Build1 publisher3 min readPublished
Hosted agent sandboxes from OpenAI, Perplexity and Gobare disagree on what survives idle time
Misha, who works on Gobare, found in public docs that OpenAI may delete a sandbox idle for an hour and Gobare pauses one after five minutes. Perplexity promises nothing between responses, so jobs that wait on a human or serve a preview have to fit one of those clocks.
The Engineer · Build desk
What happened
- Within one response, Perplexity runs every execution in the same container, so files and packages from an earlier step are still there for a later one.
- OpenAI turns files written to /workspace/outputs into immutable artifacts that stay downloadable after the sandbox expires, and the rest of the workspace is lost.
- Perplexity publishes a container price of $0.03 per session with a billing window of up to 20 minutes.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint A 40-minute refactor spread over several turns cannot be held in one Perplexity container, and on Gobare every gap longer than five minutes sends it through a pause and wake.
- exposure An approval step that sits with a human for more than an hour outlasts OpenAI's idle window, so anything the agent did not publish to /workspace/outputs is gone when the approver answers.
- decision Choosing OpenAI's hosted API for its sandbox ties the job to OpenAI models; choosing the SDK keeps model choice open but moves the sandbox lifecycle, and its outages, onto the buyer.
- cost At about $0.09 an hour, Perplexity's container is the public price other runtimes get compared with, though it covers only work that fits inside a single response.
Each provider starts a different clock when the agent goes quiet. Misha's comparison opens on a 30-second demo in which all three runtimes look alike [2]. The shortest idle clock among them is five minutes [6]. He disclosed that he works on Gobare and says he quoted the other two vendors' public docs, checked 27 September 2026, instead of describing them [1]. Pick without checking, and "you find out an hour into a job, at the step where the work was supposed to land," he wrote [20].
His first test case is a 40-minute refactor across several turns, with gaps while a dispatcher does other work [8]. On Perplexity each response starts clean, so the job cannot live in one container. Any state that crosses a response boundary has to be rebuilt by the caller on previous_response_id [9]. On OpenAI's hosted sandbox the gaps are safe as long as none of them reaches an hour with no activity and no keep-alives [4]. On Gobare, five quiet minutes pause the workspace and the next message wakes it [6], inside a cap of two hours of active time [19].
Human approval is harder. The machine waits "maybe for hours," Misha wrote [10]. Past an hour, OpenAI can delete the sandbox, and only what the agent wrote to /workspace/outputs remains [4][5]. Gobare wakes a paused workspace on the next message, but after a reclaim the agent gets back only what a best-effort snapshot caught. That snapshot leaves out installed dependencies, git history and running processes, and it skips any workspace over 300MB [7]. A preview server is a running process. If the sandbox goes, the preview URL goes with it [11].
I think OpenAI's hosted contract is the easiest of the three to design against. It has one clock and one directory that outlives it [4][5]. An agent that publishes results to /workspace/outputs as it goes loses only scratch state when the hour runs out. That is a clean piece of engineering, and it is honest about what it keeps.
OpenAI's naming adds a second decision. The Agents SDK is open source, runs in the caller's process and reaches other models through LiteLLM in Python or the AI SDK adapter in TypeScript [12]. The hosted Agents API is the one with the sandbox. It is built around OpenAI models, with others on what Misha calls a second-class path [13]. Self-hosting hands the lifecycle back to the caller. With Daytona as the provider, the post says auto_stop_interval=0 has to be set, or the sandbox stops mid-agent [14].
Perplexity's published price is $0.03 per session with a billing window of up to 20 minutes [15]. Three windows an hour puts that at about $0.09 an hour [16]. Misha argues the runtime does not need to be profitable for Perplexity because its money is in the search index, and that this sets the floor for everyone else [17]. That is his inference; Perplexity has not said it. The price applies to a long job only if the job fits inside one response, and the refactor and approval cases do not [8][10].
He is specific about Gobare's own limits too. It runs on a single sandbox provider, E2B, with no bring-your-own compute. "That is real lock-in, and the fairest criticism of us here," he wrote [18].
What to watch
- Whether OpenAI documents a way to keep a hosted sandbox alive past the one-hour idle window for approval-gated runs.
- Whether Gobare adds a sandbox provider beyond E2B or extends its snapshot to cover installed dependencies.
- Whether Perplexity offers session state that survives a response boundary, or changes its $0.03 session price.