Cognition's SWE-2 is 64% cheaper than a number it published for someone else — and it loses the one benchmark Devin is sold for

A real model, priced in a currency nobody published

Start with the part that is not in dispute. SWE-2 is a substantial piece of work and a bigger jump than SWE-1.7 was in July.

Cognition took Kimi K3, Moonshot’s 2.8-trillion-parameter open-weights base that had already been reinforcement-trained for agentic coding, and pushed its own RL stack on top — the first time it has done that at multi-trillion scale. The result adds five to six points over the base on many benchmarks, which suggests the K3 frontier was not exhausted by its original training. SWE-2 is also Cognition’s first model with configurable reasoning effort at medium, high and max, and the interesting engineering claim is that all three points on the cost-performance curve came out of one reinforcement learning run shaped by a cost penalty, rather than three separately trained variants.

Against its own predecessor the numbers are emphatic: SWE-2 medium matches or beats SWE-1.7 on FrontierCode while using 58% fewer turns and costing 81% less on average. Fewer turns is the honest efficiency metric for an agent, and it moved a long way.

Then the launch reaches the price, and the price is not there.

The only dollar figures are someone else’s

Cognition’s claim is that SWE-2 hits 50.0% on FrontierCode 1.1 Main — within a point of Claude Fable 5.1 at 50.9% — while costing 64% less to run at that score. To make that concrete, it publishes what the comparison costs: Fable 5.1 Medium at $3.28 per FrontierCode task, Fable 5.1 Max at $12.83.

It publishes nothing equivalent for SWE-2. No per-token rate. No per-task figure. No standalone API. No model card, and no stated context window.

You can back the number out — 64% below $3.28 implies roughly $1.18 per task — and that arithmetic is fine as far as it goes. It is an inference from a percentage, and it is not a price you can put in a budget, because SWE-2 is not sold as a model. It is a capability inside Devin, and Devin bills in ACUs — Agentic Computing Units, Cognition’s normalised bundle of VM time, model inference and networking, historically around fifteen minutes of autonomous work apiece.

That gap matters more than it first appears. A cheaper model reduces the inference share of an ACU. It does not automatically reduce how many ACUs your task consumes, and the ACU is what the invoice counts. Cognition has reported an internal unit-cost improvement — genuine, and good engineering — and has not said how much of it reaches a customer. That is the same shape of problem as Cursor’s always-on agent billing, where the spend controller sits on the vendor’s side of the meter.

The row it loses is the row Devin is sold for

Three of those four rows support the near-frontier story. The fourth does not, and it is not a random miss.

Terminal-Bench 4 is the harder successor benchmark, built around long-horizon, multi-step terminal work. SWE-2 scores less than half what either frontier model manages. The shape across the whole table is consistent: strong to excellent on well-scoped tasks, sharply weaker as the horizon lengthens.

That shape is awkward for this specific product. Devin’s pitch has always been asynchronous delegation — hand it a multi-hour task and walk away — which is exactly the regime Terminal-Bench 4 probes. And it interacts badly with ACU billing, because a failed long run still burns ACUs. A model that is 64% cheaper per attempt and needs two or three attempts on your hardest tickets is not cheaper. The cost claim and the weak row are measuring the same system from opposite ends, and only one of them made the headline.

To be fair to Cognition: SWE-1.7 scored 7.6% on the same benchmark. 27.3% is a 3.6x improvement. The trajectory is right; the level is not yet competitive.

Vendor-run, unreplicated, best-of-effort

No third party has reproduced any row of that table.

Two methodology notes belong next to it. Cognition states it reports the best score across reasoning-effort settings for every model. Applied evenly, that is fair — and it means the table compares best cases, while your team will run defaults. Second, each model was scored in its native harness (Claude Code for Anthropic’s models, the Devin CLI for others), and where a public result already existed, Cognition used it instead of running its own. The announcement does not mark which cells are internal and which are published, which makes independent verification harder than it needs to be.

This is not an accusation of bad faith; it is the ordinary condition of vendor benchmarks, and the reason private test sets and repriced index runs exist. The same caution applies here that applied when evaluation gaming showed up in METR’s work on GPT-5.6 Sol: a launch table is a hypothesis about your workload, not a measurement of it.

The base model has a week it did not ask for

One coincidence is worth naming plainly rather than leaving for someone else to discover.

SWE-2 is post-trained from Kimi K3. On the same day SWE-2 launched, Anthropic published a threat report naming Moonshot, K3’s maker, among seven China-based labs it accuses of illicit distillation — including the allegation that Moonshot relayed live Kimi customer requests to Claude without telling customers.

Two things are true simultaneously, and conflating them would be wrong in both directions.

Devin’s data path is not implicated. Cognition serves SWE-2 itself. K3’s weights are published. Running published weights relays nothing to Moonshot or anyone else — that distinction is the whole point of open weights.

The provenance chain is now contested, and a commercial buyer at the end of it has a fair indemnity question: if a base model’s capabilities are later held to be the product of a breach of a third party’s terms, who covers the downstream user? Anthropic’s allegations are vendor-authored, unreplicated and undenied. The proportionate response is a contract clause and a written answer from Cognition, not a veto — the same conclusion this desk reached about licensed model-factory arrangements when continuity, not capability, was the exposure.

What to do with this

1. Pilot on the promotion, and measure ACUs per completed task. Not per attempt, not per benchmark point. Completed tasks are what you buy; everything else is what you are shown.
2. Split your evaluation by task horizon. Well-scoped ticket work is where SWE-2’s near-Fable scores are most credible. Multi-hour autonomous delegation is where Terminal-Bench 4 predicts trouble, and where retries convert a cheap model into an expensive one.
3. Run defaults, not best-of-effort. The launch table reports each model’s best effort setting. Your production config will not be that, for any vendor on the sheet.
4. Ask what happens when the free month ends. A model with no published price, inside a meter you do not control, is a renegotiation that has not happened yet. Get the post-promotional cost in writing before you move a team onto it.
5. Keep a frontier fallback wired in. If your hardest work is the long-horizon kind, the honest read of Cognition’s own numbers is to keep paying Fable 5.1 or GPT-6 Astra for it and route the rest to SWE-2.
6. File the provenance question. One email to Cognition about base-model indemnity, answered in writing, costs nothing and closes a gap your procurement team will otherwise find later.

For the wider field, the Devin review tracks where this sits in the agent stack, best AI coding tools covers the alternatives, and the Claude Code and Cursor reviews cover the two harnesses SWE-2 is priced against. Cognition’s own trajectory — a $26B valuation and an ARR line built on Devin — is the context for why it keeps training its own models rather than renting someone else’s.

Questions or corrections? Email Pick Right. Want the full list? See all news.