What engineering leaders and their teams report about Codex, with every finding tied to its evidence.
Read the findingsPractitioner sentiment, leaders and teams
+55 Strong positive from 33 sources that take a side
How this figure is madeFrom 31 authors in 33 discussions. Use a task pilot to measure accepted work and review time.
Apr 2025 to Oct 2026.
136 sources from practitioners, by month and sentiment.
Teams praise the code Codex writes in 7 of the 8 sources that discuss it, and the day-to-day experience in 8 of 14. Their cost complaints are surprise bills on team plans: credit use that spikes overnight past monthly caps, and overage billed above API rates. Quesma could not turn off fast mode for the whole organization.
Positive, mixed and negative sources
The score counts them as 19 positive, 10 mixed and 2 negative votes, one per author. See How the numbers work.
See the evidenceBelow: 2 praise themes, 3 concerns, 3 close to agreement and 3 disagreements.
Counted for Engineering leaders and teams, from 44 sources. Change group
Every capability practitioners describe positively, ranked by how many sources say so. Open one to read who said it.
Counted for Engineering leaders and teams.
Positive, 8
Positive
iforgotmypasswo, Hacker News commenter
Comment on: Gemini 4 ArgonHacker News comment
As an engineering manager, says US$500 a month for 40 or more hours a week of Astra has made their team's jobs much easier, and calls the Codex App Server amazing.
Practitioner account, pseudonymous; opinion; not counted, takes no position
Positive
Riley Collins, software developer at Atomic Object
Review: Codex Tools for Improving AccessibilityCompany blog post, Atomic Object (Atomic Spin)
With a broken wrist while moving to a project as its lone developer, relied on Codex dictation and appshots for several weeks to keep working. Dictation sometimes misheard words, and the developer expects to keep using it after healing.
"Over the past several weeks, I have found Codex tools for improving accessibility very useful…"
Practitioner account, name and role on their organization's own page; direct use; counts toward Operational control
Positive
Claire Morrison and Katie Fifer, Technically Curious
Codex Is ExcellentNewsletter post, Technically Curious (Substack)
Two long-time Claude Code users now favor Codex with GPT-5.5 for detailed work and long threads thanks to better compaction, while still preferring Claude for original design and writing.
Practitioner account, self-described; direct use; counts toward Delivered outcomes, Reliability, Operational control, Adoption and upkeep
Positive
Dan Shipper, co-founder and CEO of Every, with staff writer Katie Parrott
Vibe Check: We Tested OpenAI's New Codex AppNewsletter, Every
Every's team found the Codex desktop app the first coding GUI that feels like progress; the CEO moved to a 50-50 split with Claude Code, though one GM prefers terminal handoffs.
Practitioner account, self-described; direct use; not counted, takes no position
Positive
Nick Kilpatrick, software developer at Atomic Object
Get Into Agentic Development From the Shallow EndCompany blog post, Atomic Object (Atomic Spin)
Set up Codex command rules, sandbox settings and a guard hook to allow routine Git and project file edits, block recursive deletion and .env access, and review pushes; then logged approval requests via Codex hooks to propose allowlist rules he reviews and tests first.
"I implemented a similar model for Codex using its command rules, sandbox configuration, and a guard hook."
Practitioner account, name and role on their organization's own page; direct use; counts toward Operational control
Mixed
Katie Parrott, staff writer, Every
Every: Vibe Check, Opus 5.5 Is Pulling Our Codex Converts Back to ClaudeOrganization write-up
Two Every staff who had switched to Codex in recent months were reconsidering after Opus 5.5; one still prefers the Codex harness, and the CEO still uses Codex for about 80% of his work.
"I prefer the Codex harness, but I’d rather have a worse harness if I can work with a better model."
Mixed
Jacek Migdal, Quesma
OpenAI Codex pricing: the US$270 PR a US$200/month sub covers dailyCompany blog post, Quesma Blog
Quesma put Codex on OpenAI Business seats and one typical PR burned the US$270 overflow within a day; seat plans are far cheaper than tokens, but fast mode cannot be disabled org-wide.
"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."
Practitioner account, self-described; structured test; counts toward Operational control, Cost predictability and limits
Positive
Steve Sewell, co-founder and CEO of Builder.io
Codex vs Claude Code: which is the better AI coding agent?Company blog post, Builder.io blog
Builder.io's CEO prefers Codex mainly for its GitHub app, whose auto reviews find real bugs where Claude Code's GitHub reviews disappointed their dev team, plus pricing and limits.
Vendor material, self-described; direct use; counts toward Delivered outcomes, Operational control, Cost predictability and limits
Mixed or negative, 6
Mixed
Alexander Stolle, Director Data & AI at XING
Alexander Stolle - From Intentional Sprawl to a Deliberate Stack: Rolling Out CodexConference talk, Agentic Hamburg (YouTube)
XING consolidated the whole company on ChatGPT Enterprise and Codex, and accepted trade-offs from token limits to missing IDE autocompletion to get there.
"Codex IDE is not supporting autocomp completion at all and interestingly the the whole test group looking at Codex before they didn't check this at all so because no one of us was using this so right you're using codeex CLI or…"
Practitioner account, self-described; ran the rollout; counts toward Operational control
Mixed
iforgotmypasswo, choosing their company's primary AI provider
Comment on: I-have-ADHD: A skill to stop coding agents from burying the answerHacker News comment
After testing Astra, the author is switching the company's primary AI provider for its communication and judgment over Claude, while saying Codex has the worse UX.
Practitioner account, pseudonymous; direct use; not counted, takes no position
Mixed
jillesvangurp, developer at a company that helps small teams adopt agentic AI
Comment on: Claude for Small BusinessHacker News comment
A developer onboarding their non-developer CEO to Codex says it lets the CEO handle website and business tasks themselves, though it is painful because of git, repositories and permission friction.
Practitioner account, pseudonymous; ran the rollout; counts toward Operational control, Adoption and upkeep
Mixed
Mitch Alderson, Render
Testing AI coding agents (2025): Cursor vs. Claude, OpenAI, and GeminiIndependent benchmark, Render engineering blog
Scored Codex CLI 5 out of 10 on Render's vibe-coding test after 4 follow-up error prompts, and says its UX felt "somewhat primitive" next to the tools tested before it; in an existing service repository its connection retry code was idiomatic, easy to read and correct first try.
"Codex performed admirably, but the UX felt somewhat primitive compared to the other tools tested so far…"
Practitioner account; structured test; counts toward Delivered outcomes, Adoption and upkeep
Mixed
Beehive Innovations, makers of 2Do and BusyCal (GitHub organization account)
AGENTS.md is silently truncated and instructions near the end ignoredGitHub issue on openai/codex
Codex cut AGENTS.md at 32 KB, about 600 to 800 lines, without any warning, so instructions near the end never reached the model; Claude Code warns when CLAUDE.md is too large.
"Users assume Codex is unreliable or that AGENTS.md doesn't work properly, when in reality their instructions are being silently clipped."
Practitioner account, pseudonymous; direct use; counts toward Delivered outcomes
Positive
Dan Shipper, co-founder and CEO of Every
Vibe Check: Codex, OpenAI's New Coding AgentNewsletter, Every
Every's CEO shipped a Cora feature through Codex within an hour; the team found it rough but useful, best for tech leads working on existing codebases and less suited to new apps.
Practitioner account, self-described; direct use; counts toward Delivered outcomes, Operational control
Counted for Engineering leaders and teams.
Positive, 7
Positive
Naresh Jain, founder of Specmatic
Are Harnesses Worth the Cost?Article, martinfowler.com (Exploring Gen AI series)
Running Codex ten times on a detailed API brief, GPT-6 Sol at low effort passed all 178 acceptance tests in every run with or without a harness, and the harness made those runs about 7 times more expensive. GPT-6 Luna succeeded in three unharnessed runs and in every harnessed run, at 33% less per run than Sol.
"Using Codex with GPT-6 Sol at low reasoning effort, the agent produced an application that passed all 178 automated acceptance tests in each of the 10 runs, both with and without the harness."
Practitioner account, name and role on an outside page; structured test; counts toward Delivered outcomes
Positive
safety1st, Hacker News commenter
Comment on: Using an open model feels surprisingly goodHacker News comment
Their team runs Codex and Claude side by side; Codex produces far less unrequested code, which matters for human-reviewed, sensitive business logic.
Practitioner account, pseudonymous; direct use; not counted, repeats a stronger account
Positive
sreekanth850, Hacker News commenter
Comment on: GPT 6.1 Sol: Near-Astra intelligence for a fifth of the priceHacker News comment
Their team runs Codex in VS Code only on work they have agreed in an impact study and implementation plan with full regression tests. It is slow, and they report about 10 times the output with clean, consistent code.
Practitioner account, pseudonymous; direct use; counts toward Operational control
Positive
Claire Morrison and Katie Fifer, Technically Curious
Codex Is ExcellentNewsletter post, Technically Curious (Substack)
Two long-time Claude Code users now favor Codex with GPT-5.5 for detailed work and long threads thanks to better compaction, while still preferring Claude for original design and writing.
Practitioner account, self-described; direct use; counts toward Delivered outcomes, Reliability, Operational control, Adoption and upkeep
Mixed
James Reggio, CTO at Brex
Full Story of Brex’s AI Hail MaryPodcast episode show notes and transcript, Latent Space
Brex buys a few seats of several coding tools and lets engineers pick. Its CTO reports that engineers saw Codex as better at code generation but a little slower, and that fewer of them now switch, with some staying on Claude Code because they know it well.
"…when Codex came out and everybody was like, oh, Codex is better at CodeGen, but it’s a little bit slower."
Practitioner account, name and role on an outside page; opinion; not counted, takes no position
Positive
tunesmith, director-level engineering manager
Comment on: Your job is to deliver code you have proven to workHacker News comment
A director had Codex investigate a bug their LLM-averse team had blamed on a vendor; Codex prepared a PR, and the team reviewed it and agreed it was a valid fix.
Practitioner account, pseudonymous; direct use; not counted, repeats a stronger account
Mixed
Mitch Alderson, Render
Testing AI coding agents (2025): Cursor vs. Claude, OpenAI, and GeminiIndependent benchmark, Render engineering blog
Scored Codex CLI 5 out of 10 on Render's vibe-coding test after 4 follow-up error prompts, and says its UX felt "somewhat primitive" next to the tools tested before it; in an existing service repository its connection retry code was idiomatic, easy to read and correct first try.
"Codex performed admirably, but the UX felt somewhat primitive compared to the other tools tested so far…"
Practitioner account; structured test; counts toward Delivered outcomes, Adoption and upkeep
Mixed or negative, 1
Negative
trueno, Hacker News commenter
Comment on: Claude Opus 4.7Hacker News comment
Ran a three-day Codex pilot with a coworker; it fell well short of Claude on accuracy and problem solving and is not replacing their Claude tooling.
Practitioner account, pseudonymous; direct use; not counted, repeats a stronger account
Every concern practitioners raise, most raised first.
Counted for Engineering leaders and teams.
Raised as a concern, 7
Mixed
Alexander Stolle, Director Data & AI at XING
Alexander Stolle - From Intentional Sprawl to a Deliberate Stack: Rolling Out CodexConference talk, Agentic Hamburg (YouTube)
XING consolidated the whole company on ChatGPT Enterprise and Codex, and accepted trade-offs from token limits to missing IDE autocompletion to get there.
"Codex IDE is not supporting autocomp completion at all and interestingly the the whole test group looking at Codex before they didn't check this at all so because no one of us was using this so right you're using codeex CLI or…"
Practitioner account, self-described; ran the rollout; counts toward Operational control
Mixed
Jacek Migdal, Quesma
OpenAI Codex pricing: the US$270 PR a US$200/month sub covers dailyCompany blog post, Quesma Blog
Quesma put Codex on OpenAI Business seats and one typical PR burned the US$270 overflow within a day; seat plans are far cheaper than tokens, but fast mode cannot be disabled org-wide.
"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."
Practitioner account, self-described; structured test; counts toward Operational control, Cost predictability and limits
Mixed
chang-kr, GitHub user in an organization workspace
Codex credits consumed ~38,000 to 39,000 overnight with zero tasks issued: recurring bug, cap not enforced, no warning givenGitHub issue on openai/codex
In an organization workspace, Codex twice consumed about 38,000 to 39,000 credits against monthly caps of 5,000 and 10,000, once overnight with no tasks running. The company issued an internal notice about abnormal credit use.
"This appears to be affecting multiple accounts within our organization. An internal company-wide notice has been issued regarding abnormal Codex credit consumption."
Practitioner account, pseudonymous; direct use; counts toward Operational control, Cost predictability and limits
Mixed
Scott Chacon, co-founder of GitButler
Grit: rewriting Git in Rust with agentsCompany blog post, GitButler (Butler's Log)
GitButler's co-founder ported Git to Rust by running several coding agents for months. In /goal mode Codex kept working where Claude often stalled, though he found both goal modes slower than his own prompting loop, and his Codex subscription tokens ran out in the final days.
"Codex running in /goal mode seemed to keep working and getting stuff done."
Practitioner account, name and role on their organization's own page; direct use; counts toward Delivered outcomes, Reliability, Cost predictability and limits
Negative
z2, Hacker News commenter
Comment on: Outsourcing plus local AI will soon become more economical vs. frontier labsHacker News comment
The author's company saw OpenAI credit usage double on ChatGPT Business although only 2% of staff use Codex, with overage billed above API rates.
Practitioner account, pseudonymous; second-hand; not counted, repeats a stronger account
Mixed
submeta, Hacker News commenter
Comment on: AI subscriptions are a ticking time bomb for enterpriseHacker News comment
After a company rollout of ChatGPT with Codex, the author used it happily for two months until IT flagged 400 million tokens and a bill of at least a thousand euros.
Practitioner account, pseudonymous; direct use; counts toward Cost predictability and limits
Positive
Jayson DeLancey, Semgrep
Why Engineers Are Switching to CodexCompany blog post, Semgrep
Jayson DeLancey writes that a lot of tool shuffling has been landing on Codex for Semgrep's engineering team, with engineers citing less babysitting on certain types of tasks, cheap retries and AGENTS.md. One engineer now does 90% of their work through Codex; one described burning 94% of a five-hour limit.
"One engineer on our team switched about six months ago, around the GPT-5.5 release, and now does 90% of their work through Codex."
Vendor material; ran the rollout; counts toward Operational control, Adoption and upkeep
Discussed without concern, 6
Positive
Naresh Jain, founder of Specmatic
Are Harnesses Worth the Cost?Article, martinfowler.com (Exploring Gen AI series)
Running Codex ten times on a detailed API brief, GPT-6 Sol at low effort passed all 178 acceptance tests in every run with or without a harness, and the harness made those runs about 7 times more expensive. GPT-6 Luna succeeded in three unharnessed runs and in every harnessed run, at 33% less per run than Sol.
"Using Codex with GPT-6 Sol at low reasoning effort, the agent produced an application that passed all 178 automated acceptance tests in each of the 10 runs, both with and without the harness."
Practitioner account, name and role on an outside page; structured test; counts toward Delivered outcomes
Positive
iforgotmypasswo, Hacker News commenter
Comment on: Gemini 4 ArgonHacker News comment
As an engineering manager, says US$500 a month for 40 or more hours a week of Astra has made their team's jobs much easier, and calls the Codex App Server amazing.
Practitioner account, pseudonymous; opinion; not counted, takes no position
Mixed
Elie Bakouch, research engineer at Prime Intellect
We Let Claude Code and Codex Race Human Researchers - Elie Bakouch, Prime IntellectConference talk, AI Engineer (YouTube)
In a multi-day research speedrun Codex never went idle and beat the human record, though it burned more tokens than Claude Code and invented no new optimizer.
"…and then I was just like okay continue explore new direction hey just go again for 10 hours and then say yeah I cannot beat the recall and so on. So basically onethird of the time the cloud code agent was idle…"
"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…"
Practitioner account, self-described; structured test; counts toward Delivered outcomes
Positive
sreekanth850, Hacker News commenter
Comment on: Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?Hacker News comment
Their three-person team left Claude Code, which was slow and failed to fix issues, for Codex; with two Plus accounts per person, work is easier and free of usage anxiety.
Practitioner account, pseudonymous; direct use; counts toward Delivered outcomes, Cost predictability and limits
Positive
Rogerio Chaves, Co-founder and CTO at LangWatch
Codex vs Claude Code: Which Is the Better Background Agent? (2026)Company engineering blog post, LangWatch
After months of running both at LangWatch, Chaves keeps Claude Code as the builder for its background agents. Codex is the reviewer, surfacing the most P1 and P2 findings, such as a spend query missing its tenant filter. Automatic Codex review on every pull request stays off because of cost.
"After months of running both, our preference has been boringly consistent: Claude Code is better at doing the work, and Codex is the better reviewer of it."
Practitioner account, name and role on their organization's own page; ran the rollout; counts toward Delivered outcomes, Operational control
Positive
Steve Sewell, co-founder and CEO of Builder.io
Codex vs Claude Code: which is the better AI coding agent?Company blog post, Builder.io blog
Builder.io's CEO prefers Codex mainly for its GitHub app, whose auto reviews find real bugs where Claude Code's GitHub reviews disappointed their dev team, plus pricing and limits.
Vendor material, self-described; direct use; counts toward Delivered outcomes, Operational control, Cost predictability and limits
GILD's fact checks, which don't count as sources
Checked Oct 12, 2026.Codex limits vary by model and task. OpenAI estimates 15 to 160 local messages per five hours on GPT-6.1 Sol for Plus and Business, and Pro plans have no five-hour limit. Credits extend usage past the limit. A report of running out does not establish what another plan or model would allow. Pricing
Open them in the recordCounted for Engineering leaders and teams.
Raised as a concern, 3
Mixed
James Reggio, CTO at Brex
Full Story of Brex’s AI Hail MaryPodcast episode show notes and transcript, Latent Space
Brex buys a few seats of several coding tools and lets engineers pick. Its CTO reports that engineers saw Codex as better at code generation but a little slower, and that fewer of them now switch, with some staying on Claude Code because they know it well.
"…when Codex came out and everybody was like, oh, Codex is better at CodeGen, but it’s a little bit slower."
Practitioner account, name and role on an outside page; opinion; not counted, takes no position
Negative
Alex Ghiculescu, co-founder of Tanda
AI coding: plateauing but also acceleratingNewsletter post, Alex Ghiculescu (Substack)
Tanda's co-founder tried Codex in the cloud and found the feedback loop too slow to work with. Some prompts return a pull request that needs no changes, but any follow-up seems to rebuild the whole environment, which he calls a night and day difference from Claude Code on his machine.
"The implementation is so slow that it feels unworkable."
Practitioner account, self-described; direct use; counts toward Delivered outcomes
Mixed
Scott Chacon, co-founder of GitButler
Grit: rewriting Git in Rust with agentsCompany blog post, GitButler (Butler's Log)
GitButler's co-founder ported Git to Rust by running several coding agents for months. In /goal mode Codex kept working where Claude often stalled, though he found both goal modes slower than his own prompting loop, and his Codex subscription tokens ran out in the final days.
"Codex running in /goal mode seemed to keep working and getting stuff done."
Practitioner account, name and role on their organization's own page; direct use; counts toward Delivered outcomes, Reliability, Cost predictability and limits
Discussed without concern, 2
Mixed
Alexander Stolle, Director Data & AI at XING
Alexander Stolle - From Intentional Sprawl to a Deliberate Stack: Rolling Out CodexConference talk, Agentic Hamburg (YouTube)
XING consolidated the whole company on ChatGPT Enterprise and Codex, and accepted trade-offs from token limits to missing IDE autocompletion to get there.
"Codex IDE is not supporting autocomp completion at all and interestingly the the whole test group looking at Codex before they didn't check this at all so because no one of us was using this so right you're using codeex CLI or…"
Practitioner account, self-described; ran the rollout; counts toward Operational control
Positive
Steve Sewell, co-founder and CEO of Builder.io
Codex vs Claude Code: which is the better AI coding agent?Company blog post, Builder.io blog
Builder.io's CEO prefers Codex mainly for its GitHub app, whose auto reviews find real bugs where Claude Code's GitHub reviews disappointed their dev team, plus pricing and limits.
Vendor material, self-described; direct use; counts toward Delivered outcomes, Operational control, Cost predictability and limits
GILD's fact checks, which don't count as sources
Checked Oct 12, 2026.GPT-6.1 Sol runs in Standard and Fast modes where supported, and in Ultrafast on Pro US$500 and eligible Enterprise and Edu plans; faster modes draw on usage limits at a higher rate. Pricing
Open them in the recordCounted for Engineering leaders and teams.
Raised as a concern, 1
Positive
johnjwang, Hacker News commenter testing Codex with their team at Assembled
Comment on the Hacker News launch thread for Codex (May 2025)Hacker News comment
Says Codex shines at running dozens of small edits in parallel, like an unlimited pool of junior engineers, though output still needs work to be production ready.
Practitioner account, pseudonymous; direct use; not counted, repeats a stronger account
Discussed without concern, 8
Positive
tunesmith, director-level engineering manager
Comment on: Your job is to deliver code you have proven to workHacker News comment
A director had Codex investigate a bug their LLM-averse team had blamed on a vendor; Codex prepared a PR, and the team reviewed it and agreed it was a valid fix.
Practitioner account, pseudonymous; direct use; not counted, repeats a stronger account
Positive
Jayson DeLancey, Semgrep
Why Engineers Are Switching to CodexCompany blog post, Semgrep
Jayson DeLancey writes that a lot of tool shuffling has been landing on Codex for Semgrep's engineering team, with engineers citing less babysitting on certain types of tasks, cheap retries and AGENTS.md. One engineer now does 90% of their work through Codex; one described burning 94% of a five-hour limit.
"One engineer on our team switched about six months ago, around the GPT-5.5 release, and now does 90% of their work through Codex."
Vendor material; ran the rollout; counts toward Operational control, Adoption and upkeep
Positive
ndriscoll, Hacker News commenter
Comment on: Google fixed more Chrome bugs in June than over the past two years, thanks to AIHacker News comment
An engineer built a team AI docs repo and a listener so Codex, triggered by PagerDuty alerts, investigates code and dashboards and posts bug-fix PRs to Slack for the on-caller to review.
Practitioner account, pseudonymous; ran the rollout; counts toward Delivered outcomes, Adoption and upkeep
Positive
safety1st, Hacker News commenter
Comment on: Using an open model feels surprisingly goodHacker News comment
Their team runs Codex and Claude side by side; Codex produces far less unrequested code, which matters for human-reviewed, sensitive business logic.
Practitioner account, pseudonymous; direct use; not counted, repeats a stronger account
Positive
Rogerio Chaves, Co-founder and CTO at LangWatch
Codex vs Claude Code: Which Is the Better Background Agent? (2026)Company engineering blog post, LangWatch
After months of running both at LangWatch, Chaves keeps Claude Code as the builder for its background agents. Codex is the reviewer, surfacing the most P1 and P2 findings, such as a spend query missing its tenant filter. Automatic Codex review on every pull request stays off because of cost.
"After months of running both, our preference has been boringly consistent: Claude Code is better at doing the work, and Codex is the better reviewer of it."
Practitioner account, name and role on their organization's own page; ran the rollout; counts toward Delivered outcomes, Operational control
Positive
Someone1234, Hacker News commenter
Comment on: Ask HN: Who has seen productivity increases from AIHacker News comment
A team feeds legacy app definitions through a multi-step Codex pipeline that outputs roughly 90% working React pages; developers fix and review, raising throughput from one to three pages a week.
Practitioner account, pseudonymous; ran the rollout; counts toward Delivered outcomes
Positive
Kelsey Brennan, software developer at Atomic Object
Why I’m Sold on AI for Code ReviewsCompany blog post, Atomic Object (Atomic Spin)
Kelsey Brennan's client project team runs Codex through codex exec in GitHub Actions to review pull requests against the project's written rules. Over a few months, 60 to 70 percent of its comments were solid findings worth addressing, though a person still needs project context to judge them.
"We have a GitHub Actions workflow that generates the prompt and passes it to Codex using codex exec…"
"…over the last few months I’ve found that 60 to 70% of the AI’s comments are solid findings that I choose to address…"
Practitioner account, self-described; direct use; counts toward Delivered outcomes
Positive
Ryan Cox, Senior Director of Infrastructure and Security at Temporal
Improving our Java SDK with Codex by OpenAIEngineering blog post, Temporal blog
Temporal's engineers used Codex to clear long-standing Java SDK issues, batching backlog tasks before lunch or end of day, and added AGENTS.md files to prepare repos for agents.
Practitioner account, self-described; direct use; counts toward Delivered outcomes
Statements this group's sources agree on, then the ones close to agreement.
Close to agreement: each needs 9 supporting sources
Codex is evaluated as an agent that carries engineering tasks to a pull request, with planning and review around it, and never as a completion tool.
7 supporting sourcesA maintained AGENTS.md, shared skills, and written review rules make Codex's results more consistent across engineers and sessions.
6 supporting sourcesThe most reliable reported workflows give Codex a well-scoped repository task with checks it can run, such as tests that define done.
5 supporting sourcesQuestions the sources split on, with the count on each side.
Evidence remains mixed
Codex as the main tool, 7 sources
Some moved their main work to Codex for code that follows instructions closely and reviews that catch real bugs.
Claude Code as the main tool, 7 sources
Others keep Claude Code for front-end and design work or because it plans and asks before building, and use Codex for review or second opinions.
Evidence remains mixed
Local CLI, IDE, or app, 6 sources
Most work through the CLI, the IDE extension, or the desktop app on their own machine, where they can watch and steer each change.
Delegated cloud tasks, 2 sources
Others hand well-specified tasks to Codex Cloud and review the pull requests later, often several at a time.
Evidence remains mixed
Broader delegation, 4 sources
Some give Codex long unattended runs, overnight or across many issues, once acceptance checks are written down.
Tighter steering, 2 sources
Others plan first and work in small, supervised steps so wrong assumptions do not spread through larger changes.
The three groups, then every practitioner by the context they work in, and how emphasis changes with scale.
Three groups drawn from the same sources. Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. Signals open on Engineering leaders and teams.
44 sources. Leaders' accounts cover shared conventions, review load, spend as adoption spreads, and what changes when Codex moves from individual use into a team or company standard.
92 sources. Builders' accounts cover code quality, speed, plan limits, and how Codex performs in self-directed development.
136 sources. This group gives the broadest market read. It mixes individual work with team rollouts, and the two can reach different conclusions.
136 sources from practitioners, by the context they work in. Filled rows count for this group.
Emphasizes Code quality and plan limits
Emphasizes How much to delegate and what reaches review
Emphasizes Review load and shared AGENTS.md conventions
Emphasizes Which vendor to standardize on, and at what cost
Emphasizes Managed defaults and company-wide spend
What changes with scale. Solo developers weigh the code Codex writes against how fast they hit their plan's caps. Team and organization accounts focus on how much of a task Codex carries and what that costs at company scale.
Where this group's sources were published.
Counted for Engineering leaders and teams, from 44 sources. Change group
65 sources for this group, 21 of them documentation
Older evidence stays when it still applies. Sources tied to a changed plan, limit, or model remain marked Needs a recheck.
See sources by originWho each source is, how it knows, whether it still applies and what it counts toward.
Newest first, on its own page. Every count above opens its sources there.
Read all 157 sourcesAbout this read
A running review of what engineering teams report about Codex, from individual builders to company-wide rollouts. Each finding links to its sources, and each conflict shows both sides.
Counts cover only the sources GILD reviewed.
Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. GILD files each source in a group from how the account describes its author. A role or employer marked self-described is what the source states about themselves. Documentation checks facts for every group and never counts toward sentiment.
A source's date and whether it still applies are separate: a recent post can describe an old version, and an old one can still hold.
Flagged for a recheck, of the sources that take a side: leaders and teams 6 of 33, individual builders 22 of 73, all practitioners 28 of 106.
Each row opens its sources for the group you are reading.
| Finding | Engineering leaders and teams | Individual builders | All practitioners | Sources |
|---|---|---|---|---|
| Current readCurrent sentiment | 21 positive, 10 mixed, 2 negative | 38 positive, 25 mixed, 10 negative | 59 positive, 35 mixed, 12 negative | Sources |
| PraiseCode Quality | 7 of 8 positive | 20 of 25 positive | 27 of 33 positive | Sources |
| PraiseDeveloper Experience | 8 of 14 positive | 12 of 22 positive | 20 of 36 positive | Sources |
| PraiseLong-Running Autonomy | 3 of 3 positive (under the minimum) | 9 of 14 positive | 12 of 17 positive | Sources |
| PraisePull Request Review | 3 of 3 positive (under the minimum) | 9 of 9 positive | 12 of 12 positive | Sources |
| PraiseCodebase Understanding | 3 of 3 positive (under the minimum) | 7 of 8 positive | 10 of 11 positive | Sources |
| PraiseValue for Money | 3 of 4 positive (under the minimum) | 8 of 11 positive | 11 of 15 positive | Sources |
| PraiseFast Runs | 0 of 1 positive (under the minimum) | 6 of 11 positive | 6 of 12 positive | Sources |
| PraiseStability | 0 of 1 positive (under the minimum) | 1 of 5 positive | 1 of 6 positive | Sources |
| ConcernUsage Limits / Cost | 7 of 13 raise it | 15 of 26 raise it | 22 of 39 raise it | Sources |
| ConcernRequires Strong Review | 1 of 9 raise it | 8 of 14 raise it | 9 of 23 raise it | Sources |
| ConcernNeeds Close Steering | 0 of 0 raise it (under the minimum) | 6 of 9 raise it | 6 of 9 raise it | Sources |
| ConcernSpeed | 3 of 5 raise it | 10 of 17 raise it | 13 of 22 raise it | Sources |
| ConcernInconsistent Results | 2 of 2 raise it (under the minimum) | 13 of 14 raise it | 15 of 16 raise it | Sources |
| ConcernCommunication & Steering | 0 of 2 raise it (under the minimum) | 7 of 10 raise it | 7 of 12 raise it | Sources |
| ConsensusWritten conventions make results consistent | 6 supporting (below threshold) | 8 supporting (below threshold) | 14 supporting (below threshold) | Sources |
| ConsensusStrongest on scoped, repository-level work | 5 supporting (below threshold) | 12 supporting (below threshold) | 17 supporting (below threshold) | Sources |
| ConsensusSeveral tasks at once | 3 supporting (under the minimum) | 15 supporting (below threshold) | 18 supporting (below threshold) | Sources |
| ConsensusEvaluated as an engineering agent | 7 supporting (below threshold) | 8 supporting (below threshold) | 15 supporting (below threshold) | Sources |
| ConsensusFaster on self-directed work | 2 supporting (under the minimum) | 5 supporting (below threshold) | 7 supporting (below threshold) | Sources |
| DisagreementAmong practitioners who use both, is Codex or Claude Code the better main tool? | Codex as the main tool 7, Claude Code as the main tool 7 | Codex as the main tool 14, Claude Code as the main tool 14 | Codex as the main tool 21, Claude Code as the main tool 21 | Sources |
| DisagreementShould Codex work run on the engineer's machine or as delegated cloud tasks? | Local CLI, IDE, or app 6, Delegated cloud tasks 2 | Local CLI, IDE, or app 25, Delegated cloud tasks 11 | Local CLI, IDE, or app 31, Delegated cloud tasks 13 | Sources |
| DisagreementHow much should Codex complete before a human checkpoint? | Broader delegation 4, Tighter steering 2 | Broader delegation 7, Tighter steering 5 | Broader delegation 11, Tighter steering 7 | Sources |
| Current readEvidence mix | 65 sources | 113 sources | 157 sources | Sources |
GILD's checks of claims against official documentation, kept apart from the source counts.
