion-parameter mixture, 104B of it active per token. SWE-2 Flash is post-trained from Inkling-Small, 276B total and 12B active, released under Apache 2.0 in July. Cognition's own size chart puts the new model at 1x and SWE-2 at 10.1x, which is a polite way of drawing an eightfold gap in active parameters.

That swap is what makes the price claim mechanically plausible rather than promotional. Serving cost tracks active parameters and tokens generated far more than it tracks a benchmark score, and a 12B-active model is in a different serving class from a 104B-active one. You do not need to trust Cognition's efficiency narrative to accept that a model this much smaller can be sold at a fifth of the price. The interesting question is what it kept.

Cognition frames the swap nationally. The post's first line calls SWE-2 Flash "a fully American coding model", and it goes on to say the model "outperforms all American open-weight models, even models that are 3x larger." Both statements are true as written, and both are narrower than they sound. "Fully American" describes a supply chain — an American lab's base weights, post-trained by an American company — not a moat, because Inkling-Small is Apache 2.0 and downloadable by anyone, including the competitors the claim is aimed at. The part Cognition actually controls is the mid-train and the RL, and that is the part it is not shipping.

The contradiction sitting in the Origins & Cyber section

The most quotable passage in the release is not a benchmark. It is the section where Cognition explains why the weights are staying home.

It begins by describing a customer who needs more than "a capable model behind an API": enterprises and governments that need "observability to host and control deployments on their own infrastructure, including on-prem", and researchers who need to study how a model behaves. It then states the goal plainly — "One of our goals with SWE-2 Flash is to provide an option to industry partners and researchers that allows transparent access to weights along with the ability to fine-tune and run it in sensitive environments."

Two paragraphs later: "we are not sharing the weights publicly yet", with a partner program promised "in a few weeks", and the reason given is cybersecurity risk. Cognition deliberately trained for it. The same section says it "targeted cybersecurity work in our training, raising the model's ExploitBench score from 11.3% to 34.8%", then acknowledges that "these same capabilities can also be used by malicious attackers" and withholds the weights.

Read as policy, the logic is coherent but uncomfortable: the on-prem customer the section describes cannot have the model that was trained for them, and the stated blocker is a capability Cognition chose to add. Read as evidence about the model, there is a sharper detail in the table. SWE-2 Flash scores 34.8 on ExploitBench against SWE-2's 36.1. The model Cognition will not release is within 1.3 points of the flagship it already sells. If size were the safety argument, that number would look different.

The recipe changed too, and it is the part worth copying

Cognition says this is the first time it has added a mid-training stage before reinforcement learning. The stage used "a few trillion tokens" of high-quality scientific articles, blogs, code and other sources, and ran in three passes so that "the model would see the scarcest, highest-quality data later in training." The mixtures were chosen with token-efficiency experiments: candidate mixes were run against a held-out validation set and compared on how many tokens each needed to reach the same loss as a baseline.

Then came the RL pass. Cognition is unusually candid about where the gains came from: "While we made some tweaks to our RL recipe from SWE-2, most of the improvements came from the much higher quality and diversity of our RL tasks." It describes auditing instructions, environment setup and verifiers, and curating new task types across a wider range of capabilities. That is a data-and-verifier story, not a scale story, and it is the most transferable thing in the release.

The scale of the jump supports the reading. Against its own base, SWE-2 Flash moves Terminal-Bench 4.0 from 0.9 to 20.0 and FrontierCode 2.0 from 3.2 to 24.0. Whatever the post-training cost, it bought roughly twenty times the base model's score on one benchmark and seven times on another, on a 12B-active model. What Cognition does not publish is the token count for the mid-train beyond "a few trillion", the compute, or any ablation isolating the mid-train from the RL. The claim that the new stage caused the gain is Cognition's; the mechanism is not independently checkable.

The numbers, and where the headline claim gets thin

All of the following are Cognition's own, from its launch table, and none has been independently reproduced. FrontierCode is Cognition's benchmark, written with outside maintainers and graded by Cognition.

• FrontierCode 2.0 — SWE-2 Flash 24.0 vs SWE-2 23.8

• FrontierCode 1.1 Main — 45.5 vs 50.0

• DeepSWE 1.1 — 60.1 vs 73.0

• Terminal-Bench 4.0 — 20.0 vs 27.3

• Terminal-Bench 2.1 — 89.2 vs 92.8

• SWE-Atlas QA — 50.0 vs 56.3

• SWE-Bench Pro v2 Hard — 89.9 vs 90.6

• ExploitBench — 34.8 vs 36.1

One row is a win and seven are losses, most of them small. The Flash is level with the flagship on FrontierCode 2.0, within a point on SWE-Bench Pro v2 Hard, and thirteen points behind on DeepSWE 1.1. That shape — close on bounded tasks, further behind on long-horizon ones — is the most useful thing in the table, and it is the subject of our separate comparison of the two models.

The "3x larger" claim deserves its own paragraph because the margin is thinner than the sentence suggests. The American open-weight comparator that comes closest is Muse Spark Open, 616B total and 42B active. On DeepSWE 1.1 it scores 60.0 against SWE-2 Flash's 60.1. One-tenth of a point. The claim survives because the Flash's other margins are real — 20.0 against 8.0 on Terminal-Bench 4.0, 24.0 against 6.2 on FrontierCode 2.0 — but the headline "even models that are 3x larger" rests, on the row where that model is strongest, on a rounding error.

There is also a column in the same table that cuts the other way. DeepSeek V4.1 Flash 1017, at 763B total and 16B active, beats the Flash on four of the eight rows: 77.0 to 60.1 on DeepSWE 1.1, 71.0 to 50.0 on SWE-Atlas QA, 42.3 to 20.0 on Terminal-Bench 4.0, and 91.0 to 89.2 on Terminal-Bench 2.1. Two of those margins are larger than anything the Flash wins anywhere in the table. The "all American open-weight models" claim is scoped by nationality, and the strongest comparator in the table sits outside the scope.

What the missing harness line costs you

SWE-2's September launch came with a methodology appendix. It named the scaffold used for each vendor — Claude Code for Anthropic models, Codex for OpenAI models, Grok Build for xAI — and stated that every model's best score across reasoning-effort settings was reported. The Flash post contains none of that. No harness, no effort setting, no statement of whether the numbers are best-of-N, and no footnote explaining the table.

That omission matters more for a small model than a large one, because the Flash's argument is that post-training closed a gap. If the two Cognition models were run in different scaffolds, part of the 0.2-point FrontierCode difference is the harness. If SWE-2's row is its best effort setting and the Flash has one setting, the comparison is a best case against an unknown. Three disclosures would fix it: the harness per row, the effort setting per row, and one independent run of Terminal-Bench 4.0 by someone who is not Cognition.

What you can call this week

Nothing here. SWE-2 Flash has no endpoint, and no third-party platform carries it — there is no identifier to route. What you can call are the models it is being compared to, and the gap between the two situations is the practical story of the release.

DeepSeek V4 Flash is routed on OrcaRouter as deepseek/deepseek-v4-flash, with the 1M-token context and 384K maximum output that DeepSeek publishes and Cognition does not publish for the Flash. Kimi K3, the base SWE-2 was post-trained from, is routed on OrcaRouter at $3.00 per million input tokens and $15.00 per million output tokens as kimi/kimi-k3. One API covers more than 200 models at the provider's list price with no markup on top, so a vendor's price change reaches you the same day rather than on the router's schedule. If you are testing whether a 12B-active post-train is good enough for your queue, that test has to run against a 12B-active model you can actually call, with a fallback chain behind it so a trial does not become a production dependency.

The narrow reading of SWE-2 Flash is the one the evidence supports. A 12B-active model, post-trained on a mid-train and a better-audited RL set, lands within 4.5 points of a 104B-active flagship on FrontierCode 1.1 Main and level on FrontierCode 2.0, at a fifth of the price inside one product. If that holds up under independent testing, the lesson is not that small models caught up. It is that in coding agents the marginal return on parameters has fallen below the marginal return on verifier quality — and the companies that can generate and audit hard tasks, rather than the ones that can afford the biggest base, will set the pace. That is a bigger claim than a 0.2-point win, and it is the one worth watching for the next independent run.