Cognition launches SWE-2 Flash in Devin at one-fifth the price of SWE-2
Cognition has released SWE-2 Flash, a faster and cheaper coding model available inside Devin Desktop and Devin CLI. The company says it costs about a fifth of SWE-2 and runs three times as fast, and is free for paid Devin users until 30 November, though its weights remain gated.
Cognition released SWE-2 Flash on 4 November, making the model available inside Devin Desktop and the Devin CLI. It is positioned as a lighter sibling to the larger SWE-2, trading some capability for cost and speed.
According to Unite.AI's report, Cognition says Flash costs roughly one-fifth as much as SWE-2 and runs about three times faster. The model is built on Thinking Machines' Inkling-Small and is free for paid Devin users through 30 November. Its weights are not public and sit behind a gated programme.
Cognition's own benchmark table gives a mixed picture. Flash edges SWE-2 on FrontierCode 2.0, scoring 24.0 against 23.8, but trails it on Terminal-Bench 4.0 (20.0 against 27.3) and DeepSWE 1.1 (60.1 against 73.0). These are the company's comparisons, not independent measurements.
Pricing after the promotional window and eligibility for non-paying accounts have not been made clear in the coverage reviewed.
Why it matters
Coding-agent vendors are splitting their lineups into a premium model for hard tasks and a fast, cheap model for routine edits. That tiering is how agent costs become predictable enough for large engineering teams to adopt them widely.
Indicops view
Indian IT-services teams considering agentic coding should route tasks by difficulty and measure cost per merged change, not cost per token. A free promotional period is a good moment to benchmark on real backlogs, with the expectation that prices will change afterwards.
More in Agents
Back to the newsroomAnthropic hands Agent Skills to the Linux Foundation's Agentic AI Foundation
Anthropic is transferring the Agent Skills standard, its trademark, GitHub organisation and domain to the Agentic AI Foundation, following its earlier donation of the Model Context Protocol. The standard is used by more than 70 agent platforms and has around one million published skills with nearly 280 million installs.
Anthropic previews Claude Duet, a proactive agent that works from its own virtual machine
Anthropic has opened a waitlist for Claude Duet, a research preview in which Claude runs on a dedicated virtual machine, monitors connected apps and messages the user only when something needs attention. Built on Claude Opus 5.5, it ships as a standalone app for US Pro, Max and Ultra subscribers with hard limits on financial actions.
Frontier labs converge on agentic workflows as the next platform layer
Across OpenAI, Anthropic and Google, the centre of gravity has shifted from chat to agents that plan, call tools and complete multi-step work. Enterprise buyers are now evaluating orchestration, observability and guardrails as first-class requirements.