Devin now supports ChatGPT sign-in: Fusion with Astra is 39% cheaper

Devin now supports ChatGPT sign-in: a Plus or Pro subscription covers the cost of OpenAI models in Devin Cloud, Devin Desktop and Devin CLI. Usage counts against the Codex and ChatGPT Work limits that are already included in the plan. Devin is also publishing savings figures. According to the Artificial Analysis Coding Agent Index, Fusion mode, which pairs GPT-6 Astra with SWE-2, costs $4.54 per benchmark run, versus $7.47 for Codex running Astra alone. That is a 39% difference.

How the connection works

Once you link your ChatGPT account, your subscription pays for the OpenAI-model share of each session, while other models are charged to your Devin quota. You can set the limit for Devin in ChatGPT settings: give it everything or only part of your allowance. If the limit runs out, sessions do not break off. They keep running on Devin's quota until the limit resets.

In Devin Cloud, OpenAI models are part of the mix the platform uses across its Lite, Normal, Ultra and Fusion modes. Devin takes the connected plan into account when choosing models and picks OpenAI more often, so that more of the work runs on the limit you have already paid for. Devin says GPT models are especially good at computer use. Devin's testing agent can use them when it opens an app in the browser and clicks through the interface to check its own work. GPT-6 Sol is also available there, with research preview status.

On a local machine, things are simpler. In Devin Desktop and Devin CLI, you choose any GPT model yourself, and requests are billed to your ChatGPT plan.

Fusion: Astra plans, SWE-2 writes

To make the ChatGPT limit last longer, Devin offers Fusion, with GPT-6 Astra as the lead model and SWE-2 as the assistant. Astra plans and reviews, and hands most of the implementation to SWE-2, which Devin describes as free. Astra spends fewer tokens per task, so the subscription drains more slowly.

Figures from the Artificial Analysis Coding Agent Index, as cited by Devin:

• Codex with Astra (max): 61.6 points, $7.47 per run (taken as 100 for cost).
• Devin Fusion, Astra + SWE-2: 58.9 points, $4.54. That is 61% of the baseline cost, i.e. −39%.
• Claude Code with Fable 5.1 (max): 62.2 points, $12.36.
• Devin Fusion, Fable 5.1 + SWE-2: 61.7 points, $7.90. Savings of 36%.
• Claude Code with Opus 5 (max): 60 points.
• Muse Code (Muse Spark 1.3) and Opencode (GLM-5.3): 54 points each.
• Kimi Code CLI (Kimi K3): 52. Grok Build (Grok 4.6): 47. Claude Code with Qwen3.8 Max and Codex with DeepSeek V4: 43 each. Antigravity SDK (Gemini 3.8 Flash): 42.

The savings cost a few points of quality. Fusion on Astra scored 58.9 versus 61.6 for Codex, a gap of 2.7 points. The Fable 5.1 pairing trails the leader only slightly: 61.7 versus 62.2. The caveat is obvious: Devin itself published the data on its own blog. The index is third-party, but the choice of configurations and the run conditions can only be verified independently.

Who this is useful for

The benefit has two audiences. The first is developers who already have ChatGPT Plus or Pro. Previously they had to pay for OpenAI models twice: in the subscription and in the agent's quota. Now a single limit works in both places. The second is teams on Devin Teams, who can split costs between the subscription and the platform's quota.

The practical order of steps is this. Connect ChatGPT in Devin settings and immediately set the limit for Devin in ChatGPT settings, so the agent does not burn through the whole subscription in a couple of long sessions. In Desktop and CLI, it makes sense to start routine tasks with Fusion on the Astra + SWE-2 pairing and keep plain Astra for hard cases. On your own repository, compare cost and quality on 5–10 typical tasks rather than relying on the index's average score. The feature is available on Devin Pro, Max and Teams plans. Details on eligible plans and billing in different modes are in the documentation.