TODO for AI
vs 
OpenAI Codex
By Marcell Havlik, founder of TODO for AI · Updated
OpenAI's coding agent: an open-source CLI plus a cloud sandbox inside ChatGPT. Here is where each one wins, with numbers, and who should pick which.
| Fact | OpenAI Codex | TODO for AI |
|---|---|---|
| Interface | Terminal, ChatGPT web, IDE | Web, desktop, mobile, CLI, IDE (Zed, JetBrains, VS Code), voice |
| Models | OpenAI only | Any provider: Claude, GPT, Gemini, Grok, Ollama |
| Source | Partly open | Edge + CLI public |
| Starts at | $8/mo | Free |
Key takeaways
- Models. OpenAI Codex model choice: OpenAI only. TODO for AI runs any provider, Claude, GPT, Gemini, Grok or a local Ollama model, on one plan, and one model can review another's work.
- Benchmark. TODO for AI scored 89.9% on Terminal-Bench 2.1 with Opus 5.5 (high) (own harness, 80/89 tasks, 2026-10-08; not an official leaderboard submission). OpenAI Codex is listed at 84.3% with Opus 5.5 (high) on the public leaderboard.
- Price. OpenAI Codex pricing: Included with ChatGPT: Go $8, Plus $20, Pro $100 or $200. Usage limits per window. TODO for AI has a free plan (Claude Sonnet only), then Starter $20, Pro $100 or Ultra $200 a month with any provider included.
- Where the work happens. OpenAI Codex gets one task done in one person's terminal. TODO for AI is one board your team and agents work through together, on your own PC, a cloud VM and your logged-in Chrome.
- Who should pick which. Pick OpenAI Codex if you are OpenAI-only by policy. Pick TODO for AI if you want to run GPT-5.6 and have Claude review it.
OpenAI Codex
OpenAI only
TODO for AI
Any providerWe are not competing for the same job
OpenAI Codex gets the task done.
TODO for AI organises the whole company’s work.
OpenAI Codex answers one question: how do I get this task done, in one person's terminal? So do the others — Cursor inside the editor, Claude Code in one person's session, Lindy by asking its assistant.
We answer a different one: where does the work live? One board per project, and a project for each part of the company — engineering, marketing, ops, finance, one per client. Your colleagues work on it. So do your agents, as many as you want to run, each with its own model, tools and permissions. Items carry the people who worked them and the agent that ran them, and the board filters by person, so what the company has in flight is one screen instead of a Slack thread.
That is not a team plan bolted onto an AI tool. It is the product. The agents are how items get closed; the board is what the company runs on.
What OpenAI Codex is, and where it stops
Codex is OpenAI's answer to Claude Code: an open-source CLI for local work and a cloud sandbox you launch from ChatGPT for parallel tasks. On GPT-5.6 it is unusually token-efficient, and being bundled into ChatGPT Plus makes it the cheapest way for an existing OpenAI customer to get an agent.
The trade-offs mirror Claude Code's. OpenAI models only, sandboxes that are fresh every task and scoped to a repo, and nothing for the browser or the business side. The CLI being open is genuinely useful; the cloud agent that does the heavy lifting is not.
Teams already on ChatGPT Plus or Pro who want GPT-native coding.
Three reasons OpenAI Codex users look elsewhere
No Claude, no Gemini, no way to cross-check output with a different model family.
Each cloud task starts clean. Logins, installed tools and half-finished state do not carry over.
It is built to produce pull requests. Anything outside a git repository is not a task it can take.
Six dimensions that decide it
OpenAI Codex 14/30TODO for AI 29/30TODO for AI · Opus 5.5 (high)89.9%OpenAI Codex · Opus 5.5 (high)84.3%Same 89-task set, one trial each. Ours excludes the tasks the model refused on safety grounds (neither pass nor fail); theirs is their public leaderboard entry. Different model too, so treat it as indicative, not a controlled A/B.
How does OpenAI Codex compare to TODO for AI, feature by feature?
| Feature | OpenAI Codex | TODO for AI |
|---|---|---|
Persistent cloud VM An always-on machine. Files, logins and sessions survive between runs. | Ephemeral sandbox | Yes |
Model choice | OpenAI only | Any provider |
Cross-vendor review Review GPT output with Claude, or the other way around. | No | Yes |
Drives your real browser Browser extension (Chrome, Firefox) plus cloud browsers. | No | Yes |
Business tasks, not only code SEO, email, ads, social, CRM via CLI tools. | No | Yes |
Any CLI as an integration | MCP | Any binary |
Voice control | No | Yes |
People and agents on one board Colleagues and agents work the same list. Items show who contributed and which agent ran them; filter the board by person. | No | Yes |
Many agents, each configured Each agent gets its own model, prompt, tools and permissions. Pick which one takes an item. | Profiles + AGENTS.md | Unlimited, per task |
Projects for the whole company Engineering, marketing, ops, finance: separate projects and agents, one team and one bill. | Repo folders | Unlimited, shared |
Team roles, pooled usage | Per seat | Roles + pooled |
Tool permissions | Basic | Allow, approve, block |
Open runtime | CLI only | Edge + CLI public |
Same ladder, different meter
Message and feature caps per plan, counted per surface. Deep research and agent runs have their own quotas.
One pool priced in dollars of model cost, spendable on any vendor. Session and weekly windows pace it, and a 1.3-1.5x plan multiplier applies, so a $20 plan is about $15 of list-price API. Team seat prices go into one shared pool.