Cognition launches SWE-2 Flash with a free November window in Devin
SWE-2 Flash is available in Devin Desktop and Devin CLI, with free use for Pro, Max and Teams subscribers through November 30, 2026. Cognition announced the coding model on November 4. Its launch thread specifies the eligible paid plans, so readers should not assume the offer includes every Devin account or workflow.
The immediate attraction is a cheaper candidate for recurring engineering work. For teams that need downloadable weights, the release has a different limitation: Cognition has not made them public.
What SWE-2 Flash changes
According to Cognition’s launch report, Flash builds on Thinking Machines’ Inkling-Small, with 276 billion total parameters and 12 billion active. Cognition added mid-training before reinforcement learning, using a few trillion tokens and placing scarce, high-quality material later in training.
The company reports 24.0 on FrontierCode 2.0 and 20.0 on Terminal-Bench 4.0, compared with 3.2 and 0.9 for its base model. Results vary by task suite: its table puts Flash below SWE-2 on DeepSWE 1.1 and Terminal-Bench 4.0, despite a slight lead on FrontierCode 2.0. These are vendor-reported results, not ByteForward tests.
Free access and the price claim
Cognition’s official announcement thread claims a fivefold cost reduction and threefold speed increase versus SWE-2. The launch page does not supply a per-token price or a complete post-promotion rate card. Those relative figures are insufficient to calculate a monthly bill.
The Devin Desktop page lists Pro at $20 per month, Max at $200 per month, and Teams at $80 per month plus $40 per full seat. These are subscription prices, separate from the temporary Flash offer. Existing Windsurf users receive Devin Desktop through the normal update path, with their plan and settings carried over.
Before relying on the promotion, check the selected model, account tier and billing screen in the actual Desktop or CLI session. The announcement names those two products. It does not establish that cloud sessions, every usage limit or all other model charges are waived.
Reading the speed comparison carefully
The launch page’s animation replays shortened runs on one Terminal-Bench task at simulated rates of 200 and 60 tokens per second. It illustrates the speed claim rather than providing an independent, end-to-end latency test.
For a coding task, fast text generation is only one part of completion time. Repository exploration, tool calls, test execution and human review can dominate. A useful comparison starts both models from the same commit, gives them the same instructions and records the time until a reviewed change is ready.
Use several kinds of work rather than repeating one easy prompt. A documentation correction, a narrowly scoped bug fix and a change that requires new tests reveal different failure modes. Record whether the patch stays within scope, whether tests detect the original defect and how much manual repair remains.
ByteForward’s guide to the independent FrontierCode 1.1 leaderboard explains why benchmark version, harness and grading rules belong beside a score. That older leaderboard should not be treated as independent confirmation of today’s Flash results.
Weights and deployment remain separate questions
Cognition says it is developing secure weight access for trusted industry partners and research labs, citing cybersecurity concerns. It promises further details in the coming weeks. The announcement does not establish public download, general on-premise availability or a redistribution license.
That leaves two different adoption decisions. An existing paid Devin user can evaluate the hosted model during the stated promotional window. A team whose requirements include self-hosting, fine-tuning or a specific data-control agreement still needs the program’s eligibility and contractual details before planning a deployment.
For now, the strongest practical use of the offer is a bounded trial on representative repository work. Keep the current model as a comparison, retain normal review and test gates, and judge total effort per accepted change. A lower inference price only helps when the resulting work is usable.
