Comparison · facts verified 2026-09-07
Same software-engineering goal, different boundary. Codex is OpenAI's cloud software-engineering agent: give it a task through ChatGPT, CLI, or GitHub and it works in an isolated cloud sandbox. Orbi turns a GitHub Issue into a reviewed PR on your machine, using the model endpoint you choose. This page separates the trade-offs without pretending one fits every team.
THE COMPETITOR, FROM OFFICIAL MATERIALS
OpenAI introduced Codex as a research preview in May 2025: a cloud software-engineering agent designed to handle multiple tasks in parallel. Each task runs in its own cloud sandbox with the repository preloaded, and the result can be a pull request. Codex is available through ChatGPT, the Codex CLI, and GitHub integrations, and uses OpenAI's Codex model family.
Sources: OpenAI's Introducing Codex announcement and Codex documentation — verified 2026-09-07.
THE CLOUD-OR-LOCAL DECISION
Codex cloud vs local is a choice about where the agent runs: OpenAI's managed sandbox, or the machine in front of you. The cloud route starts a task from ChatGPT or GitHub and runs it in an isolated sandbox with the repository preloaded, so nothing is set up locally and several tasks can run in parallel. The local route is the Codex CLI in your own terminal: it works against your local repository and uses the tools already installed on your machine, so the checkout and your credentials stay where they are — but you supply the machine, you start the run, and you stay with it while it works. What the cloud route takes in return is covered in the next section.
Orbi is the third option: local execution with cloud-style delivery. You write a GitHub Issue and add the ai-ready label; Orbi claims it, implements in a Git worktree on your machine, runs the tests, reviews the change, and opens one pull request for an ordinary merge — unattended from Issue to merge, with no prompt to paste and no run to supervise. The code, keys, and run evidence stay on your infrastructure, and the model endpoint is yours to choose (BYOK). The Codex guide covers pairing it with a ChatGPT/Codex subscription; if you would rather not operate the runner, Orbi Cloud runs the same Issue-to-merge workflow on a hosted runner.
So: pick cloud for zero setup and managed parallelism, pick the local CLI when the repository and credentials must not leave your machine, and pick Orbi when you want that local boundary without standing over the run.
Sources: OpenAI's Codex CLI documentation (terminal, local repository, installed tools) and Codex documentation; Orbi's documentation — the CLI page verified 2026-10-18, the rest 2026-09-07.
THE OWNERSHIP BOUNDARY
A cloud sandbox can remove local setup and make parallel work convenient. It also means the vendor controls the service price, model availability, account access, and infrastructure through which the preloaded repository and task context pass.
| Risk | Codex cloud | Orbi's trade-off |
|---|---|---|
| Pricing | Cloud access is tied to ChatGPT plans; Codex CLI/API usage can have separate API billing. Plan and usage policy can change. | Self-hosting needs your hardware; model usage follows your chosen provider's terms. |
| Degradation | Codex model iterations, availability, limits, and throttling are OpenAI decisions. | BYOK supports any OpenAI-compatible endpoint, so switching provider is a configuration choice. |
| Audit / data | The repository is preloaded into an OpenAI cloud sandbox for the task. | Code, credentials, state, and run evidence stay on your infrastructure. |
| Account access | ChatGPT, CLI, GitHub integration, and OpenAI account access are part of the path. | No OpenAI account is required to run the self-hosted core. |
Sources: OpenAI's announcement and Codex docs; Orbi's documentation — verified 2026-09-07. These are ownership trade-offs, not claims about service quality.
DIFFERENCES THAT DECIDE THE CHOICE
The useful question is not which agent is universally better. It is whether you want a ready cloud execution surface or a GitHub-native delivery system whose models and data boundary you control.
| Dimension | Orbi | OpenAI Codex cloud |
|---|---|---|
| Task entry | GitHub Issue + ai-ready label | ChatGPT sidebar, CLI, or GitHub integration |
| Runs where | Your machine and Git worktree | OpenAI cloud sandbox, one isolated environment per task |
| State | GitHub Issues, labels, branches, and PRs | Task and repository context in OpenAI's service |
| Models | BYOK, any OpenAI-compatible endpoint; provider terms vary | OpenAI Codex model family |
| Delivery | Independent review, then ordinary PR merge | Task output can include a PR |
| Cost | Self-hosted hardware plus chosen model usage; no $/session-hour | ChatGPT plan for cloud; separate API billing may apply to CLI/API use |
| Auditability | Readable and auditable end to end | Vendor service and cloud infrastructure |
Sources: OpenAI Codex docs; OpenAI launch announcement; Orbi docs — verified 2026-09-07.
CODEX COST · PRICES CHECKED 2026-10-17
Codex has no separate price. It comes with a ChatGPT plan, or you sign in with an API key and pay API rates per token. OpenAI's Codex pricing page lists these plans:
| Plan | Price | Codex in that plan |
|---|---|---|
| Free | $0/month | Quick coding tasks; GPT-6 Luna in the desktop app, subject to rollout |
| Go | $8/month | Lightweight coding tasks; GPT-6 Luna in the desktop app, subject to rollout |
| Plus | $20/month | Codex on the web, in the CLI, in the IDE extension and on iOS; GPT-6.1 Sol and GPT-6 Luna; automatic code review and Slack; extra usage with ChatGPT credits |
| Pro | $100, $200 or $500/month | Everything in Plus; no five-hour limit at present, though weekly limits may apply; Ultrafast mode on the $500 plan |
| Business | $20 per user/month billed annually, 2+ users; $25 billed monthly | Admin controls, SAML SSO and MFA; larger cloud VMs; extra usage with credits |
| Enterprise & Edu | Contact sales | Business features plus SCIM, EKM, RBAC, audit logs and data residency controls |
| API key | API token rates | CLI, SDK and IDE extension; no cloud features such as GitHub code review or Slack |
A plan buys an allowance, not a fixed number of tasks. For Plus and standard Business seats, OpenAI estimates 15–160 local messages every five hours on GPT-6.1 Sol and 350–3,000 on GPT-6 Luna. Cloud tasks can use more of the allowance, and weekly limits may also apply. OpenAI says not to estimate included tasks from API token prices.
Past the allowance, Plus and Pro users can buy ChatGPT credits, and Business, Edu and Enterprise workspaces on flexible pricing buy workspace credits. At Standard speed, GPT-6.1 Sol uses 50 credits per million input tokens, 2.5 per million cached input tokens and 250 per million output tokens. OpenAI's pricing page gives no fixed dollar price per credit; it depends on your plan.
Your current limits are on the usage dashboard. Source: Codex pricing, checked 2026-10-17.
API PRICING
Standard short-context rates per million tokens, from OpenAI API pricing:
| API model | Input | Cached input (cache reads) | Output |
|---|---|---|---|
gpt-6.1-sol | $2.00 | $0.10 | $10.00 |
gpt-6-luna | $0.10 | $0.01 | $0.50 |
gpt-6-astra | $10.00 | $1.00 | $50.00 |
On gpt-6.1-sol, a prompt over 272K input tokens is billed at long-context rates for the whole request: $4.00 input, $0.20 cached input and $15.00 output (model page). Writing a prompt into the cache costs 1.25 times the uncached input rate (prompt caching); for gpt-6.1-sol that is $2.50 per million at the short-context rate. The API's Fast mode costs twice the standard rates.
Orbi turns GitHub Issues into reviewed pull requests and runs on Pi, a coding agent. Self-hosted, it can use your ChatGPT/Codex subscription through Pi's openai-codex provider (see the Codex guide), or your own key for any OpenAI-compatible endpoint. We run our own Orbi on DeepSeek, not Codex; the cost page shows what each merged PR cost us.
Sources: Codex pricing and OpenAI API pricing, checked 2026-10-17. Prices and limits change; check both pages before you buy.
AN HONEST READ
IF YOU ARE REPLACING CODEX
Pick by the boundary you cannot move: where the agent runs, how a task starts, and who carries it to a merge. Codex answers all three with a managed cloud sandbox and a ChatGPT, CLI, or GitHub entry point. Each option below keeps the same goal and moves one of those boundaries; Orbi, the subject of this page, is the one that turns a GitHub Issue into a reviewed pull request in a worktree on your own machine, with the model endpoint you choose.
Pick it when you want Anthropic's coding agent inside your own repository workflows: a GitHub Action that answers @claude mentions and can turn issues into pull requests.
Pick it when you want a hosted agent that stops for approval: Jules waits for the plan to be approved, then codes, and leaves you the branch and the pull request to open in GitHub.
Pick it when GitHub is already the queue: you can start work from an issue and the cloud agent opens one pull request per task, inside a session capped at 59 minutes.
Pick it when you want to run the agent yourself: a self-hosted control center that runs OpenHands, Codex, Claude Code, or any ACP-compatible agent locally, in Docker, on a VM, or in the cloud.
Pick it when you want an autonomous teammate in chat: Devin takes tasks from Slack, Teams, or its web app and can hand the longer ones to its cloud, leaving draft pull requests for review.
Pick it when you already work in Cursor: cloud agents run in their own virtual machine, build and test there, and open pull requests in the repositories they change, launched from the desktop, web, or iOS app.
Sources: Claude Code GitHub Actions; Google Jules — Running tasks; About GitHub Copilot cloud agent; OpenHands repository; Introducing Devin; Cursor — Cloud Agents — all verified 2026-10-06. Each alternative is described from its own documentation; Orbi's side is documented in the Orbi docs.
PROVENANCE
The facts on this page are a snapshot. The Codex sources were checked on 2026-09-07, the Codex CLI documentation on 2026-10-18, the Codex pricing sources on 2026-10-17, and the alternatives sources were added and checked on 2026-10-06 — re-check them before making an architectural decision.
PICK BY THE BOUNDARY
Choose Codex when the ChatGPT entry point and managed cloud sandbox are worth the OpenAI dependency. Choose Orbi when GitHub is the queue and local ownership, model choice, and auditability matter more than zero setup.
One email a week about one real delivery. Unsubscribe anytime.
