SWE-2 Flash
Cognition's fast, low-cost coding model for Devin, at one-fifth of SWE-2's price
Context window
Not published
Input / 1M tokens
$0.10
Output / 1M tokens
$0.50
Provider
Cognition
Devin token pricing after the free period: $0.10 per 1M input, $0.50 output, $0.01 cache reads and $0.125 cache writes, the same for all three effort variants. Cognition's docs say self-serve use is free until December 4, 2026 at 18:00 UTC (the launch post said through November 30), and some enterprise plans get 25% off until November 18, 2026. Cognition says it costs one-fifth as much as SWE-2. · Data verified 2026-11-11
SWE-2 Flash (November 4, 2026) is a smaller, faster sibling of Cognition's SWE-2 coding model, built for the routine work software agents do: automations, code review, feature development and documentation updates. Cognition mid-trained and then RL-trained it on Thinking Machines' Inkling-Small, a 276B-parameter Mixture-of-Experts with 12B active per token, and reports about 200 tokens per second versus 60 for SWE-2. It runs only in Devin Desktop and Devin CLI, in medium, high and max effort variants. Cognition is building a program to share the weights with industry partners and researchers.
How to access it
Available in Devin Desktop and Devin CLI (not Devin cloud sessions) as three variants: swe-2-flash-medium, swe-2-flash-high and swe-2-flash-max. No standalone API or public weights; Cognition is building a gated program to share weights with partners and researchers.
Strengths
- ✓About 3x faster output than SWE-2 (200 vs 60 tokens per second, per Cognition)
- ✓$0.10/$0.50 per 1M after the free period
- ✓Terminal-Bench 4.0 rose from 0.9% for the Inkling-Small base to 20.0% after Cognition's training
- ✓Beats Muse Spark Open and Reflection Beam 1.0 on Cognition's coding evals
Best for developers who...
When to choose it (and when not to)
Reach for SWE-2 Flash when...
- →You use Devin Desktop or CLI and want fast, cheap runs for reviews, automations and docs
- →You want to cut Devin costs on routine tasks (one-fifth of SWE-2's price)
Look elsewhere if...
- ✕Long-horizon or hard coding tasks: SWE-2 scores higher on DeepSWE v1.1 (73.0% vs 60.1%) and Terminal-Bench 4.0 (27.3% vs 20.0%)
- ✕You need a standalone API, OpenRouter access or open weights
- ✕You work in Devin cloud sessions, which do not offer it
How to use it
- ›Pick the effort variant per task: medium for small edits, max for multi-file changes
- ›Keep tasks well scoped; Cognition built it for automations, code review and docs work
Quickstart
N/A// No standalone API. Select SWE-2 Flash (medium, high or max) in Devin Desktop or Devin CLISWE-2 Flash is not available through a public API or model routers. Use it through Devin Desktop or Devin CLI only.
API model id: swe-2-flash-medium
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Terminal-Bench 4.0 | 20.0% | Cognition-reported; SWE-2 27.3%, Inkling-Small base 0.9% |
| DeepSWE v1.1 | 60.1% | Cognition-reported; SWE-2 73.0% |
| FrontierCode 2.0 | 24.0% | Cognition-reported; SWE-2 23.8% |
| Terminal-Bench 2.1 | 89.2% | Cognition-reported; SWE-2 92.8% |
| SWE-Bench Pro v2 Hard | 89.9% | Cognition-reported variant, not comparable with standard SWE-bench Pro; SWE-2 90.6% |
Tools powered by SWE-2 Flash
Compare SWE-2 Flash
Compare SWE-2 Flash with any other model
Build a comparison →All model comparisons →