SWE-2 Flash: Cognition's Three Times Faster Coding Model and What the Trade-off Actually Costs
Cognition announced SWE-2 Flash on November 4, 2026: a 276 billion parameter mixture of experts model, adapted from Thinking Machines Lab's Inkling Small, that runs roughly three times faster and costs about one fifth as much as Cognition's larger SWE-2. It wins on one of three published benchmarks and loses on the other two. Weights are not public, and Cognition says that is deliberate. Here is what the actual numbers say about when the speed is worth the score you give up.
SWE-2 Flash: Cognition's Three Times Faster Coding Model and What the Trade-off Actually Costs
| Benchmark | Inkling Small (base) | SWE-2 Flash | SWE-2 |
|---|---|---|---|
| Terminal-Bench 4.0 | 0.9 | 20.0 | 27.3 |
| FrontierCode 2.0 | 3.2 | 24.0 | 23.8 |
| DeepSWE 1.1 | 17.4 | 60.1 | 73.0 |
| Reported speed | not stated | about 3 times SWE-2 | baseline |
| Reported cost | not stated | about one fifth of SWE-2 | baseline |
| Weights | public (July 30, 2026) | not public | not public |
Cognition announced SWE-2 Flash on November 4, 2026, pitched explicitly at routine software agent work rather than uniform benchmark leadership, which is an unusually honest framing for a model launch. The architecture is a mixture of experts model with 276 billion total parameters and 12 billion active per token, adapted from Thinking Machines Lab's Inkling Small, whose full weights Thinking Machines released publicly on July 30, 2026. Cognition trained it with what it calls the first recipe where a large scale mid training stage precedes reinforcement learning, run across a mixture spanning a few trillion tokens. It is available now in Devin Desktop and the Devin CLI, free through November 30, 2026, aimed at automations, code review, feature development, and documentation updates rather than the hardest end of agentic coding.
The benchmark table tells a specific story once you read past the headline speed claim. Flash beats SWE-2 on exactly one of the three published evaluations, FrontierCode 2.0, by a fraction of a point, 24.0 against 23.8. On the other two, Terminal-Bench 4.0 and DeepSWE 1.1, SWE-2 wins clearly, by 7.3 and 12.9 points respectively. That is the real shape of the trade: you are not getting a strictly worse model for a lower price, you are getting a model that is worse on harder, longer horizon tasks and roughly even on one narrower benchmark, in exchange for triple the throughput and a fifth of the cost. Whether that trade is good for you depends entirely on what fraction of your actual workload looks like DeepSWE and Terminal-Bench style multi step tasks versus FrontierCode style narrower completions, and Cognition's own benchmark selection does not answer that question for you.
The speed claim is real but the methodology deserves a second look
Cognition's speed comparison replayed a Terminal-Bench 4.0 run at simulated token rates, 200 tokens per second for Flash and 60 for SWE-2, with transcripts shortened to equal length before comparing. That is a reasonable way to isolate the effect of raw generation speed on wall clock time for a fixed task, but it is explicitly a simulation of the speed difference applied after the fact, not a live head to head run of both models against the same task set in real time. It tells you the throughput difference is real and substantial, a genuine 3x generation speed gap is not a rounding error, but it does not independently validate the accuracy numbers in the same pass. The same caution applies to reading any single harness's numbers as the whole picture, since Terminal-Bench 2.1 and Terminal-Bench 4.0 have produced meaningfully different rankings for the same models depending on task difficulty and saturation, and Flash's own scores sit well below where SWE-2 and the frontier closed models land on that specific, harder benchmark.
Comparing Flash against external coding agents rather than its own sibling is harder than it should be, because Cognition's published numbers only cover its own model family and American open weight models it says Flash beats on the evaluations it reports, without naming which ones or publishing head to head numbers against Claude, GPT, or Gemini class agents. Our own comparison of SWE-2 against Claude Opus 5.5 and Gemini 4 Argon found that harness specific scores for the same model can look like they describe two different products depending on which benchmark and which effort level produced them, and that same warning applies doubly here: Flash's numbers are not directly comparable to that earlier comparison's figures without knowing whether Cognition ran them under matching harness and effort settings, which it has not disclosed.
Why the weights are not public, and what that tells you about the real audience
Cognition is explicit that Flash's weights are not available for download, and plans only a secure sharing program for trusted partners and research labs rather than a public release. The stated reason is misuse risk: Cognition says targeted cybersecurity training raised Inkling Small's ExploitBench score from 11.3 percent to 34.8 percent, a tripling of a model's measured ability to find and use exploits, and cites that specifically as the justification for withholding public weights. That is a meaningfully different posture from DeepSeek, Qwen, and other labs that have made comparably capable coding models openly downloadable this year, and it signals who Flash is actually built for: teams that want Cognition's hosted product, Devin Desktop and the CLI, running against Cognition's infrastructure, not teams that want to self host a cheap, fast coding model on their own hardware. If self hosting a lower cost coding agent is your actual goal, Flash is not a candidate regardless of how attractive its price looks on paper, because you cannot run it yourself.
When this applies to your stack
If your team runs a meaningful volume of routine agentic coding work, automations, code review passes, documentation updates, rather than the hardest multi step refactors, SWE-2 Flash's price and speed are worth piloting specifically against that subset of work rather than your full workload. Measure cost per completed task on your own traces, the way you would for any model choice, since that number has consistently mattered more than headline benchmark scores once retries and tool calls get counted. Keep SWE-2 or a frontier closed model in the loop for the DeepSWE and Terminal-Bench style tasks where Flash gives up real accuracy, and do not treat the free access window through November 30 as a long term pricing signal, since it is explicitly promotional. Teams evaluating whether to consolidate onto a single coding agent versus routing by task type should treat this release as evidence for routing, not consolidation.
FAQ
Is SWE-2 Flash just a cheaper, strictly worse version of SWE-2? No. It beats SWE-2 narrowly on FrontierCode 2.0 and loses clearly on Terminal-Bench 4.0 and DeepSWE 1.1. It is a different point on the speed and accuracy curve, better suited to routine tasks, not a strictly dominated product.
Can I download and self host SWE-2 Flash? No. Cognition has not made the weights public and plans only a secure sharing program for trusted partners and research labs, citing cybersecurity misuse risk from the model's training.
How much faster and cheaper is SWE-2 Flash really? Cognition reports roughly three times the generation speed and about one fifth the cost of SWE-2, based on a simulated replay of a Terminal-Bench 4.0 run at matched token rates rather than a live side by side task run.
What is Flash actually built from? A mixture of experts architecture with 276 billion total and 12 billion active parameters, adapted from Thinking Machines Lab's Inkling Small, whose weights Thinking Machines released publicly on July 30, 2026.
Sources:
More from the lab.
GitHub Copilot Fusion and the Case for Multi Model Routing in Engineering Teams
GitHub shipped Copilot Fusion to general availability on October 28, 2026, and the pitch is a reversal of how most developers choose an AI coding tool: instead of picking a model up front, you pick an effort level and Fusion decides which models handle the task and in what order. On GitHub's own Artificial Analysis Coding Agent Index numbers, Fusion lands within two points of Claude Code running Opus 5.5 at its highest effort setting, at roughly half the reported cost and time. Here is how the routing actually works and what to verify before trusting the comparison.
Cognition's SWE-2 vs Claude Opus 5.5 vs Gemini 4 Argon: What the Benchmark Numbers Actually Mean
Cognition released SWE-2 in September 2026, post trained from Kimi K3's 2.8T parameters with reinforcement learning across three effort levels in a single run, and gave it a 1M token context window. On Terminal-Bench 2.1 it posts 92.8 percent, the best verified score in its benchmark catalog entry. On Terminal-Bench 4.0 it posts 27.3 percent, 43.3 points behind Claude Sonnet 5.5's 70.6. Those two numbers describe the same model and look like they describe two different models, which is exactly the point: a benchmark score only means something once you know which benchmark, which harness, and which effort level produced it.
FrontierSWE v2 Explained: What a 20 Hour Coding Benchmark Says About Claude Code, Cursor, and Codex (October 2026)
Most agentic coding benchmarks measure minutes to a couple of hours of autonomous work, and the top tier of frontier models has started clearing them fast enough that the leaderboards stop being useful signal. Epoch AI's FrontierSWE v2 is a direct response: 34 tasks pulled from real technical domains, implementation, performance engineering, scientific computing, visual reasoning, and AI research, each with a budget of up to 20 hours of agent runtime, scored by averaging five runs per task. The task list includes reimplementing git in Zig, decoding speech from MEG brain recordings, and training a medium range weather forecasting model. None of that is a toy problem, and that is the point.