GILD verdict
Choose Codex if engineers want the agent in their editor and CI as well as the cloud.
Choose Devin if you want a multi-model agent or a dedicated single-tenant deployment. Answer.AI's month with Devin ended with 3 successes in 20 tasks in January 2025, while Goldman Sachs gives Devin to developer teams beside Copilot. Public merge data favors Codex. In a July 2026 study of popular Python repositories, with data to November 2025, 90.0% of Codex's agent pull requests merged against 37.8% of Devin's, and the study's model estimates merge probabilities of 73.5% and 43.0%, a significant gap. Either way, pilot on the same backlog and compare accepted, reviewed pull requests and total spend at the plan you would buy. A Devin Teams seat costs US$40 a month with a US$80 monthly minimum, US$15 to US$20 more than a ChatGPT Business seat at US$20 to US$25.
Pick on model choice and on where the work may run. Both take a task, work in a cloud environment and return a pull request, and both ship a local CLI and a desktop app. Codex runs OpenAI models in OpenAI-managed environments, and Devin runs Cognition's multi-model harness, with a single-tenant deployment option.
What teams report
Failures and incidents 1
Answer.AIOn Devin
Answer.AI gave Devin 20 real tasks over a month and recorded 14 failures, 3 successes and 3 inconclusive results, then finished failed tasks step by step in Cursor (Jan 8, 2025).
“The results were sobering. Out of 20 tasks, we had 14 failures, 3 successes (including our 2 initial ones), and 3 inconclusive results.”
Hamel Husain, Co-author, Answer.AIAnswer.AI, , write-up
Run both 1
Goldman SachsOn Devin
Goldman Sachs gives all its developers GitHub Copilot and began giving developer teams Devin in 2025 as a colleague to delegate work to, American Banker reported on Jul 25, 2025.
“All the developers have Github Copilot. They are getting the Devin technology. There's a shift from the developer assistant that helps you auto complete your code or generate code for a specific task, so I now have a virtual colleague that I can delegate work to.”
Marco Argenti, Chief Information OfficerAmerican Banker, , reportSecond-hand, as American Banker reported it
Evaluated 1
Prime IntellectOn Codex
Prime Intellect evaluated Codex on GPT-5.5 against Claude Code as autonomous agents on a public optimizer speedrun. Both ran ahead of the human record for most of the run, Claude Code finishing about 50 to 60 steps under it, and Codex worked almost without idling (talk of Sep 26, 2026).
“I think at the time the best uh record was like uh 2,990 step and we beat it by like uh uh 50 or 60 step for code and codeex was like 20 step above.”
Elie Bakouch, Research engineer, Prime IntellectAI Engineer, , at 11:13, from automatic captions
Choosing between Codex and Devin
Choose Codex if
- Engineers want one agent across the ChatGPT desktop app, the CLI and Codex Cloud, guided by the same AGENTS.md files.
- Engineers want the agent inside their current editor through the IDE extension, as well as in the cloud.
- You want CI automation through codex exec or the Codex GitHub Action.
Choose Devin if
- You want models from several providers, including Cognition's own SWE-2, behind one agent.
- Your security review requires the work environment in a single-tenant VPC connected over AWS PrivateLink.
- You want GPT usage in Devin billed to engineers' ChatGPT Go, Plus, or Pro plans.
Side by side
| Dimension | Codex | Devin |
|---|---|---|
| Where it runs | ChatGPT desktop app, CLI, IDE extension, and Codex Cloud | Devin Cloud sessions from the web, Slack, Teams, or API, plus Devin CLI and Devin Desktop |
| Individual priceList prices in US dollars. Included usage differs by plan and is not directly comparable. | Included in ChatGPT Plus: US$20 per month | Pro: US$20 per month; Max: US$200 per month |
| Team price | Included in ChatGPT Business: US$25 per user per month billed monthly, or US$20 billed annually (2+ users); Premium seats US$125, or US$100 billed annually, with five times the usage | Teams: US$40 per full seat per month, with a US$80 monthly minimum and free flex seats on shared credits |
| Beyond the plan | ChatGPT credits, or an API key billed at API rates | Prepaid on-demand credits that roll over; Enterprise is billed in ACUs at contract rates |
| Models | OpenAI models, including GPT-6.1 Sol and GPT-6 Luna | Models from Anthropic, OpenAI, Google, and open-source providers such as DeepSeek, Kimi, and GLM, plus Cognition's SWE models, with SWE-2 offered for a limited time in Devin Cloud's agent selector |
| Enterprise deployment | Codex Cloud in OpenAI-managed environments, with per-environment internet access settings | Multi-tenant Enterprise Cloud, or a dedicated single-tenant VPC; the agent's model service stays in Cognition's cloud |
| Project instructions | AGENTS.md, layered from the repository root down, up to 32 KiB combined by default | AGENTS.md, with the first 16 KiB of each file loaded automatically, plus Knowledge |
| Pull request review | @codex review on connected GitHub repositories | Devin Review on GitHub, GitLab, and connected Azure DevOps repositories |
| Estimated merge probability (model, closed PRs) | 73.5%, estimated in a July 2026 study of 309 Python repositories with 100+ stars, PRs to 20 November 2025; 5,451 of 6,057 merged raw, 90.0%, or 72.0% without the one repository that holds 3,981 of its rows. 64% in AIDev-pop, repositories with more than 500 stars, PRs to 22 June 20251819 | 43.0% in the same model, behind Codex (p<0.001); 512 of 1,355 merged raw, 37.8%; 49% in AIDev-pop, where Devin ran mostly in less popular repositories1819 |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
Leaders and teams rate Codex +55, 16 points above Devin at +39. Most of the gap is in complaints, with 4 of Devin's 24 sources negative against 2 of Codex's 33.
The score counts the authors from leaders and teams that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so Codex's +55 rests on votes split 19 positive, 10 mixed and 2 negative and Devin's +39 on 13, 6 and 4.
A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Use a task pilot to measure accepted work and review time.
Footprint on the Agentic Software Development stack
They differ on three of 12 layers. Codex owns two that Devin does not. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Layer | Codex | Devin |
|---|---|---|
| Build | ||
| Codebase context | Core | Supports |
| Agent runtime | Core | Supports |
| Validate | ||
| Application security | Touches | Supports |
The 9 layers where they match
| Layer | Codex | Devin |
|---|---|---|
| Plan | ||
| Work planning | Supports | Supports |
| Build | ||
| Coding | Core | Core |
| Tool execution | Core | Core |
| Validate | ||
| Debugging | Supports | Supports |
| Testing | Supports | Supports |
| Code review | Supports | Supports |
| Ship | ||
| Source control | Touches | Touches |
| CI/CD | Touches | Touches |
| Operate | ||
| Observability | Touches | Touches |
What practitioners praise
Among leaders and teams, Devin alone has four praise themes that reach 5 sources from 3 authors. Codebase understanding and Async delegation lead, at 11 of 11 and 9 of 9. Both carry Developer Experience. It is praised in 8 of the 14 Codex sources that discuss it, against 5 of 7 for Devin. Code Quality reaches 5 sources from 3 authors only for Codex, at 7 of 8. Devin needs a pilot on the same repository tasks before you choose on it.
| Praised for | Codex | Devin |
|---|---|---|
| Developer Experience | 8 of 14 | 5 of 7 |
| Code Quality | 7 of 8 | Not a theme in its sources |
| Codebase understanding | 3 of 3 so far | 11 of 11 |
| Async delegation | Not a theme in its sources | 9 of 9 |
| End-to-end execution | Not a theme in its sources | 7 of 8 |
| Migrations & repetitive changes | Not a theme in its sources | 5 of 5 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among leaders and teams. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Devin's sources raise Requires Strong Review more often than Codex's. It is raised in 1 of the 9 Codex sources that discuss it, against 7 of 8 for Devin. Usage Limits / Cost and Speed reach 5 sources from 3 authors only for Codex, at 7 of 13 and 3 of 5. Devin has 3 of 4 on Usage Limits / Cost so far and needs a pilot on the same repository tasks before you choose on Speed. Unpredictable task success reaches 5 sources from 3 authors only for Devin, at 3 of 6. Codex needs a pilot on the same repository tasks before you choose on it.
| Raised as a concern | Codex | Devin |
|---|---|---|
| Requires Strong Review | 1 of 9 | 7 of 8 |
| Usage Limits / Cost | 7 of 13 | 3 of 4 so far |
| Speed | 3 of 5 | Not a theme in its sources |
| Unpredictable task success | Not a theme in its sources | 3 of 6 |
19 sources
- Codex docs: Pricing
- ChatGPT Business pricing
- Codex docs: Models
- Codex docs: IDE extension
- Codex docs: Cloud
- Codex docs: Cloud environments
- Codex docs: AGENTS.md
- Codex docs: GitHub code review
- Codex docs: Codex GitHub Action
- Devin docs: Introducing Devin
- Devin docs: Self-serve plans
- Devin docs: Enterprise billing
- Devin docs: Use your ChatGPT plan in Devin
- Devin docs: Enterprise deployment
- Devin docs: Devin CLI
- Devin docs: AGENTS.md files for Devin
- Devin docs: Devin Review
- arXiv: How Do AI Coding Agents Contribute to Software Development? An Empirical Study of Agentic Pull Requests (Jul 23, 2026)
- arXiv: The Rise of AI Teammates in Software Engineering 3.0, the AIDev dataset (Jul 2025)