Codex+55 Strong positive

Operator Insights

What are engineering leaders learning from deploying Codex?

16 accounts: 1 failure, incident or rollback, 2 switched, 6 run it beside another tool, 2 evaluated it, 1 adopted it and 4 reported by others

Failures, incidents and rollbacks1

Armin RonacherDeveloper, creator of FlaskEarendil
One developer's unattended agent experiment, September 2026.
Armin Ronacher's Thoughts and Writings, published Sep 7, 2026Read the write-up (opens Armin Ronacher's Thoughts and Writings in a new tab)
Observation
A 35-hour unattended Codex run burned about US$1,200 of tokens and delivered nothing of value.
Takeaway
Long unattended runs fail without checks the agent must pass along the way. Cap spend and review progress well before day two.
What happened
Ronacher left the Codex harness with GPT-6 Astra running unattended for 35 hours. It burned around 1 billion tokens, about US$1,200 at raw API prices, and produced 79 commits, and he judged that it delivered nothing of value (Sep 7, 2026).
“35 hours later, the factory has delivered absolutely nothing of value…”
Armin Ronacher
- Public practitioner write-up
- Self-reported, with token and cost figures
Deployment details
One developer's unattended agent experiment, September 2026.
- Public practitioner write-up
- Self-reported, with token and cost figures
Team
One developer
Duration
35 hours
Workload
Unattended agent build
Product version
Codex harness
Model
GPT-6 Astra
Measurement
Tokens, raw API cost and commits

Switched to or from it2

Katie ParrottStaff writerEvery
Every's staff across evals, design and leadership, in the months before Sep 22, 2026.
every.to, published Sep 22, 2026Read the write-up (opens every.to in a new tab)
Observation
Every's staff had switched to Codex in recent months, and by Sep 22, 2026 Opus 5.5 had them reconsidering.
Takeaway
The Codex harness holds users on its own merits, and the model race can still pull them away at each release.
What happened
Every's head of evals Mike Taylor and designer Tyler Nishida had switched from Claude Code to Codex in recent months, and by Sep 22, 2026 Opus 5.5 had them reconsidering. Taylor still prefers the Codex harness but would take a worse harness for a better model, and CEO Dan Shipper still splits his use 80/20 toward Codex.
“I prefer the Codex harness, but I’d rather have a worse harness if I can work with a better model.”
Mike Taylor, Every
- Public organization write-up
- Self-reported
Deployment details
Every's staff across evals, design and leadership, in the months before Sep 22, 2026.
- Public organization write-up
- Self-reported
Team
Every staff
Duration
Recent months before Sep 22, 2026
Workload
Product and editorial engineering
Product version
Codex and Claude Code
Model
GPT-6.1 Sol era; Opus 5.5

Anurag SinghJul 27, 2026XDA DevelopersA daily Claude user moved his coding to Codex on ChatGPT Plus for more room before the limit.
Anurag SinghTech journalistXDA Developers
One writer's own coding projects, July 2026.
Its own announcement, published Jul 27, 2026Read the write-up (opens the announcement in a new tab)
Observation
A daily Claude user moved his coding to Codex on ChatGPT Plus for more room before the limit.
Takeaway
At the US$20 tier, Codex's allowance lets a heavy user finish more work per session.
What happened
After six months of daily Claude Code use, Singh switched his coding to Codex with GPT-5.6 Sol on the US$20 ChatGPT Plus plan in July 2026. He can work much longer before hitting the limit, so he finishes more projects without moving them to another agent.
“However, I’ve been able to work for much longer before hitting the limit, which has given me a better chance of finishing a project without moving it to another coding agent.”
Anurag Singh
Public practitioner write-up
Self-reported
Deployment details
One writer's own coding projects, July 2026.
Public practitioner write-up
Self-reported
Team
One writer
Workload
Personal coding projects
Product version
Codex on ChatGPT Plus
Model
GPT-5.6 Sol

Run it beside another tool6

DHHCTO37signals
DHH's own development on Omarchy and 37signals work, August 2026.
Lex Fridman Podcast, published Aug 26, 2026Listen to the episode (opens Lex Fridman Podcast in a new tab)
Observation
At 37signals, Codex at xhigh effort reviews every job Claude Code finishes.
Takeaway
Codex earns a seat as the reviewer even on teams whose main agent is another vendor's.
What happened
DHH drives his agents in Claude Code and ends every job with a Codex review at xhigh effort, his standard procedure, calling Codex his favorite checker because it keeps finding problems. He runs Codex in a pane beside Claude, and sometimes OpenCode too.
“Codex is very good. I use it all the time. It’s my favorite checker. It keeps finding stuff. It’s great. But one of the things I’ve found that I like so much about the current Claude models is they’re really good writers out of the box.”
DHH, at 3:23:46 in the recording
- Podcast episode
- Self-reported
Deployment details
DHH's own development on Omarchy and 37signals work, August 2026.
- Podcast episode
- Self-reported
Team
DHH's own development
Workload
Omarchy and 37signals code
Product version
Codex beside Claude Code
Model
Codex at xhigh effort reviewing

SuperhumanAug 18, 2026Merantix AI CampusSuperhuman's code reviewer calls Codex for a second opinion on complex pull requests.
Yaroslav (Yarik)Tech lead for the coding-agent rolloutSuperhuman
Superhuman engineering's internal code reviewer, August 2026.
Merantix AI Campus, published Aug 18, 2026Watch the talk (opens Merantix AI Campus in a new tab)
Observation
Superhuman's code reviewer calls Codex for a second opinion on complex pull requests.
Takeaway
A second model family on hard reviews catches what one vendor's agents share as blind spots.
What happened
Superhuman's internal reviewer runs a Claude Code coordinator over a swarm of review agents and launches Codex as a second reviewer on especially complex pull requests. Superhuman keeps maintaining it in place of Anthropic's Code Review because it can tailor it.
“…we have a Claude code coordinator that launches um a swarm of agents and also for especially complex merge request or pull request or merge request it launches Codex for a second opinion.”
Yaroslav (Yarik), Superhuman, at 20:46, from automatic captions
Conference talk
Self-reported
Deployment details
Superhuman engineering's internal code reviewer, August 2026.
Conference talk
Self-reported
Team
Superhuman engineering
Workload
Internal code review
Product version
Codex beside Claude Code
DatabricksJun 24, 2026Latent SpaceDatabricks runs Codex and Claude Code behind one internal wrapper.
Matei ZahariaCo-founderDatabricks
Databricks developer infrastructure, June 2026.
Latent Space, published Jun 24, 2026Listen to the episode (opens Latent Space in a new tab)
Observation
Databricks runs Codex and Claude Code behind one internal wrapper.
Takeaway
Wrapping both agents keeps the model choice open without two separate rollouts.
What happened
Databricks' developer infrastructure team built Isaac, a wrapper on Codex and Claude Code that lets engineers run either agent in web sandboxes or on their own machines.
“…they built something called Isaac, that’s like a wrapper on Claude Code and Codex, and, lets you use them either on the web in, like, sandboxes or, just on your dev machine or on your laptop or whatever.”
Matei Zaharia, Databricks, at 2:22 in the recording
Podcast episode
Self-reported
Deployment details
Databricks developer infrastructure, June 2026.
Podcast episode
Self-reported
Team
Databricks engineering
Workload
Databricks product development
Product version
Codex and Claude Code behind Isaac
BetterUpNov 24, 2025AI EngineerBetterUp's AI enablement lead added Codex beside Claude Code to avoid lock-in.
Brian JohnPrincipal full-stack engineer, AI enablement for R&DBetterUp
BetterUp R&D AI enablement, November 2025.
AI Engineer, published Nov 24, 2025Watch the talk (opens AI Engineer in a new tab)
Observation
BetterUp's AI enablement lead added Codex beside Claude Code to avoid lock-in.
Takeaway
Codex fits unattended, hands-off tasks; plan extra setup to reproduce sub-agent workflows from other tools.
What happened
John added the Codex CLI beside Claude Code so he is not locked in to one tool or one model family, and built a sub-agent workaround for it. He finds Codex slower and serial, built for hands-off unattended work, and its sandbox resisted the workaround.
“Codeex does not have the ability to run things asynchronously like claw does. So, this is slower. And Codeex in general, if you've used it, I think you find it's slower overall than than Cloud Code. But I think that's really kind of intentional. seems like Codex is really kind of meant to be more of like a hands-off unattended type of a tool versus clog code is meant to be more kind of iterative…”
Brian John, BetterUp, at 11:28, from automatic captions
Conference talk
Self-reported
Deployment details
BetterUp R&D AI enablement, November 2025.
Conference talk
Self-reported
Team
BetterUp R&D
Workload
AI enablement across R&D
Product version
Codex CLI beside Claude Code
MetaJan 18, 2026Lenny's PodcastA Meta product manager has Codex, Claude Code and Cursor each review the same code.
Zevi ArnovitzProduct managerMeta
His own projects, January 2026.
Lenny's Podcast, published Jan 18, 2026Listen to the episode (opens Lenny's Podcast in a new tab)
Observation
A Meta product manager has Codex, Claude Code and Cursor each review the same code.
Takeaway
Running two or three agents as reviewers on the same change catches more than any one of them.
What happened
Arnovitz builds his own app with agents and runs multiple reviews: Codex, Claude Code and Cursor each review the same code, and a peer-review command reconciles their findings.
“But what's even cooler and something that I'm really proud of is I will usually do multiple reviews and I'll have Codex, which is ChatGPT's competitor to Claude Code, as well as cursor open, and I will have each of them review the code.”
Zevi Arnovitz, at 39:37 in the recording
Podcast episode
Self-reported
Deployment details
His own projects, January 2026.
Podcast episode
Self-reported
Team
One product manager's own projects
Workload
A personal study app
Product version
Codex beside Claude Code and Cursor
EveryJul 17, 2025Lenny's PodcastIn 2025, Every used Codex beside Claude Code for one-off, self-contained features.
Dan ShipperCo-founder and CEOEvery
Every, company-wide, July 2025.
Lenny's Podcast, published Jul 17, 2025Listen to the episode (opens Lenny's Podcast in a new tab)
Observation
In 2025, Every used Codex beside Claude Code for one-off, self-contained features.
Takeaway
Codex often enters as the agent for isolated tasks and grows from there.
What happened
Shipper stated in July 2025 that everyone at Every who builds something uses Claude Code, and that the team also uses Codex beside it for one-off, self-contained features.
“We also use Codex a bit, which is OpenAI's coding tool. And that's for, like, "I want a one-off, self-contained... I want to pick off this little feature."…”
Dan Shipper, at 37:17
Podcast episode
Self-reported
Deployment details
Every, company-wide, July 2025.
Podcast episode
Self-reported
Team
Every
Workload
Product engineering
Product version
Codex beside Claude Code

Evaluated it2

QuesmaAug 28, 2026Quesma BlogOn ChatGPT Business, one typical pull request at Quesma used US$270 of overflow within 24 hours.
Jacek MigdalAuthor of the Quesma blog postQuesma
Quesma on ChatGPT Business seats, August 2026.
Quesma Blog, published Aug 28, 2026Read the write-up (opens Quesma Blog in a new tab)
Observation
On ChatGPT Business, one typical pull request at Quesma used US$270 of overflow within 24 hours.
Takeaway
Set a credit limit before the first week of Business seats, and tell engineers how fast mode bills.
What happened
Quesma tested Codex with real money on ChatGPT Business seats at US$25 a month with US$270 of overflow allowed. One typical pull request used it all within 24 hours; US$110 of the US$283 day was the fast-mode premium, and fast mode cannot be turned off across a Business organization (Aug 28, 2026).
“On the downside, it is super easy to land in fast mode. I accidentally confirmed it and ran the first half of this change at double rates: $110 of the $283 above was the fast-mode premium. Even worse, it is impossible to disable fast mode across an entire OpenAI Business organization, while Claude requires organizational approval before anyone can turn it on.”
Jacek Migdal
Public engineering write-up
Self-reported, with the bill
Deployment details
Quesma on ChatGPT Business seats, August 2026.
Public engineering write-up
Self-reported, with the bill
Team
Quesma
Duration
One day
Workload
A typical pull request
Product version
Codex on ChatGPT Business
Measurement
Overflow spend in US dollars
Prime IntellectSep 26, 2026AI EngineerIn Prime Intellect's research race, Codex almost never sat idle and burned more tokens.
Elie BakouchResearch engineerPrime Intellect
Automated research on Prime Intellect's cluster, about July 2026.
AI Engineer, published Sep 26, 2026Watch the talk (opens AI Engineer in a new tab)
Observation
In Prime Intellect's research race, Codex almost never sat idle and burned more tokens.
Takeaway
Codex suits work that has to run without someone watching, and its token use needs a budget to match.
What happened
Prime Intellect evaluated Codex on GPT-5.5 against Claude Code on Opus 4.8 as autonomous agents on a public optimizer speedrun. Both ran ahead of the human record for most of the run, and Claude Code finished about 50 to 60 steps under the 2,990-step record; Codex worked almost continuously, spawned more sub-agents and burned more tokens, while Claude Code sat idle about a third of the time.
“So basically onethird of the time the cloud code agent was idle because I had no way to basically monitor it and codeex totally the opposite just worked for all the all the time and uh yeah almost never idle never asked for question…”
Elie Bakouch, Prime Intellect, at 8:12, from automatic captions
Conference talk
Measured on one task
Deployment details
Automated research on Prime Intellect's cluster, about July 2026.
Conference talk
Measured on one task
Team
Prime Intellect research
Duration
One speedrun
Workload
Optimizer speedrun
Product version
Codex and Claude Code
Model
GPT-5.5 and Opus 4.8
Measurement
Steps to the target loss, idle time

Adopted it1

XINGSep 23, 2026Agentic Hamburg on YouTubeXING consolidated a sprawl of AI tools on ChatGPT Enterprise and Codex, and accepted token limits and lost autocompletion.
Alexander StolleDirector Data & AIXING
Company-wide rollout. ChatGPT Enterprise and Codex. Talk at Agentic Conf Hamburg, March 22, 2026.
Agentic Hamburg on YouTube, published Sep 23, 2026Watch the talk (opens Agentic Hamburg on YouTube in a new tab)
Observation
XING consolidated a sprawl of AI tools on ChatGPT Enterprise and Codex, and accepted token limits and lost autocompletion.
Takeaway
Name the trade-offs you accept before you consolidate, and judge the stack the way you judge laptops or Slack, since Stolle argues DORA metrics will not prove its value.
What happened
From 2023 to 2025 XING let many AI tools run side by side, then chose one stack for business users and engineers alike: ChatGPT Enterprise with Codex. Stolle describes a pull-based rollout, AGENTS.md files in the repositories, plain CLIs in place of complex MCP setups, and Codex defaults managed centrally on every laptop. XING accepted trade-offs from token limits to the loss of IDE autocompletion, and Stolle argues agentic tools amplify whatever state an engineering organization is already in.
Conference talk (recorded Mar 22, 2026)
No figures captured from the talk
Deployment details
Company-wide rollout. ChatGPT Enterprise and Codex. Talk at Agentic Conf Hamburg, March 22, 2026.
Conference talk (recorded Mar 22, 2026)
No figures captured from the talk
Team
XING, business users and engineers company-wide
Workload
Product engineering across XING repositories
Product version
ChatGPT Enterprise with Codex, defaults managed centrally

Reported by others4

NVIDIAApr 23, 2026NVIDIA BlogNVIDIA gave every employee a cloud VM for the Codex agent, with read-only access to production systems.
Justin BoitanoVice President and General Manager, Enterprise ComputingNVIDIA
Organization-wide. Over 10,000 employees across engineering, legal, finance, sales, and other functions. GPT-5.5 in the Codex app.
NVIDIA Blog, published Apr 23, 2026Read the write-up (opens NVIDIA Blog in a new tab)
Observation
NVIDIA gave every employee a cloud VM for the Codex agent, with read-only access to production systems.
Takeaway
Give each agent its own audited machine and read-only production access before a company-wide rollout, and ask for numbers behind any speedup claim.
What happened
NVIDIA opened GPT-5.5 in the Codex app to over 10,000 employees, engineers and non-engineers alike. NVIDIA IT gave every employee a cloud virtual machine for the agent, reached from the Codex app over SSH, so agents work with company data inside an audited sandbox. A zero-data-retention policy covers the deployment, and agents reach production systems read-only through command-line tools and Skills. Boitano reports that debugging that took days now closes in hours.
“NVIDIA IT rolled out cloud virtual machines (VMs) for every employee to run their agent safely.”
Justin Boitano
Public organization write-up
No metrics published
Model stated
Deployment details
Organization-wide. Over 10,000 employees across engineering, legal, finance, sales, and other functions. GPT-5.5 in the Codex app.
Public organization write-up
No metrics published
Model stated
Team
Over 10,000 NVIDIA employees across functions
Duration
A few weeks at publication
Workload
Debugging, experiments in multi-file codebases, and knowledge work outside engineering
Product version
Codex app connected over SSH to per-employee cloud VMs
Model
GPT-5.5
TemporalMay 17, 2025Temporal blogTemporal engineers queued Codex tasks from the backlog before lunch and reviewed the results after it.
Ryan CoxSenior Director of Infrastructure and SecurityTemporal
Open-source SDK and server repositories. Codex cloud agent, early access. May 2025.
Temporal blog, published May 17, 2025Read the write-up (opens Temporal blog in a new tab)
Observation
Temporal engineers queued Codex tasks from the backlog before lunch and reviewed the results after it.
Takeaway
Hand an agent the backlog items with a clear issue and an existing pattern to copy, and write the build, test, and review commands into AGENTS.md before you do.
What happened
Invited to test Codex, Temporal pointed it at Java SDK work it had not found time for: code coverage, documentation, and features its other SDKs already had. Given a GitHub issue and a short summary, Codex implemented the API to count workflow executions by search attributes, with an end-to-end test. The change still needed classes renamed and the test cleaned up. Engineers picked a few backlog issues before lunch or before leaving for the day, started several Codex tasks, and iterated on the results later. Temporal added AGENTS.md files with build, test, and review rules to its repositories.
“After giving Codex the Github issue and a brief summary of the requirements, it was able to implement the API, including adding an end-to-end test.”
Ryan Cox
Public engineering write-up
Linked pull requests, no metrics
Deployment details
Open-source SDK and server repositories. Codex cloud agent, early access. May 2025.
Public engineering write-up
Linked pull requests, no metrics
Team
Temporal engineers working on the SDKs and server
Workload
Java SDK features, bug fixes, coverage, and documentation, plus server and sdk-core changes
Product version
Codex cloud agent at launch
James FarisJul 29, 2026businessinsider.comDisney is preparing to adopt Codex while dropping Copilot, Business Insider reported.
James FarisReporterBusiness Insider
Disney's US technology organization, August 2026.
businessinsider.com, published Jul 29, 2026Read the report (opens businessinsider.com in a new tab)
Observation
Disney is preparing to adopt Codex while dropping Copilot, Business Insider reported.
Takeaway
Codex is entering large buyers as a second agent beside Claude, in place of the assistants staff ignored.
What happened
Business Insider reported on Jul 29, 2026, from an internal message, that Disney is preparing to adopt Codex, with no timeline given, while dropping GitHub Copilot, Kiro and Amazon Q for US staff in August and keeping Claude Enterprise and Cursor.
“Disney is preparing to adopt OpenAI's Codex coding tool, according to the screenshot. The message didn't give a timeline for the tool's introduction.”
James Faris
Second-hand news report (Business Insider, from a Disney internal message)
No timeline stated
Deployment details
Disney's US technology organization, August 2026.
Second-hand news report (Business Insider, from a Disney internal message)
No timeline stated
Team
Disney technology staff, US
Workload
Software development across Disney
Product version
Codex, beside Claude Enterprise and Cursor
Uday Kiran MedisettyOct 14, 2026The Pragmatic EngineerUber supports Codex beside Claude Code and OpenCode behind one wrapper.
Uday Kiran MedisettyDistinguished EngineerUber
Uber's AI Foundations developer platform, October 2026.
The Pragmatic Engineer, published Oct 14, 2026Read the report (opens The Pragmatic Engineer in a new tab)
Observation
Uber supports Codex beside Claude Code and OpenCode behind one wrapper.
Takeaway
Keep skills and settings in a layer the company owns, so moving to or from Codex costs engineers nothing.
What happened
The Pragmatic Engineer reported on Oct 14, 2026 that Uber supports Codex, Claude Code and OpenCode as harnesses. AIFX, an internal wrapper over all three, moves an engineer's settings, skills and plugins when they switch harness.
“Uber currently supports Claude Code, Codex, and OpenCode as harnesses. AIFX (short for AI Foundations) is a wrapper on top of all of these, so when an engineer changes harnesses, all their settings, skills and plugins move with them.”
The Pragmatic Engineer
Second-hand news report (The Pragmatic Engineer)
Organization platform described
Deployment details
Uber's AI Foundations developer platform, October 2026.
Second-hand news report (The Pragmatic Engineer)
Organization platform described
Team
Uber engineering
Workload
Uber's developer platform
Product version
Codex, Claude Code and OpenCode behind AIFX

Failures, incidents and rollbacks1

A 35-hour unattended Codex run burned about US$1,200 of tokens and delivered nothing of value.

Switched to or from it2

Every's staff had switched to Codex in recent months, and by Sep 22, 2026 Opus 5.5 had them reconsidering.

A daily Claude user moved his coding to Codex on ChatGPT Plus for more room before the limit.

Run it beside another tool6

At 37signals, Codex at xhigh effort reviews every job Claude Code finishes.

Superhuman's code reviewer calls Codex for a second opinion on complex pull requests.

Databricks runs Codex and Claude Code behind one internal wrapper.

BetterUp's AI enablement lead added Codex beside Claude Code to avoid lock-in.

A Meta product manager has Codex, Claude Code and Cursor each review the same code.

In 2025, Every used Codex beside Claude Code for one-off, self-contained features.

Evaluated it2

On ChatGPT Business, one typical pull request at Quesma used US$270 of overflow within 24 hours.

In Prime Intellect's research race, Codex almost never sat idle and burned more tokens.

Adopted it1

XING consolidated a sprawl of AI tools on ChatGPT Enterprise and Codex, and accepted token limits and lost autocompletion.

Reported by others4

NVIDIA gave every employee a cloud VM for the Codex agent, with read-only access to production systems.

Temporal engineers queued Codex tasks from the backlog before lunch and reviewed the results after it.

Disney is preparing to adopt Codex while dropping Copilot, Business Insider reported.

Uber supports Codex beside Claude Code and OpenCode behind one wrapper.