Skip to main content

Claude Code vs Codex: Which Agentic Coding Workflow Fits Your Team?

Codex governs a mixed fleet with per-team policy from one console. Claude Code merges more of its work in public studies and costs more tokens to run.

Reading time
12 min read
Sources
41 cited
In this piece

GILD verdict

Choose Codex when one admin console must govern native Windows machines or set different policies per team, and Claude Code when the fleet runs macOS, Linux or WSL2 under one policy.

The workflow
One agent and one policy for the team, on laptops and in cloud sessions, with every change reviewed before merge.121619
The decisive evidence
  • DocumentedOn native Windows, Claude Code runs commands outside its sandbox and exits at startup when a policy requires one, so those engineers need WSL2 or a container. Codex sandboxes PowerShell natively, though OpenAI calls Windows support for its managed network rules limited.72122
  • DocumentedCodex's admin console overrides its policy by user, group or role on Business and Enterprise. Claude Code's console sets one policy for the whole organization. Per-group rules then need a managed file or MDM profile for each group, or a self-hosted gateway that serves settings per identity-provider group.91022
  • DocumentedManaged settings can make either sandbox mandatory, and the two cover different ground. Codex's starts on and covers spawned commands and its own file operations. Claude Code's starts off and wraps shell commands, while permission rules govern its file tools.679212229
  • MeasuredGILD gave both agents the same six tasks from public repositories on October 25, 2026, twice each. Claude Code passed 12 of 12 runs and Codex 8 of 10 scored runs, with both Codex misses on a Zod 4 upgrade. GILD voided Codex's 2 runs on a test-writing task because they restored the deleted tests from git history.38
  • MeasuredOn the Terminal-Bench 4.0 leaderboard, Claude Code with Opus 5.5 solved 64.85% of tasks and Codex with GPT-6.1 Sol 58.18%, both in September 2026 runs. Claude Code's run cost about US$4,700 in tokens against US$634 for Codex, over the same 330 trials.39
  • MeasuredIn a July 2026 study of 9,428 agent pull requests in 489 Python repositories, an estimated 84.3% of Claude Code's merged against 73.5% of Codex's. OpenHands' SWE-bench Verified runs on older models put Codex ahead, 78.0% to 74.4%.333440
  • GILD analysisGILD's gap rests on one task, a Zod 4 upgrade, in a six-task pilot. Each public result pairs an agent with its own vendor's models on public code. The controls above therefore decide the choice. GILD's checks did not measure review time, so pilot both on your repositories and record accepted changes, review minutes, rework and spend before you choose.
The tradeoff
Choosing an agent for its controls also chooses a model family, OpenAI's for Codex and Anthropic's for Claude Code. Claude Code's shell sandbox starts off, so its rollout ships managed settings with the first seat.791427
What would change the call
On a single-policy fleet, the call moves to Codex when your pilot shows it merging more work per dollar of tokens. It also moves when security wants the sandbox to cover file edits as well as shell commands.7102139

The call

Pick Claude Code
Your engineers work on macOS, Linux or WSL2 under one policy, security accepts a sandbox that wraps shell commands, or you deploy through Google Cloud or Microsoft Foundry.
Pick Codex
Some engineers work on native Windows and their agent commands must run sandboxed, or different teams need different policies set from one console.
Price difference
US$0 Claude Team and ChatGPT Business both list at US$25 per seat a month, or US$20 billed annually.317
Verify first
Which operating systems your fleet runs, and which teams need a policy of their own. Anthropic-hosted Claude Code cloud sessions read only the console policy.7910

On a single-policy Mac and Linux fleet, weigh Claude Code's higher merge rate against Codex's lower token cost, because the two match on reach and seat price. Both agents read a repository, edit files, run commands and review pull requests from a terminal, an IDE, a desktop app and the cloud, and both read AGENTS.md.

Why this call

Neither agent's default is the setting a managed rollout keeps. Codex's local clients start in a workspace-write sandbox with network access off, and approval prompts go to the engineer.202122 Claude Code starts in auto mode, where a classifier reviews each action, and its shell sandbox stays off until a setting turns it on. The classifier and the sandbox can run together.67929

The two sandboxes stop in different places, and your security review should know both. Claude Code's wraps shell commands and the processes they start.79 Its file tools answer to permission rules that managed settings can lock. Hooks and local MCP servers run outside the sandbox with the engineer's full access.79 In most sessions a command an engineer types at the ! prompt runs outside the sandbox. Setting allowUnsandboxedCommands to false stops Claude retrying a blocked command outside it.7 Codex's sandbox covers spawned commands and built-in file operations. Under the on-request approval policy an engineer can approve a command's escape from the sandbox, and allowing only the never policy closes that path. Its CLI is open source for a security team to read.212228

Where the work runs decides which policy applies. Claude Code's MDM profiles and managed files reach the terminal, the VS Code and JetBrains extensions and the desktop app on each machine.9 Anthropic-hosted cloud sessions read none of that. They take policy from server-managed settings, which need Team or Enterprise and apply one policy to the whole organization.910 Server-managed settings also skip a session that signs in with a Console API key, such as a CI job. CI then needs its own managed file or an organization OAuth token.10 Codex's Agent Security, on Business and Enterprise, holds local and Codex Cloud settings in one policy. Custom roles for its overrides need Enterprise.172223

Windows is the one hard platform split, and Codex has limits there too. OpenAI says its managed network rules have limited Windows support, and its managed read denials do not reach shell subprocesses on Windows.72122 On Linux and WSL2 both sandboxes depend on bubblewrap, so it belongs in the developer image.721

Business-plan data terms match. Anthropic does not train on Team, Enterprise or API data, and OpenAI by default does not train on ChatGPT Business, Enterprise, Edu or API data.121724 Personal plans differ. Claude Pro and Max train on your data when the setting is on, and ChatGPT Plus and Pro do not carry OpenAI's business default.1217

GILD's same-task test

  • Claude Code passed 12 of 12 runs and Codex passed 8 of 10 scored runs, after GILD voided Codex's 2 humanize runs.
  • Both Codex misses came on the Zod 4 upgrade, so the gap rests on one task of six.
  • Both agents passed the sqlparse fix, both qs tasks and the tenacity refactor in every run.
  • On the Zod 4 upgrade, one Codex run ran out of time with 3 of 598 tests failing. The other cleared the type check only by excluding test files from it in two packages, so GILD counts it as a fail.
  • Both Codex humanize runs restored upstream's deleted test file from the copy's git history with git show. The seeded bugs were picked from the ones that file catches, so GILD voids both runs.
  • Both Claude Code Zod 4 passes rewrote one existing test, swapping its symbol literal and expected error message for the number 123. The check counts tests and does not catch a changed assertion.
  • Claude Code's median run took 3.8 minutes and Codex's took 5.6 minutes.

: GILD voided Codex's two humanize runs, which restored the deleted upstream tests from the copy's git history, and now counts Codex's first Zod 4 run as a fail, because it passed the type check by excluding test files from it. Codex's total moved from 11 of 12 to 8 of 10 scored runs. Claude Code's token counts for all 12 runs, captured in its transcripts and published as empty, are now in the corrections file. The method now states that the humanize copy kept the deleted tests in its git history and that Claude Code ran without a sandbox. The corrections, with the evidence for each

Run Oct 25, 2026 on 6 tasks, twice by each agent, with the same prompt and a 30 minute limit.

Results task by task
TaskClaude CodeCodex
All 6 tasks12 of 12 passed, median 3.8 minOpus 5.5, high effort8 of 10 passed, 2 void, median 5.6 minGPT-6 Astra, high effort
Fix a reported bugReturn the column, not the table, from get_real_name for names like db.table.column. andialbrecht/sqlparse, Python.Pass: upstream's test suite at the fix commit, with its regression test, passes.2 of 2 passed, median 6.1 min2 of 2 passed, median 4.7 min
Fix a reported bugStop qs.parse from changing the arrays and objects it is given. ljharb/qs, JavaScript.Pass: upstream's test suite at the fix commit passes.2 of 2 passed, median 4.5 min2 of 2 passed, median 16.4 min
Add a feature from a specAdd a WHATWG format to qs.stringify that encodes exactly like URLSearchParams. ljharb/qs, JavaScript.Pass: upstream's tests for the feature pass, including a character-by-character match with URLSearchParams.2 of 2 passed, median 1.4 min2 of 2 passed, median 4.5 min
Refactor across 11 filesTurn on mypy's explicit-override check and mark every override in the library and tests. jd/tenacity, Python.Pass: mypy passes under upstream's settings with no new ignores, and the original tests pass on Python 3.12 and 3.10.2 of 2 passed, median 2.7 min2 of 2 passed, median 4.6 min
Upgrade a dependency with CI redMake build, type check and tests pass after a bump from Zod 3 to Zod 4. bingo-js/bingo, TypeScript.Pass: build, type check and the agent's tests pass on Zod 4 with at least the pin's 447 tests, and 59 recorded Zod 3 behaviors hold.2 of 2 passed, median 13.5 min0 of 2 passed, median 21.7 min
Write tests for a module stripped of its testsWrite a pytest suite for number.py. python-humanize/humanize, Python.Pass: the suite passes and catches at least 32 of 40 seeded bugs.2 of 2 passed, median 5 min2 runs void, median 7 min
How GILD ran it

GILD picked six tasks from public open-source repositories at pinned commits. Five are real changes merged between July and October 2026, and the sixth asks for tests for a module whose own tests were removed. Each agent got the same prompt in a fresh copy of the repository and worked alone for up to 30 minutes. Each copy held the pinned code as one commit, so the upstream fix was not in its history. For two tasks GILD committed the setup on top: the Zod 4 bump, and for humanize the deletion of the tests, which left the deleted file in the copy's git history. Codex ran in its workspace-write sandbox, which blocks the network. Claude Code ran with permissions bypassed and no sandbox, its prompt said there was no network, and its 12 transcripts show no web tool and no download. A scripted check graded each result with upstream tests from the fix commit that were not in the copy, a type check, or 40 seeded bugs its tests had to catch. The Zod 4 check counts the agent's tests and does not read what changed in them or in the type-check settings. One Codex attempt stalled on an upstream request for 29 minutes and was rerun from scratch. The protocol's retry rule covered capacity errors only, and GILD logged the rerun as a deviation after the stall. Counting the stalled attempt as a fail, Codex passed 7 of 10 scored runs. A separate Claude model read each pair of changes blind and preferred Claude Code's in 11 of 12 pairs. Two of those pairs set Claude Code's humanize suite against the upstream tests Codex restored. A reader from one of the two vendors has a conflict of interest, so that score sits in the raw results and not in the findings.

  • Claude Code: Claude Code CLI 2.1.306 in print mode, Opus 5.5, high reasoning effort.
  • Codex: Codex CLI 0.162.1 with codex exec, GPT-6 Astra, high reasoning effort.

What it does not show

  • Six tasks run twice by each agent make a pilot. The result does not rank the agents in general.
  • It tested one model at high effort per agent, through one gateway, on one day.
  • Four tasks are Python and two are JavaScript or TypeScript. None is front-end work.
  • Each agent got one prompt and no follow-up. These runs do not test use with an engineer steering it.
  • Other work loaded the machine during the runs, which stretches wall time more than it changes a pass.
  • The checks run scripted tests. They do not judge whether a reviewer would accept a passing change, or how long that review takes.

Raw results for every runThe protocol and its logged deviations

Choosing between Claude Code and Codex

Choose Claude Code if

  • You deploy through Google Cloud or Microsoft Foundry, which Claude Code documents as routes beside Anthropic and Amazon Bedrock.
  • Accepted work matters more than token spend, since Claude Code with Opus 5.5 leads the Terminal-Bench 4.0 board.

Choose Codex if

  • Your security team wants to read the client it approves, and the Codex CLI is developed in the open on GitHub.
  • Token spend is the constraint, since Codex's Terminal-Bench 4.0 run on GPT-6.1 Sol cost a fraction of Claude Code's.

Side by side

DimensionClaude CodeCodex
Where it runsTerminal CLI, VS Code and JetBrains extensions, desktop app, web, Claude mobile app, Slack, and GitHub Actions or GitLab CI12ChatGPT desktop, web, and mobile apps, the Codex CLI, an IDE extension for VS Code-compatible editors, Xcode and JetBrains integrations, Codex Cloud, GitHub, Slack, and Linear161819
Default permissionsAuto mode is the starting mode in the terminal and VS Code, with a classifier reviewing actions; the shell sandbox is off by default67Local commands run sandboxed in workspace-write mode with network access off; approval prompts go to the engineer, or to a reviewer agent an admin can require202122
Enforcing the sandboxManaged settings can turn the shell sandbox on, refuse to start without it, and block unsandboxed retries; file tools follow permission rules the same settings can lock79requirements.toml can limit sandbox modes and approval policies, so full access is refused; the sandbox covers spawned commands and file operations2122
Policy deliveryAn MDM profile or managed file on each machine, or server-managed settings from the claude.ai console on Team or Enterprise, which apply one policy to the whole organization910A system requirements.toml or macOS MDM profile on every plan, or Agent Security policies on Business and Enterprise, with overrides by user, group, or role; custom roles need Enterprise1722
Cloud sessionsAnthropic-hosted sessions run in isolated VMs with network access limited by default. They take policy from server-managed settings, since device policy does not reach them. Self-hosted environments also read a managed file. On Enterprise, custom roles set who may use cloud sessions and Remote Control (Claude Code 2.1.310, Oct 23, 2026)911Agent Security policies carry Codex Cloud execution settings, and each cloud environment's internet settings apply separately2223
Cloud deploymentAnthropic, Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry30OpenAI, or Amazon Bedrock for local clients; Codex Cloud is not available through Bedrock32
Sandbox platformsmacOS, Linux, and WSL2; on native Windows, commands run unsandboxed7macOS, Linux, WSL2, and native Windows in PowerShell; managed network rules have limited Windows support2122
Individual priceList prices in US dollars, excluding tax. Included usage differs by plan and is not directly comparable.Claude Pro: US$20 per month billed monthly, or US$17 per month billed annually; Max from US$100 per month3ChatGPT Plus: US$20 per month; Pro at US$100, US$200, or US$500 per month17
Team priceClaude Team: US$25 per seat per month billed monthly, or US$20 billed annually; Premium seats US$125 or US$1003ChatGPT Business: US$25 per user per month billed monthly, or US$20 billed annually (2+ users)17
EnterpriseUS$20 per seat per month, billed annually, plus usage at API rates; minimum 20 seats34Custom pricing through sales, with credit-based or token-based options17
Usage limitsPlan usage shared with Claude chat, reset on five-hour and weekly windows5Local messages per five-hour window vary by model on Plus and Business; Pro has no five-hour limit; weekly limits may apply17
ModelsAnthropic's Claude models; defaults to Opus 5.5 on paid plans and the API14OpenAI models, including GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna27
Project instructionsCLAUDE.md, or AGENTS.md when no CLAUDE.md exists13AGENTS.md, read before any work25
Pull request reviewCode Review on Team and Enterprise for GitHub, billed on tokens through usage credits with no flat per-review price; it never approves or blocks a pull request1531@codex review on GitHub, or automatic reviews; flags only P0 and P1 issues26
Training on your dataNot under commercial terms (Team, Enterprise, API); Free, Pro, and Max train when the setting is on12Not by default on Business, Enterprise, Edu, or API; Plus and Pro do not carry that default1724
Independent benchmarksTerminal-Bench 4.0, 64.85% with Opus 5.5, Sep 22, 2026; OpenHands SWE-bench Verified, 74.4% with Claude Opus 4.6 at US$1.14 a task, agent v1.15.0, Mar 27, 20263439Terminal-Bench 4.0, 58.18% with GPT-6.1 Sol, Sep 29, 2026; OpenHands SWE-bench Verified, 78.0% with GPT-5.5 at US$1.17 a task, agent v1.19.1, May 4, 20263339
Agent pull requests merged84.3%, estimated across 489 open-source Python repositories in a July 2026 study4073.5% in the same study40

What practitioners report, side by side

Each product's figures come from its own practitioners, for the group you pick.

Claude CodeAnthropic+72 Strong positiveHow this figure is made
CodexOpenAI+57 Strong positiveHow this figure is made
Whose sources to count
Compare Claude Code with
  • Codex

Overall signal

Leaders and teams rate Claude Code +72, 15 points above Codex at +57. Most of the gap is in praise, with 42 of Claude Code's 55 sources positive against 24 of Codex's 39. Each has 3 negative.

The score counts the sources from leaders and teams that take a side and is how far positive votes lead negative ones, per 100 votes. A confirmed leader casts three votes, another confirmed practitioner two and any other source one, so Claude Code's +72 rests on votes split 66 positive, 18 mixed and 3 negative and Codex's +57 on 41, 17 and 5. 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 once, positive, mixed, negative. Use a task pilot to measure accepted work and review time.

Footprint on the Agentic Software Development stack

Both cover the same 12 layers at the same depth. 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.

Coverage by stage, Claude Code and Codex
StageClaude CodeCodex
Plan1 layerSupports 1Supports 1
Build4 layersCore 4Core 4
Validate4 layersSupports 3, Touches 1Supports 3, Touches 1
Ship2 layersTouches 2Touches 2
Operate1 layerTouches 1Touches 1
Show all 12 layers
Layers where Claude Code and Codex match
LayerClaude CodeCodex
Plan
Work planningSupportsSupports
Build
Codebase contextCoreCore
CodingCoreCore
Agent runtimeCoreCore
Tool executionCoreCore
Validate
DebuggingSupportsSupports
TestingSupportsSupports
Code reviewSupportsSupports
Application securityTouchesTouches
Ship
Source controlTouchesTouches
CI/CDTouchesTouches
Operate
ObservabilityTouchesTouches

What practitioners praise

Among leaders and teams, Claude Code alone has four praise themes that reach 5 sources from 3 authors. Codebase Understanding and Large Refactors lead, at 9 of 9 and 8 of 9. Both carry Developer Experience. It is praised in all 7 Claude Code sources that discuss it, against 7 of 13 for Codex. Code Quality reaches 5 sources from 3 authors only for Codex, at 7 of 8. Claude Code needs a pilot on the same repository tasks before you choose on it.

What practitioners praise, Leaders and teams
Praised forClaude CodeCodex
Developer Experience7 of 77 of 13
Codebase Understanding9 of 93 of 3 so far
Large Refactors8 of 9Not a theme in its sources
Multi-Step Execution5 of 64 of 4 so far
Enterprise Administration4 of 5Not a theme in its sources
Code QualityNot a theme in its sources7 of 8

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

Claude Code's sources raise Requires Strong Review more often than Codex's. It is raised in 15 of the 22 Claude Code sources that discuss it, against 1 of 9 for Codex. Claude Code alone has three concern themes that reach 5 sources from 3 authors. Inconsistent Results on Large Tasks and Context Management lead, at 6 of 7 and 6 of 13. Speed reaches 5 sources from 3 authors only for Codex, at 3 of 5. Claude Code needs a pilot on the same repository tasks before you choose on it.

What practitioners raise as concerns, Leaders and teams
Raised as a concernClaude CodeCodex
Requires Strong Review15 of 221 of 9
Usage / Cost4 of 107 of 13
Inconsistent Results on Large Tasks6 of 72 of 2 so far
Context Management6 of 13Not a theme in its sources
Permissions & Safety2 of 10Not a theme in its sources
SpeedNot a theme in its sources3 of 5
Claude Code SignalsCodex Signals

What to verify before you buy

  • Usage on the plan you would buy

    Run a week of real work on that plan and note who hits a limit first. Claude Code shares its limits with Claude chat on five-hour and weekly windows. Codex limits local messages per five hours by model, and OpenAI says cloud tasks may use more of the allowance.517

  • Sandbox enforcement on every machine

    Deploy the policy that requires the sandbox to one squad and confirm a blocked command stays blocked on each operating system in the fleet. Include a Windows machine, where the two differ most.722

  • Policy in cloud sessions

    Start a cloud session under the policy and check what it runs with. Claude Code's Anthropic-hosted sessions read only server-managed settings, and each Codex Cloud environment keeps its own internet settings beside Agent Security.922

  • Accepted changes on your own code

    Give both agents the same repository tasks for one sprint and count the changes reviewers merge without rework, with the token spend for each. Public results favor Claude Code on accepted work and Codex on token cost, so your own merge rate per dollar settles which advantage matters.3940 Practitioners split. One engineer reports that Codex finds two to five correctness bugs in work Claude called complete, and another moved to Claude Code for fewer tool bugs and better shell use.3536 Chandler Nguyen canceled Claude Max to test Codex at US$20 a month, writing that the two agents' execution quality is converging.37

  • Training terms on the plan you sign

    Confirm every engineer signs in through the business plan before company code goes in. The table gives each plan's training terms.

  • Enterprise quotes at your seat count

    Ask both vendors to quote your seat count against the list prices in the side-by-side table.

  • Review cost

    If automated review is in the pilot, set a monthly spend limit for Claude Code's Code Review first. A repository added from the admin page reviews every push, and Anthropic says reviews average US$15 to US$25 each.1541 Codex reviews when someone writes @codex review on GitHub or automatic reviews are on, and flags only P0 and P1 issues. OpenAI meters those GitHub reviews as Code Review usage, and reviews run locally draw on the general limits.1726

41 sources

  1. Claude Code docs: OverviewOfficial documentationcode.claude.com
  2. Claude Code docs: PlatformsOfficial documentationcode.claude.com
  3. Claude pricingOfficial documentationclaude.com
  4. Claude EnterpriseOfficial documentationclaude.com
  5. Claude Code docs: CostsOfficial documentationcode.claude.com
  6. Claude Code docs: SecurityOfficial documentationcode.claude.com
  7. Claude Code docs: SandboxingOfficial documentationcode.claude.com
  8. Claude Code docs: SettingsOfficial documentationcode.claude.com
  9. Claude Code docs: Deploy managed settingsOfficial documentationcode.claude.com
  10. Claude Code docs: Server-managed settingsOfficial documentationcode.claude.com
  11. Claude Code docs: Claude Code on the webOfficial documentationcode.claude.com
  12. Claude Code docs: Data usageOfficial documentationcode.claude.com
  13. Claude Code docs: MemoryOfficial documentationcode.claude.com
  14. Claude Code docs: Model configurationOfficial documentationcode.claude.com
  15. Claude Code docs: Code ReviewOfficial documentationcode.claude.com
  16. Codex docs: OverviewOfficial documentationlearn.chatgpt.com
  17. Codex docs: PricingOfficial documentationlearn.chatgpt.com
  18. Codex docs: IDE extensionOfficial documentationlearn.chatgpt.com
  19. Codex docs: CloudOfficial documentationlearn.chatgpt.com
  20. Codex docs: Agent approvals & securityOfficial documentationlearn.chatgpt.com
  21. Codex docs: SandboxingOfficial documentationlearn.chatgpt.com
  22. Codex docs: Managed configurationOfficial documentationlearn.chatgpt.com
  23. Codex docs: Agent SecurityOfficial documentationlearn.chatgpt.com
  24. OpenAI: Business data privacy, security, and complianceOfficial documentationopenai.com
  25. Codex docs: AGENTS.mdOfficial documentationlearn.chatgpt.com
  26. Codex docs: GitHub code reviewOfficial documentationlearn.chatgpt.com
  27. Codex docs: ModelsOfficial documentationlearn.chatgpt.com
  28. Codex docs: Open sourceOfficial documentationlearn.chatgpt.com
  29. Claude Code docs: Permission modesOfficial documentationcode.claude.com
  30. Claude Code docs: Enterprise deployment overviewOfficial documentationcode.claude.com
  31. Claude Code docs: Code Review costs and usageOfficial documentationcode.claude.com
  32. Codex docs: Use ChatGPT Work and Codex with Amazon BedrockOfficial documentationlearn.chatgpt.com
  33. OpenHands Index: Codex on GPT-5.5, SWE-bench Verified (submitted May 4, 2026)Public sourcegithub.com
  34. OpenHands Index: Claude Code on Claude Opus 4.6, SWE-bench Verified (submitted Mar 27, 2026)Public sourcegithub.com
  35. Hacker News: Claude Code after trying Codex (Oct 20, 2025)Public sourcenews.ycombinator.com
  36. Hacker News: Codex reviews Claude's work (Jul 3, 2026)Public sourcenews.ycombinator.com
  37. Chandler Nguyen: Claude Max to Codex (Apr 3, 2026)Public sourcechandlernguyen.com
  38. GILD same-task test: Claude Code and Codex, Oct 25, 2026GILD Researchgildhq.com
  39. Terminal-Bench 4.0 leaderboard (Oct 23, 2026)Public sourcetbench.ai
  40. arXiv: How Do AI Coding Agents Contribute to Software Development? An Empirical Study of Agentic Pull Requests (Jul 23, 2026)Public sourcearxiv.org
  41. Claude: Code Review for Claude CodeOfficial documentationclaude.com
Read the research