What Is Devin SWE 2? 50.0% FrontierCode
Sofia Marenco
Model Evaluation Lead

TLDR50.0% FrontierCode and 92.8% Terminal-Bench 2.1 explain Devin SWE 2, including its reported cost, access, and limitations.
Devin SWE 2 Is Devin's Coding Model — 50.0% FrontierCode at Reported 64% Lower Cost
Devin SWE 2 is Cognition’s coding model for the Devin software-engineering agent, post-trained from Kimi K3 to execute software tasks with a reported 50.0% FrontierCode 1.1 Main score. Cognition released it on September 10, 2026, with access first reported in Devin Desktop and Devin CLI, followed by rollout to Devin Web and Fusion. The model is proprietary, not open-weight, and has no confirmed public API or token rate card. Reports describe a Devin Pro plan priced at $20 per month and a limited promotion offering free or unlimited SWE-2 usage, but the promotion’s exact terms remain unconfirmed.
Key Takeaways
- Devin SWE 2 is Cognition’s successor to SWE-1.7 and is designed for coding-agent work inside Devin.
- The model is reported to use Kimi K3, a 2.8-trillion-parameter mixture-of-experts base with about 104 billion active parameters per token.
- Cognition-reported results include 50.0% on FrontierCode 1.1 Main and 92.8% on Terminal-Bench 2.1.
- Cognition reports 64% lower cost than Claude Fable 5.1 on one comparison and approximately one-quarter the cost of GPT-6 Astra.
- Early community testing reports a 262,000-token context window, but that specification is not confirmed.
- Access is centered on Devin Desktop, Devin CLI, Devin Web, and Fusion rather than downloadable weights or a standalone public API.
What Is Devin SWE 2?
Devin SWE 2 is a software-engineering model developed by Cognition for Devin. It is intended to explore repositories, plan changes, edit code, run terminal tasks, write tests, and complete longer agent workflows. The model is not positioned as a general-purpose chat model with an independent consumer endpoint.
The available release coverage describes SWE-2 as a post-trained version of Kimi K3. That base is reported as a 2.8-trillion-parameter mixture-of-experts model with approximately 104 billion active parameters per token. These architecture figures come from launch coverage and should be treated as vendor-reported specifications rather than independently audited measurements.
Cognition’s stated goal is cost-efficient coding-agent performance. The model is meant to reduce unnecessary turns, reach a substantive code edit sooner, and adjust effort for the difficulty of a task. Devin SWE 2 is therefore best understood as an agent component, not simply as a model leaderboard entry.
The release status is materially different from an early leak. SWE-2 has been reported as released in Devin Desktop and Devin CLI, with Web and Fusion rollout following. However, public reporting does not include a directly linked Cognition model card or public API reference, so some technical details remain provisional.
Devin SWE 2 at a Glance
| Specification | Current information |
|---|---|
| Developer | Cognition |
| Type | Coding model for the Devin software-engineering agent |
| Base model | Kimi K3, reported as a 2.8-trillion-parameter MoE model |
| Active parameters | Approximately 104 billion per token, based on launch coverage |
| Modality | Software-engineering model using text, code, and tool interactions; exact modality specification not yet confirmed |
| Context window | 262,000 tokens according to an early community report; unconfirmed |
| Reported release date | September 10, 2026 |
| Reported pricing | Devin Pro described at $20 per month; SWE-2 promotion terms are not yet confirmed |
| Public API | Not yet confirmed |
| Availability | Devin Desktop and CLI; Web and Fusion rollout reported |
| License | Proprietary; weights are not publicly available |
The most important qualification in this table is the difference between a reported capability and a published specification. The context window, promotional pricing, and several user-observed behaviors come from community posts. They should not be treated as stable product guarantees.
How Devin SWE 2 Works / What Makes It Different
The first differentiator is K3 Post-Training. Cognition reportedly applied its reinforcement-learning recipe to Kimi K3 instead of releasing the base model unchanged. Third-party launch coverage says the post-training added roughly 5–6 benchmark points across several evaluations, although the improvement has not been independently reproduced in public reporting so far.
The second is the Single-Run Effort Ladder. Release coverage describes three effort levels: medium, high, and max. The levels were reportedly trained in one reinforcement-learning run, allowing Devin to trade compute and persistence against speed without presenting them as entirely separate models.
The third is Move-Soon Behavior. Cognition reports that medium-effort SWE-2 runs reduced average FrontierCode steps from 127 steps to 53 steps. The first substantive edit reportedly arrived at a median of 18 steps, compared with 48 steps for SWE-1.7. Those numbers describe Cognition’s evaluation setup, not a universal result for every repository.
Devin SWE 2 is a cost-oriented coding-agent model whose main design target is useful progress with fewer unnecessary turns.
SWE-2 also fits into the Devin-Only Harness. The model’s planning, terminal execution, permissions, session state, and repository context are handled by Devin. This explains why comparing SWE-2 with a standalone chat model can be misleading: the user is evaluating a model-plus-agent system rather than only generated text.
Fusion adds a related Fusion Sidekick mode. Third-party coverage describes Fusion as a harness in which a frontier lead model handles planning and review while SWE-2 handles cheaper mechanical work. The two models reportedly maintain separate caches and exchange structured briefs, reducing the cost of switching models during a task. Fusion availability is separate from the initial Desktop and CLI release.
What You Can Do With Devin SWE 2
SWE-2 is aimed at practical repository work rather than isolated code completion. Suitable tasks include issue implementation, test scaffolding, repository exploration, bug fixes, refactoring, and terminal-driven setup. The model can also be used as a lower-cost executor under a stronger planning model in Fusion.
Early users have reported using Devin’s CLI to coordinate more than 10 agents across Linear and GitHub issues. That workflow is an individual report, not an official concurrency guarantee. Another user reported running 26 SWE-2 sub-agents under an unlimited plan, but the relationship between plan tier, parallel requests, and actual limits remains unresolved. The orchestration report is available here
The usage pattern can extend beyond web applications. One community post said SWE-2 Max built an iOS app that runs Gemma4-E2B locally on an iPhone. That result is an unverified user demonstration, so it is better treated as a possible workflow than as a supported product capability. The iOS demonstration claim is documented here
One user also claimed to have consumed more than 1 billion SWE-2 tokens and described the model as unusually good value during the promotion. That observation indicates strong early adoption interest, not a benchmark result. The usage and value claim appears in this post
How Devin SWE 2 Compares
The available benchmark comparisons are Cognition-reported and mixed. SWE-2 is close to Claude Fable 5.1 on FrontierCode 1.1 Main, ahead of Fable 5.1 on Terminal-Bench 2.1, and far behind it on Terminal-Bench 4.
| Model | FrontierCode 1.1 Main | Terminal-Bench 2.1 | Terminal-Bench 4 |
|---|---|---|---|
| Devin SWE 2 | 50.0% | 92.8% | 27.3% |
| Claude Fable 5.1 | 50.9% | 91.4% | 55.8% |
| GPT-6 Astra | 53.3% | 89.9% | 57.9% |
| Kimi K3 base | 44.2% | 88.3% | 21.5% |
On Cognition’s reported cost comparison, SWE-2 was 64% cheaper than Fable 5.1 and approximately one-quarter the cost of GPT-6 Astra. The cost figures are tied to the vendor’s evaluation methodology and should not be read as general API prices. For additional context on Astra’s release evidence, see the GPT-6 Astra release analysis.
On Cognition-reported evaluations, SWE-2 combines a 50.0% FrontierCode 1.1 Main score with a 92.8% Terminal-Bench 2.1 score.
The comparison also shows why a single ranking is inadequate. SWE-2’s 27.3% on Terminal-Bench 4 is substantially below Fable 5.1 at 55.8% and GPT-6 Astra at 57.9%. Community reports similarly disagree about persistence: some testers describe reliable long tasks, while others report repeated stops, slow fixes, or the need to send “continue.”
Availability: How to Access Devin SWE 2
The primary access path is Devin itself. SWE-2 was reported available in Devin Desktop and Devin CLI at release, with Devin Web and Fusion receiving rollout afterward. Public reporting does not confirm a standalone API endpoint, downloadable weights, or a provider-neutral model identifier.
Pricing is less settled. Multiple community posts describe SWE-2 as free or unlimited for users on qualifying Devin plans during the launch period. A third-party report describes Devin Pro as a $20-per-month plan, while another user described free access through October. The exact eligibility rules, rate limits, parallel-request limits, and end date are not confirmed.
Engineers comparing the underlying coding-model category can also review Kimi K3 on kie.ai, although Kimi K3 is the reported base model and is not the same product as Devin SWE 2. Kie.ai does not currently host Devin SWE 2.
What We Don't Know Yet
Several important questions remain open:
- The 262,000-token context window is an early community report, not a confirmed model-card specification.
- No independently reproduced benchmark package confirms Cognition’s FrontierCode, DeepSWE, or Terminal-Bench results.
- The public pricing model is unclear. The $20 monthly Pro figure, “unlimited” wording, and promotion end date may describe different plans or temporary conditions.
- The maximum number of parallel agents and the limits for Desktop, CLI, Web, and Fusion have not been published so far.
- It is not confirmed whether a public API, enterprise endpoint, or standalone SDK will become available.
- User reports conflict on long-horizon reliability. Some describe strong delivery, while others report 20-minute delays, 40-minute task durations, or unresolved rework.
- There is no verified basis for claiming that SWE-2 is better than Claude Opus 5, Kimi K3, Codex, Fable 5.1, or Astra across most coding tasks.
Frequently Asked Questions
What is Devin SWE 2?
Devin SWE 2 is Cognition’s coding model for the Devin software-engineering agent, post-trained from Kimi K3. It is delivered through Devin rather than as an open-weight model or confirmed public API.
Is Devin SWE 2 free?
Devin SWE 2 was reported free and unlimited for some Devin Pro, Max, and Teams subscribers during a launch promotion. The promotion’s exact terms, limits, and end date are not yet confirmed.
How much does Devin SWE 2 cost?
Devin SWE 2 has no confirmed public per-token price. Community and third-party reports described access through a Devin Pro plan priced at $20 per month, with promotional access terms still unconfirmed.
Is Devin SWE 2 open source?
Devin SWE 2 is not open source based on the available release coverage, and its weights are not publicly available. The model is reported to run inside Devin’s proprietary harness.
What is Devin SWE 2’s context window?
Devin SWE 2 has an early community-reported context window of 262,000 tokens, but that figure is unconfirmed. Cognition has not supplied a confirmed context-window specification in the available material.
Devin SWE 2 vs Claude Opus 5?
Devin SWE 2 has been described by one user as being at an Opus 5 level, but no independent standardized comparison establishes that ranking. Cognition-reported benchmark results show competitive scores on some coding evaluations and substantial gaps on others.
Devin SWE 2 is best treated as a promising proprietary coding-agent component, not a universally dominant coding model, until independent evaluations and stable pricing are available.
What to watch next. Track whether Cognition publishes a model card or API, whether the 262,000-token context claim becomes official, and whether the free or unlimited promotion receives clear rate-limit and end-date terms.
Try related models on Kie.ai: OpenAI Codex, Claude Opus 5.5 and Claude Haiku 5.5.
About Sofia Marenco
Sofia stress-tests new models on coding and reasoning benchmarks and reports what holds up.
View all posts by Sofia Marenco