On September 10, the @devindesktop account announced that SWE-2 — Cognition’s frontier coding model — is now available in Devin Desktop and Devin CLI, with unlimited usage for Pro and Teams subscribers through October 2026. The announcement came one day after Cognition shipped SWE-2 to Devin Cloud, but with a critical difference: Cloud removed SWE-2’s free tier, while Desktop and CLI are offering an all-you-can-eat trial month.
The divergence reflects Cognition’s evolving pricing strategy. Frontier models are no longer free across the board, but Desktop and CLI users get a temporary window to validate the upgrade — and Cognition gets a month of production telemetry on how much SWE-2 costs to serve when usage is uncapped.
What changed on Cloud vs Desktop/CLI
When SWE-2 launched on Devin Cloud on September 9, Cognition made it available under two tiers:
- Free tier users: No access to SWE-2. Free users are limited to SWE-1.7 (the previous-generation model).
- Pro and Teams users: SWE-2 included, but subject to usage quotas (specific limits weren’t disclosed, but community reports suggest roughly 100–150 SWE-2 tasks per month on Pro, higher on Teams).
When SWE-2 launched on Desktop and CLI the next day, the pricing was different:
- Free tier users: No access (same as Cloud).
- Pro and Teams users: Unlimited SWE-2 usage through October 31, 2026 — no quotas, no throttling.
The Desktop/CLI announcement explicitly called out the October 31 deadline, which suggests Cognition will reintroduce quotas in November, likely matching Cloud’s limits. But for one month, Desktop and CLI users on paid plans can run as much SWE-2 as they want.
The Cloud free tier change: why now?
SWE-2 is the first major model launch where Cognition removed free-tier access from day one. Previous models — including SWE-1.7, Opus 5, and Fable 5.1 — were available to free users with rate limits. SWE-2 is Pro/Teams only.
The reason is cost. SWE-2 runs at roughly $3.11 per task on FrontierCode 1.1 Main, per Cognition’s September 9 announcement. That’s cheaper than Opus 5 ($4.30/task) but more expensive than SWE-1.7 ($2.10/task). For a free user running 10–20 tasks per week, that’s $30–$60/month in inference costs that Cognition was previously subsidizing. At $1 billion ARR, the company is done subsidizing.
The shift also aligns with Cognition’s enterprise push. Free users are great for growth, but they don’t convert to $500/month team seats. By gating SWE-2 behind Pro/Teams, Cognition is telling free users: If you want frontier performance, you need to pay. That’s a harder line than most AI coding tools have drawn — GitHub Copilot still offers a free tier for students and open-source maintainers, and Cursor’s free tier includes limited access to GPT-4o — but it’s consistent with Cognition’s focus on enterprise customers who care more about quality than price.
What “unlimited” means (and doesn’t mean)
The Desktop/CLI announcement says “unlimited usage for Pro/Teams for the next month,” but there are practical limits:
-
API quota limits: Desktop and CLI users running SWE-2 via their own Anthropic or OpenAI API keys are still subject to those providers’ quotas. If you hit your Anthropic rate limit, SWE-2 stops working — Cognition’s “unlimited” policy doesn’t override that.
-
Inference speed: SWE-2 runs at 1,000 tokens/second on Cerebras (per Cognition’s July announcement), but typical tasks generate 5,000–15,000 tokens, so each task takes 5–15 seconds of inference time. “Unlimited” doesn’t mean instantaneous — you’re still waiting for the model to generate output.
-
Local compute: Desktop users running SWE-2 locally (instead of calling Cognition’s API) need enough RAM and GPU memory to load the model. SWE-2 is reportedly 70–90 billion parameters (unconfirmed, but consistent with its performance tier), which requires at least 48GB of VRAM to run at full speed. Most consumer GPUs (RTX 4090, M2 Max) can’t handle that, so “unlimited” only applies if you’re using Cognition’s hosted inference.
In practice, “unlimited” means no artificial quotas imposed by Cognition — but you’re still constrained by your API provider, your hardware, and the laws of physics.
Method: The Desktop/CLI availability and unlimited usage policy are from the @devindesktop X post on September 10. The Cloud pricing change (free tier removal) is inferred from community reports on Reddit and X that free-tier users lost access to SWE-2 after September 9. SWE-2's cost per task ($3.11) is from Cognition's September 9 FrontierCode announcement. The October 31 deadline is explicitly stated in the @devindesktop post. Inference speed (1,000 tokens/second on Cerebras) is from Cognition's July SWE-1.7 announcement, which used the same Cerebras deployment for inference. The 70–90B parameter estimate is from third-party analysis of SWE-2's inference latency and cost, cross-referenced with known parameter-to-cost ratios for other models.
References
- SWE-2 now in Devin Desktop and CLI — @devindesktop, September 10, 2026.
Editorial from Devin Central — news coverage by a Devin Ambassador, not Cognition. Devin is a trademark of Cognition.