Language modelAvailable in Devin

SWE-2

Cognition

Released: September 10, 2026
Data date: October 7, 2026

SWE-2 is Cognition’s coding model for Devin. The provider describes a central training change: one RL run optimizes all reasoning-effort levels together and includes cost in the reward. The model is available in Devin Desktop and CLI, with access through Devin Web and Fusion rolling out at launch.

Specifications and access

Specification	Value and source
Access	Available in Devin Desktop and CLI at announcement; rollout to Devin Web and Fusion was announced.Source
Training stage	Post-trained from Kimi K3; Cognition trains all reasoning-effort levels in one RL run.Source
Public weights	The announcement documents Devin access but no public weight release.Source

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

FrontierCode 1.1 Main (Show measurement, test conditions, and source)

Source value: 50
Score: 50
Metric: reported score
Unit: %
Benchmark version: 1.1 Main
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition's table reports a SWE-2 result in its Devin coding-model evaluation. It does not disclose the complete model-serving, agent-harness, tool, or decoding configuration for the table row; treat as provider-reported system/service performance, not an isolated raw language-model run.
Context: Provider-reported; sample count and sufficient harness details to separate the bare model from the deployed coding-agent path are not published.

Source Cognition SWE-2 launch announcement

DeepSWE (Show measurement, test conditions, and source)

Source value: 73
Score: 73
Metric: reported score
Unit: %
Benchmark version: 1.1
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition's table reports a SWE-2 result in its Devin coding-model evaluation. It does not disclose the complete model-serving, agent-harness, tool, or decoding configuration for the table row; treat as provider-reported system/service performance, not an isolated raw language-model run.
Context: Provider-reported; sample count and sufficient harness details to separate the bare model from the deployed coding-agent path are not published.

Source Cognition SWE-2 launch announcement

Terminal-Bench (Show measurement, test conditions, and source)

Source value: 92.8
Score: 92.8
Metric: reported score
Unit: %
Benchmark version: 2.1
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition's table reports a SWE-2 result in its Devin coding-model evaluation. It does not disclose the complete model-serving, agent-harness, tool, or decoding configuration for the table row; treat as provider-reported system/service performance, not an isolated raw language-model run.
Context: Provider-reported; sample count and sufficient harness details to separate the bare model from the deployed coding-agent path are not published.

Source Cognition SWE-2 launch announcement

Terminal-Bench (Show measurement, test conditions, and source)

Source value: 27.3
Score: 27.3
Metric: reported score
Unit: %
Benchmark version: 4
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition's table reports a SWE-2 result in its Devin coding-model evaluation. It does not disclose the complete model-serving, agent-harness, tool, or decoding configuration for the table row; treat as provider-reported system/service performance, not an isolated raw language-model run.
Context: Provider-reported; sample count and sufficient harness details to separate the bare model from the deployed coding-agent path are not published.

Source Cognition SWE-2 launch announcement

Propaganda and censorship evaluation, pass rate (overall) (Show measurement, test conditions, and source)

Source value: 98
Score: 98
Metric: judge pass rate
Unit: %
Benchmark version: Cognition trustworthiness evaluation (2026)
Category: model-output-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition prompted SWE-2 with 145 politically sensitive China-related questions in English, Simplified Chinese, and Traditional Chinese; GPT-5.6 Luna judged whether answers were substantive without adopting the official PRC position. Empty answers and execution/grading errors were excluded; exact accepted denominator is not reported.
Context: Vendor-run, model-output judge test. The source does not report the effective denominator after excluded responses/errors, so sampleCount is omitted.

Source Cognition SWE-2 launch announcement

Propaganda and censorship evaluation, pass rate (English) (Show measurement, test conditions, and source)

Source value: 99.8
Score: 99.8
Metric: judge pass rate
Unit: %
Benchmark version: Cognition trustworthiness evaluation (2026)
Category: model-output-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition prompted SWE-2 with 145 politically sensitive China-related questions in English, Simplified Chinese, and Traditional Chinese; GPT-5.6 Luna judged whether answers were substantive without adopting the official PRC position. Empty answers and execution/grading errors were excluded; exact accepted denominator is not reported.
Context: Vendor-run, model-output judge test. The source does not report the effective denominator after excluded responses/errors, so sampleCount is omitted.

Source Cognition SWE-2 launch announcement

Propaganda and censorship evaluation, pass rate (Simplified Chinese) (Show measurement, test conditions, and source)

Source value: 95.2
Score: 95.2
Metric: judge pass rate
Unit: %
Benchmark version: Cognition trustworthiness evaluation (2026)
Category: model-output-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition prompted SWE-2 with 145 politically sensitive China-related questions in English, Simplified Chinese, and Traditional Chinese; GPT-5.6 Luna judged whether answers were substantive without adopting the official PRC position. Empty answers and execution/grading errors were excluded; exact accepted denominator is not reported.
Context: Vendor-run, model-output judge test. The source does not report the effective denominator after excluded responses/errors, so sampleCount is omitted.

Source Cognition SWE-2 launch announcement

Propaganda and censorship evaluation, pass rate (Traditional Chinese) (Show measurement, test conditions, and source)

Source value: 99.1
Score: 99.1
Metric: judge pass rate
Unit: %
Benchmark version: Cognition trustworthiness evaluation (2026)
Category: model-output-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Cognition SWE-2 launch announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Cognition prompted SWE-2 with 145 politically sensitive China-related questions in English, Simplified Chinese, and Traditional Chinese; GPT-5.6 Luna judged whether answers were substantive without adopting the official PRC position. Empty answers and execution/grading errors were excluded; exact accepted denominator is not reported.
Context: Vendor-run, model-output judge test. The source does not report the effective denominator after excluded responses/errors, so sampleCount is omitted.

Source Cognition SWE-2 launch announcement

FrontierCode 1.1 Main score (medium) (Show measurement, test conditions, and source)

Source value: 43.09
Score: 43.09
Metric: weighted aggregate of rubric items; runs failing blocker criteria receive zero
Unit: %
Benchmark version: 1.1 Main
Category: agent-system
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cognition official FrontierCode 1.1 leaderboard data
Status: active
Retrieved at: 2026-10-07
Methodology: Official benchmark leaderboard. Cognition's official FrontierCode 1.1 leaderboard record for SWE-2, using the Devin harness at medium reasoning effort. The Main subset contains 100 tasks. The current revision zeroes runs flagged for unfair internet use. The published dataset does not give a per-effort trial count or a more detailed runtime configuration.
Context: Cognition-reported Devin agent-system result, not an isolated language-model evaluation. No per-effort trial count is published.

Source Cognition official FrontierCode 1.1 leaderboard data

FrontierCode 1.1 Main score (high) (Show measurement, test conditions, and source)

Source value: 47.2
Score: 47.2
Metric: weighted aggregate of rubric items; runs failing blocker criteria receive zero
Unit: %
Benchmark version: 1.1 Main
Category: agent-system
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cognition official FrontierCode 1.1 leaderboard data
Status: active
Retrieved at: 2026-10-07
Methodology: Official benchmark leaderboard. Cognition's official FrontierCode 1.1 leaderboard record for SWE-2, using the Devin harness at high reasoning effort. The Main subset contains 100 tasks. The current revision zeroes runs flagged for unfair internet use. The published dataset does not give a per-effort trial count or a more detailed runtime configuration.
Context: Cognition-reported Devin agent-system result, not an isolated language-model evaluation. No per-effort trial count is published.

Source Cognition official FrontierCode 1.1 leaderboard data

FrontierCode 1.1 Main pass rate (medium) (Show measurement, test conditions, and source)

Source value: 48.33
Score: 48.33
Metric: share of trials satisfying every blocker rubric criterion, averaged over the selected tasks
Unit: %
Benchmark version: 1.1 Main
Category: agent-system
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cognition official FrontierCode 1.1 leaderboard data
Status: active
Retrieved at: 2026-10-07
Methodology: Official benchmark leaderboard. Cognition's official FrontierCode 1.1 leaderboard record for SWE-2, using the Devin harness at medium reasoning effort. The Main subset contains 100 tasks. The current revision zeroes runs flagged for unfair internet use. The published dataset does not give a per-effort trial count or a more detailed runtime configuration.
Context: Cognition-reported Devin agent-system result, not an isolated language-model evaluation. No per-effort trial count is published.

Source Cognition official FrontierCode 1.1 leaderboard data

FrontierCode 1.1 Main pass rate (high) (Show measurement, test conditions, and source)

Source value: 52.35
Score: 52.35
Metric: share of trials satisfying every blocker rubric criterion, averaged over the selected tasks
Unit: %
Benchmark version: 1.1 Main
Category: agent-system
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cognition official FrontierCode 1.1 leaderboard data
Status: active
Retrieved at: 2026-10-07
Methodology: Official benchmark leaderboard. Cognition's official FrontierCode 1.1 leaderboard record for SWE-2, using the Devin harness at high reasoning effort. The Main subset contains 100 tasks. The current revision zeroes runs flagged for unfair internet use. The published dataset does not give a per-effort trial count or a more detailed runtime configuration.
Context: Cognition-reported Devin agent-system result, not an isolated language-model evaluation. No per-effort trial count is published.

Source Cognition official FrontierCode 1.1 leaderboard data

Page 1 of 2

Sources and data date

Type	Evidence and data date
Research status	Research date October 7, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources	independent-evaluation-and-harness Original wording The Cognition launch post and first-party FrontierCode 1.1 dataset now expose exact SWE-2 Devin-harness results, including effort and Main/Extended subsets. No independent standalone SWE-2 model evaluation with a disclosed exact checkpoint and reproducible external harness was found. Cognition does not publish complete runtime details or per-effort trial counts for the leaderboard rows.
Not found in the inspected sources	context-dependent-vulnerability-values Original wording Cognition describes the 1-to-5 score and confidence intervals but the readable source text does not provide numeric SWE-2 values for individual framings.
Additional source	Cognition SWE-2 announcement (retrieved October 7, 2026) · Cognition SWE-2 announcement · Editorial description reviewed October 7, 2026
Additional source	Cognition official FrontierCode 1.1 leaderboard data (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)