ScientifyFirst decide what the task needs: GPUs, how long it runs, what software it needs and whether it waits on people.
Writing analysis scripts or processing small data that finishes in minutes
Codex or Claude Code, local or cloud
A CPU environment is enough. Running locally keeps data on your machine.
GPU work such as molecular simulation, structure prediction or model training
Scientify or your own GPU server
None of the three cloud coding environments documents GPU support.
Tasks that run for hours or days
Scientify or your own server
Claude Code cloud stops background commands after about 30 minutes by default; local runs need your computer to stay on.
Tasks that need GROMACS, LAMMPS or other slow-to-build scientific software
Scientify
Scientify preinstalls them and self-tests on GPUs. Claude Code cloud only caches setup scripts that finish in about 5 minutes.
Tasks that pause for an advisor's or collaborator's decision
Scientify
It pauses when idle and keeps the full state. Codex Cloud keeps task state for 7 days; Claude Code cloud may reclaim idle VMs.
Maintaining a paper's code repository and opening PRs
Codex Cloud or Claude Code cloud
Both work on GitHub repositories, and Claude Code cloud also handles CI failures and review comments on PRs.
Criteria are ordered by their impact on research computing. Competitor details come from official documentation; items the vendor does not document are marked “Not publicly documented”.
Codex Cloud
- GPU
- Not publicly documented
- Preinstalled
- codex-universal general-purpose image
- Run time and idling
- No published task time limit
- State retention
- Up to 7 days after the last turn starts or resumes
- Software installed mid-session
- Not carried into new tasks
- Installing software
- Environment setup script, applied after publishing
- Default network
- Package-manager preset; other domains must be added one by one
- SSH and private networks
- No SSH; Tailscale is the only supported VPN
- CPU / memory / disk
- Plus: 2 vCPU / 8 GiB / 8 GiB; Pro and above: 4 vCPU / 16 GiB / 32 GiB
- Environment model
- Publish an environment once; each task starts an isolated workspace from the same snapshot
- Unit of work
- GitHub repository
- Plans and cost
- ChatGPT Plus and above; no separate VM charge
Claude Code cloud
- GPU
- Not publicly documented
- Preinstalled
- Common language toolchains, Docker, PostgreSQL 16, Redis 7
- Run time and idling
- Background commands stop after about 30 minutes by default (configurable); the VM pauses after a few idle minutes
- State retention
- After the VM is reclaimed, only conversation history is restored; background processes are not
- Software installed mid-session
- Not carried into other sessions
- Installing software
- Setup script runs as root; cached only if it finishes in about 5 minutes
- Default network
- Trusted allowlist; can be switched to none, full or custom
- SSH and private networks
- Not publicly documented
- CPU / memory / disk
- 4 vCPU / 16 GB / 30 GB
- Environment model
- One virtual machine per session
- Unit of work
- GitHub repository, or an uploaded local repository
- Plans and cost
- Pro, Max, Team and some Enterprise seats; shares usage limits
Jules
- GPU
- Not publicly documented
- Preinstalled
- Common language toolchains, Docker
- Run time and idling
- Not publicly documented
- State retention
- Environment snapshot kept per repository
- Software installed mid-session
- Only what the setup script writes to the snapshot is kept
- Installing software
- Setup script, followed by a snapshot
- Default network
- Internet access
- SSH and private networks
- Not publicly documented
- CPU / memory / disk
- Not publicly documented
- Environment model
- One short-lived virtual machine per task, with an environment snapshot per repository
- Unit of work
- GitHub repository
- Plans and cost
- Free: 15 tasks per day, 3 concurrent; higher on Pro and Ultra
Scientify
- GPU
- Rented per task by the agent after you grant permission, billed per second
- Preinstalled
- Domain scientific software and databases: GROMACS, LAMMPS, OpenMM, ColabFold and more
- Run time and idling
- Runs while a task is executing, with no maximum duration
- State retention
- Pauses when idle and keeps the full state for resuming
- Software installed mid-session
- Kept
- Installing software
- The agent installs dependencies in the workspace and they persist
- Default network
- No outbound allowlist
- SSH and private networks
- SSH to your own servers; the agent never sees private keys
- CPU / memory / disk
- The agent runs in a cloud computer; compute runs on rented GPU instances
- Environment model
- One long-lived cloud computer per user
- Unit of work
- Project workspace: literature, data, code and results
- Plans and cost
- Pay-as-you-go; models cost about 30% of standard API pricing; new users get $5 worth of free credits
Codex Cloud
OpenAI's cloud coding environment. Since September 2026 it uses reusable environments: you publish an environment once, and each task starts from the same filesystem snapshot.
Research tasks it fits
- Maintaining a paper's code repository with parallel edits and tests.
- CPU data processing and unit tests that finish in minutes.
Limits
- No SSH, and no computer or browser use.
- Task state is kept for up to 7 days; commit important changes to git.
Claude Code cloud
Isolated virtual machines hosted by Anthropic, one per session, generally available since September 23, 2026. Sessions can start from the web, mobile, desktop app or terminal.
Research tasks it fits
- Data pipelines that depend on databases or Docker services.
- Analysis scripts that need root to install system libraries, where installation finishes in about 5 minutes.
Limits
- Setup scripts must finish in about 5 minutes or the environment is not cached.
- Idle VMs can be reclaimed; background processes and subagents are not restored.
Jules
Google's asynchronous coding agent. Each task runs in a short-lived virtual machine, with an environment snapshot per repository.
Research tasks it fits
- Small, independent repository changes such as adding tests or tidying scripts.
- Trying a cloud coding agent for free.
Limits
- VM size and task duration are not published.
- The free plan is limited to 15 tasks per day.
Scientify
A scientific agent built on Codex that runs in an isolated cloud computer. Its unit of work is a project workspace with literature, data, code and results.
Research tasks it fits
- Simulations, structure prediction and model training that need GPUs.
- Projects that need preinstalled scientific software, long run times and a persistent state.
Limits
- Built for research tasks, not organized around GitHub repositories and PRs.
- For day-to-day software development, the other three products fit better.
- CPU only, finishes in minutes
- Use Codex or Claude Code, local or cloud.
- Needs GPUs or runs for hours or longer
- Use Scientify or your own server.
- Depends on large scientific software you do not want to build
- Use Scientify.
- Day-to-day maintenance of a paper's code repository
- Use Codex Cloud or Claude Code cloud, and hand computational experiments to Scientify.
Can I run molecular dynamics on Codex or Claude Code cloud environments?
Short CPU tests work; production runs do not fit. None of the three documents GPU support. Claude Code cloud stops background commands after about 30 minutes by default and only caches setup scripts that finish in about 5 minutes, and building the GROMACS GPU version usually takes longer.
Do tasks stop when I close my laptop?
No. All three keep cloud tasks running after you close your laptop. They differ when idle: Claude Code cloud pauses the VM after a few idle minutes and may reclaim it later; Codex Cloud keeps task state for 7 days.
Is software I install during a session still there next time?
In Codex Cloud and Claude Code cloud, only software installed by the environment setup is kept; installs made mid-session do not carry over to new tasks or sessions. Scientify's cloud computer keeps the dependencies you install.
How should I split paper code and computational experiments?
Hand repository edits, tests and PRs to Codex or Claude Code; hand experiments that need GPUs, long run times or large scientific software to Scientify. Data and figures from the experiments then go back into the repository.
Scope of this comparison
- This page is based on official public documentation. No head-to-head tests were run.
- This page does not compare model capability or code quality.
- All three products change often. The official documentation is authoritative.