SWE-2 Flash Benchmarks, Specifications & Availability

SWE-2 Flash

Explore SWE-2 Flash from Cognition: published specifications, source-linked vendor benchmarks, independent evaluator coverage and recorded pricing when available.

Published specifications

Provider: Cognition
Access: Proprietary
License: Proprietary
Context window: Not published
Total parameters: 276B
Active parameters: 12B
Released: 2026-11-04
Modalities: text
Family: SWE-2

Announcement

Model notes

Cognition’s fast coding model, released 4 November 2026 in Devin Desktop and Devin CLI (free through 30 November 2026 per the launch page). Mid-trained and then trained with reinforcement learning on Thinking Machines’ Inkling-Small; 276B total / 12B active per Cognition’s size table. Cognition says it is not sharing the weights publicly yet and is building a program to share them securely with trusted industry partners and research labs; no Hugging Face weights at this check (the cognition-ai organization lists only Kevin-32B), so access stays closed. Cognition describes it as one-fifth the price and three times the speed of SWE-2, but publishes no per-token price, context length or modality list; text is the only modality its evaluations use. The launch page’s 200 versus 60 tokens per second demo is a simulated replay, not a measured API speed. No OpenRouter route, so no price row. The Devin model list did not yet name SWE-2 Flash on 5 November 2026.

Compare SWE-2 Flash with other models →Compare subscription plans Explore data coverage

Pricing · OpenRouter

No OpenRouter pricing is recorded for this model. Missing rates are not free.

Official / vendor benchmarks

Default headline records. Own-vendor, peer-vendor and third-party provenance remain visible in evidence; configurations may differ.

Official / vendor headline scores; expand evidence for every record
Benchmark / evaluator	Headline score	Evidence
FrontierCode 2.0	24%max effort	All 2 recorded results & sources 23.2% · raw 23.17 % Alternative · low effort · Own vendor Source/record date: 2026-11-04 Cognition primary public data.json (https://cognition.com/data/frontier-code-2.0/data.json), FrontierCode 2.0 main, 132 tasks; weighted new_score x100, not unweighted correct. Harness: devin. {"cost": 0.203}. Cost is source-observed mean USD per rollout at list API pricing, not an API rate. Data checked 5 November; evaluation date unpublished. Headline is the best-scoring published effort, as Cognition displays it. Cognition’s first 5 November copy of the file also listed a high-effort run (18.77, cost 0.2185); the current file has withdrawn it, so it is not stored. https://cognition.com/frontiercode 24% · raw 24 % Headline · max effort · Own vendor Source/record date: 2026-11-04 Cognition primary public data.json (https://cognition.com/data/frontier-code-2.0/data.json), FrontierCode 2.0 main, 132 tasks; weighted new_score x100, not unweighted correct. Harness: devin. {"cost": 0.2024}. Cost is source-observed mean USD per rollout at list API pricing, not an API rate. Data checked 5 November; evaluation date unpublished. Headline is the best-scoring published effort, as Cognition displays it. Cognition’s first 5 November copy of the file also listed a high-effort run (18.77, cost 0.2185); the current file has withdrawn it, so it is not stored. https://cognition.com/frontiercode
FrontierCode 1.1 Main	45.5%unknown effort	The recorded result & source 45.5% · raw 45.5 % Headline · unknown effort · Own vendor Source/record date: 2026-11-04 Cognition SWE-2 Flash launch table (4 November 2026), row “FrontierCode 1.1 Main”, column “SWE-2 Flash”. Cognition publishes no methodology for this table; its SWE-2 post says it reports a model’s public result where one exists and otherwise runs it in the harness it was developed for (Claude Code for Anthropic, Codex for OpenAI, Devin CLI for open-weight models) at its best reasoning effort. Effort per cell unstated. https://cognition.com/blog/swe-2-flash
DeepSWE v1.1	60.1%unknown effort	The recorded result & source 60.1% · raw 60.1 % Headline · unknown effort · Own vendor Source/record date: 2026-11-04 Cognition SWE-2 Flash launch table (4 November 2026), row “DeepSWE 1.1”, column “SWE-2 Flash”. Cognition publishes no methodology for this table; its SWE-2 post says it reports a model’s public result where one exists and otherwise runs it in the harness it was developed for (Claude Code for Anthropic, Codex for OpenAI, Devin CLI for open-weight models) at its best reasoning effort. Effort per cell unstated. https://cognition.com/blog/swe-2-flash
Terminal-Bench 4.0	20%unknown effort	The recorded result & source 20% · raw 20 % Headline · unknown effort · Own vendor Source/record date: 2026-11-04 Cognition SWE-2 Flash launch table (4 November 2026), row “Terminal-Bench 4.0”, column “SWE-2 Flash”. Cognition publishes no methodology for this table; its SWE-2 post says it reports a model’s public result where one exists and otherwise runs it in the harness it was developed for (Claude Code for Anthropic, Codex for OpenAI, Devin CLI for open-weight models) at its best reasoning effort. Effort per cell unstated. https://cognition.com/blog/swe-2-flash
Terminal-Bench 2.1	89.2%unknown effort	The recorded result & source 89.2% · raw 89.2 % Headline · unknown effort · Own vendor Source/record date: 2026-11-04 Cognition SWE-2 Flash launch table (4 November 2026), row “Terminal-Bench 2.1”, column “SWE-2 Flash”. Cognition publishes no methodology for this table; its SWE-2 post says it reports a model’s public result where one exists and otherwise runs it in the harness it was developed for (Claude Code for Anthropic, Codex for OpenAI, Devin CLI for open-weight models) at its best reasoning effort. Effort per cell unstated. https://cognition.com/blog/swe-2-flash
SWE-Atlas QnA	50%unknown effort	The recorded result & source 50% · raw 50 % Headline · unknown effort · Own vendor Source/record date: 2026-11-04 Cognition SWE-2 Flash launch table (4 November 2026), row “SWE-Atlas QA”, column “SWE-2 Flash”. Cognition labels it “SWE-Atlas QA”. Cognition publishes no methodology for this table; its SWE-2 post says it reports a model’s public result where one exists and otherwise runs it in the harness it was developed for (Claude Code for Anthropic, Codex for OpenAI, Devin CLI for open-weight models) at its best reasoning effort. Effort per cell unstated. https://cognition.com/blog/swe-2-flash
SWE-bench Pro v2-Hard	89.9%unknown effort	The recorded result & source 89.9% · raw 89.9 % Headline · unknown effort · Own vendor Source/record date: 2026-11-04 Cognition SWE-2 Flash launch table (4 November 2026), row “SWE-Bench Pro v2 Hard”, column “SWE-2 Flash”. Cognition publishes no methodology for this table; its SWE-2 post says it reports a model’s public result where one exists and otherwise runs it in the harness it was developed for (Claude Code for Anthropic, Codex for OpenAI, Devin CLI for open-weight models) at its best reasoning effort. Effort per cell unstated. https://cognition.com/blog/swe-2-flash
ExploitBench	34.8%unknown effort	The recorded result & source 34.8% · raw 34.8 % Headline · unknown effort · Own vendor Source/record date: 2026-11-04 Cognition SWE-2 Flash launch table (4 November 2026), row “ExploitBench”, column “SWE-2 Flash”. Cognition publishes no methodology for this table; its SWE-2 post says it reports a model’s public result where one exists and otherwise runs it in the harness it was developed for (Claude Code for Anthropic, Codex for OpenAI, Devin CLI for open-weight models) at its best reasoning effort. Effort per cell unstated. https://cognition.com/blog/swe-2-flash

Independent evaluators

Evaluator harnesses are distinct from vendor measurements. Missing coverage is not a failed test.

No independent evaluator results are recorded for this model.

Read how we select and source scores or the comparison guide.