Cognition SWE-1.7

Cognition is based on the software engineering agent model trained by Kimi K2.7 Code, approaching the cutting edge and greatly reducing costs.

Agent Vibe Coding software engineering agent

https://www.cognition.ai/

Visit Website

In-depth Report

Cognition (the parent company of Devin) released SWE-1.7 on July 8-9, 2026, which is the strongest software engineering agent model it has trained to date. It is based on the open source model Kimi K2.7 Code of Moonshot AI as the base, and is superimposed on Cognition's self-developed large-scale reinforcement learning post-training, and is directly tuned for long-period asynchronous tasks in the Devin environment. The official did not claim to completely surpass the strongest closed-source model, but put the selling point in the deployable range that is "several points worse and an order of magnitude cheaper." The cost of a single task on the self-built FrontierCode benchmark is about US$1.97. All benchmark numbers are from Cognition's own reports, and third-party independent retesting has not yet occurred.
Cognition is an applied AI lab best known for building its own software engineer, Devin. Its SWE model family is specially designed for agent software engineering work. It is delivered through Devin's web, desktop and CLI three terminals, and is optimized for long-span, asynchronous coding tasks. SWE-1.7 is the first version of the family to be officially documented. The previous version was SWE-1.6 released in October 2025 (focusing on model experience and speed, and was free for three months). The company's team is small but high-density. The founding members hold 10 IOI gold medals and come from Cursor, Scale AI, Modal, Google DeepMind and other institutions.
SWE-1.7 is only provided in a hosted form within the Devin platform. It does not release open source weights and has no independent model API. It is available on Devin three terminals, with computing power provided by AI chip company Cerebras, and inference speeds of up to 1,000 tokens per second. The training recipe has four key components: the first is entropy preservation, using top-p sampling plus sampling distribution replay to solve the KL divergence mismatch between training and inference, and cooperates with the Muon optimizer to eliminate trainer non-determinism; the second is multi-cluster training, spanning four data centers on three continents, the trainer is centered, the inference engine is distributed, the compression weight increment volume is reduced by more than 99%, 1-2 minutes of cross-continental synchronization, and inference faults are solved by NVIDIA Dynamo Re-routing; the third is high-quality data planning, automatically executing tests, filtering tasks with low learning signals, and preventing reward hackers; the fourth is long-span self-compaction (self-compaction). When the context is close to the limit, the model summarizes itself and continues running. A single rollout can last up to 6 hours, and it is combined with alternating length penalty to compress redundancy on simple tasks. In terms of behavior, compared to the Kimi K2.7 base, it will do more exploratory code reading before making changes, provide a more condensed reasoning chain, and perform root cause-level defect repairs.
SWE-1.7 is not sold separately, and its capabilities are provided with Devin subscriptions; some pricing clues come from the Windsurf Cascade documentation: SWE-1.7 Lightning is priced per million tokens at $2.50 for input, $1.00 for cached input, and $12.50 for output. Free version users can only use the previous generation SWE-1.6. For products like Devin that charge based on task consumption, the unit cost determines the ability to scale more than the top score. Cognition emphasizes the single task cost of $1.97 as the core business number.
As a model that has just been released and is only available within the Devin platform, there is still little public first-hand user feedback, and social media discussions are more focused on speed and positioning. Developer dabit commented that 1000 tok/s "gives the asynchronous agent work a sense of real-time", describing an intermediate state that is "technically asynchronous, but so fast that you are willing to stare at it"; he also pointed out that more reasoning leads to longer runs, but it will amplify the impact of changes. The generally recognized voice is that it moves the competitive point from the "top score" to the "cost-performance Pareto front", which is realistically attractive to teams that charge by task and pursue large-scale throughput.
Industry media generally accept its positioning as “close to the cutting edge, but an order of magnitude cheaper.” Multiple reviews pointed out that SWE-1.7 scored 42.3% on FrontierCode 1.1 Main, slightly lower than 43.0% of GPT-5.5 and 46.5% of Claude Opus 4.8, but much higher than the 30.1% of the base Kimi K2.7 Code and the 9.4% of the previous generation SWE-1.6; the score on Terminal-Bench 2.1 81.5% (86.9% for Opus 4.8); 77.8% on SWE-Bench Multilingual multi-language software engineering, exceeding GPT-5.5’s 76.8%. The conclusion is that it does not lead closed-source opponents in all dimensions, but it makes a difference in multi-language generalization and unit cost. Cognition also used this release to challenge a popular judgment: the benefits of an open source base that continues to do RL after sufficient post-training will quickly peak - and SWE-1.7 still achieves substantial gains over Kimi K2.7, which has undergone multiple rounds of RL.
The biggest controversy arises from the origin of the pedestal. Kimi K2.7 Code is an open source model from a laboratory in mainland China. Cognition specially published the article "Measuring the Trustworthiness of Open-Source-Derived Models" for this purpose, using a set of multi-language publicity and review evaluations (145 politically sensitive questions, 5 samples per question, covering English/Simplified/Traditional Chinese, scored according to six axes) to test the credibility of the model, and emphasized the selection of Kimi K2.7 Because it has both coding ability and neutrality, neutrality can be further improved in subsequent training. This approach is regarded by the industry as an honest operation that "puts geological risks on the table", but it also reminds buyers: the model is only provided through the Devin platform, which requires self-hosting, strict model routing, or access to a team with its own technology stack. The access method itself is a limitation. In addition, all benchmarks are self-reported by Cognition and without independent third-party retesting, the numbers should be viewed with caution.
It is suitable for teams that have already adopted or are preparing to adopt Devin, bill based on task consumption, and pursue large-scale software engineering throughput, especially in scenarios that are suppressed by the unit cost of cutting-edge models. It is not suitable for teams that require local self-hosting, controllable model weights, or access to their own application stacks - these users should consider open source weight routes such as GLM and Kimi. Alternatives include Anthropic’s Claude series, OpenAI’s GPT-5.5/Codex, and multi-model coding agents such as Cursor and Windsurf.
SWE-1.7 is a demonstration of Cognition's "open source base + self-developed RL + inference collaborative design": the score did not top the list, but it pushed the cost curve a big step downward to the left. For pay-per-task programming agency products, this may be more weighty than a few benchmark percentages; it is just that the self-reported numbers are tied to the platform and still need to be independently retested and tested over time.

User Reviews

CGonzalez_2024

—

Frankly speaking, I only care about one thing: whether the bugs it fixes are root cause level. The report says that it is better than the base for root cause repair and exploratory reading before making changes. This has a greater impact on daily experience than simple scores. In the past, when we used other agents, we often changed the appearance and tested it. As a result, the module next door exploded due to the same root cause. This is really a waste.

6coch1

—

The choice of Kimi K2.7 Code for the base is worth talking about in itself. The open source model can increase by more than ten points after being trained twice by a top RL team, indicating that the upper limit of RL has not been reached.

Emma.Martin_X

—

Multi-cluster RL's cross-continent weight incremental synchronization looks very cool, but we are more concerned about its sandbox design to prevent reward hackers. The report mentions network isolation and zero rewards for cheating, which is the right direction.

StarkNetStarReyes

—

In the terminal, you can watch the agent work in real time at 1000 tok/s. This asynchronous but fast "intermediate state" is much better than pure background running experience, and the anxiety of waiting is significantly reduced. In the past, if I lost a long task, I would switch to doing something else. When I came back and found that I had gone astray, I had to start over again; now I can see it reading the code and setting the probe in half a minute, and if something is not right, I can stop it on the spot.

KimberlyKristensen

—

It is a real requirement that self-compression can last up to 6 hours of rollout. When we used other agents to run long tasks, the context often exploded and the state was lost midway. This design is suitable for the problem.

Michael.ChavezSr474

—

The placement of this score is very smart: don’t grab the top spot, grab the cost curve. For a team like ours that charges by task and digests hundreds of PRs a month, the unit price is much more important than those few percentage points. If the cost of a FrontierCode task of US$1.97 is true, compared with the cost of each rollout like Opus 4.8, the gap will be exaggerated when scaled up.

琉璃448

—

Cognition's approach to publicity/censorship is quite honest, which is a plus compared to labs that pretend not to use Chinese open source bases.

SJacksonSr

—

All benchmarks are reported by Cognition themselves. FrontierCode is still their own benchmark. Wait and wait for third-party re-testing. Don’t rush to write it into the selection report. In particular, OpenAI itself has previously abandoned the Verified version due to SWE-bench pollution problems, which shows that there is a lot of room for self-reported lists to be watered down. It is more stable to wait for independent audits to come out before drawing conclusions.

Bobby.Johnson520278

—

There is no open source weight, no independent API, and we can only use Devin. This is a flaw for us. We need self-hosting and strict model routing, and platform binding is directly eliminated.

PatrickHendersonQ

—

The free file is still only for SWE-1.6. If you want to try 1.7, you have to pay. Let two people in the team subscribe and run real tasks for a week before deciding.

Brenda.Collins346

—

The figure of $1.97 per task is important, but it has to be compared with our own "single consolidated change cost". Just looking at the benchmark score is meaningless.

Jean_Hill_66

—

I switched to SWE-1.7 in Devin and ran the migration tasks of two old warehouses. It felt really fast, and the psychological burden of the inferior results of 1000 tok/s was much lighter.

LJonesX

—

SWE-Bench Multilingual 77.8% exceeds GPT-5.5's 76.8%. Multilingual generalization really makes a difference and is very useful for our cross-language warehouse.

Sawyer530

—

In comparison, I prefer to think of it as a "low-cost long-term task executor" rather than the "strongest coding model". If you think about the positioning clearly, you won't be obsessed with the top score.

BPerez_7

—

It jumps from 9.4% of SWE-1.6 to 42.3%, a four-fold increase in one generation. This range makes me more curious about the training formula than the slight gap between it and Opus.

Cognition SWE-1.7

In-depth Report

User Reviews

Similar Products