Two coding agents compared for coordinated repository changes, with the decision narrowed to workflow surface, model path, execution controls, usage billing, and private-code data policy. This guidance is based on company pages linked from the profiles below.
Codex:Oldest source collected Claude Code:Oldest source collected
Which tool fits your work?
Start with Codex if
You want one coding agent across local repository work, cloud delegation, terminal, editor, and code review workflows; the CLI can also use local OSS providers through Ollama or LM Studio.
Look elsewhere if: You need managed cloud agents to run on infrastructure you control, need a fixed task allowance, or want cloud tasks while authenticating only with an API key.
Your development workflow centers on a terminal, repository tools, and coordinated file changes.
Look elsewhere if: You need a dedicated AI-first editor rather than an agent that also integrates with IDEs, or cannot send code context to a supported hosted model service.
Free and Go include limited trial access. Plus is $20 per month. Pro starts at $100 per month, with current tiers offering 5× or 20× Plus usage. Business is $20 per user monthly when billed annually or $25 month to month, with at least two users. Enterprise and Edu use sales-led pricing. OpenAI's newer rate card says Plus, Pro, and Business accounts plus most Enterprise-family accounts use token-based rates, while its Codex pricing page still says only Business and new Enterprise accounts have migrated; check the Usage panel for the meter that applies to your account. Some Plus and Pro users can add credits, while others may need to upgrade or wait for a reset. All users can run extra local tasks with an API key at standard API rates. Credit use depends on the model, input and output token mix, speed, and any additional agents. OpenAI publishes a cross-plan planning estimate of roughly $100–$200 per developer per month, with large variance; it does not define an included allowance or predict an individual bill.
Claude Code access is included with Pro, Max, every Team seat, and eligible Enterprise seats. Pro, Max, Team, and seat-based Enterprise have included usage limits. Current usage-based Enterprise costs $20 per user monthly, billed annually, with a 20-seat self-serve minimum or 50-seat sales-assisted minimum. Its seat fee includes no usage; every token across Claude, Claude Code, and Cowork is billed separately at standard API rates. Pro is $20 monthly or $200 yearly. Max 5x is listed at $100 per month and Max 20x at $200 per month. Anthropic's plan guide and pricing FAQ say both Max tiers are monthly only, but its pricing comparison table lists monthly and annual billing; confirm the available billing interval before subscribing. Team Standard is $20 per seat monthly when billed annually or $25 month to month, with a two-seat minimum; Premium is $100 or $125 and adds 5× more usage. New and self-serve Enterprise seats include Claude Code; older Enterprise plans may require a Chat + Claude Code or Premium seat. Individual Pro, Max 5x, and Max 20x subscribers can prepay usage credits to continue Claude Code terminal use after included limits at standard API rates, with a configurable monthly spending cap. Pay-as-you-go API usage through a Console account is a separate path.
Ways to use it
Web, macOS, Windows, Linux, iOS, Terminal, VS Code
Available through Codex in ChatGPT, the desktop app, CLI, IDE extension, web, and iOS, subject to plan, platform, and workspace settings. The Codex IDE extension also supports Cursor and Windsurf; OpenAI documents separate integrations for Xcode and JetBrains IDEs. Local CLI, IDE, and SDK use can instead authenticate with an API key; API-key access does not include cloud features such as GitHub code review or Slack.
Terminal, IDE, web, mobile, and GitHub access. Anthropic now provides Slack access as Claude Tag, a Team and Enterprise beta that a Primary Owner or Owner must set up for selected channels.
Where it runs
Local workflows run commands on the user's device under configurable sandbox, approval, and network policies; the CLI can also use OSS models served by Ollama or LM Studio. Cloud tasks run in isolated OpenAI-managed environments; setup scripts have internet access, while agent internet access is off by default unless configured.
Terminal and IDE sessions execute on the developer's machine and send model requests to hosted APIs. Cloud sessions execute on Anthropic-managed infrastructure by default. Team and Enterprise organizations can instead enable public-beta self-hosted environments, which move the session process and repository checkout to infrastructure they operate while Anthropic's control plane and model inference remain at api.anthropic.com. Cloud sessions are available in research preview for Pro, Max, Team, and Enterprise Premium or Chat + Claude Code seats; self-hosting is off by default.
Start with Codex when you want one agent across its desktop app, CLI, VS Code extension, web, iOS, and OpenAI-managed cloud tasks. Its local CLI, IDE, and SDK can use an API key, but API-key access excludes cloud features such as GitHub code review. Start with Claude Code when you want a terminal-centered agent with native VS Code and JetBrains extensions plus desktop, web, and mobile clients. The linked product pages establish availability, not feature parity across clients, so verify the exact client and workspace policy your team would use.
Does the model request need to stay on the local machine?
Codex CLI documents a local OSS path through Ollama or LM Studio. That is separate from Codex cloud tasks, which run in OpenAI-managed environments. Claude Code's terminal client runs locally but sends prompts and model outputs over the network for inference. Anthropic documents hosted model paths through its API and cloud-provider integrations—for example Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, and Claude Platform on AWS. The checked company pages do not establish an offline or local-model path. A local agent is not necessarily a local model, so verify the exact client, model, and provider before opening a repository with a hard network-isolation requirement.
Which approval and execution boundary fits the repository?
For local work, Codex separates a platform-enforced sandbox from approval policy: routine commands can run inside configured filesystem and network boundaries, while crossing them can require approval. Local Claude Code uses allow, ask, and deny rules; its Manual mode asks before file edits and most shell commands, and its optional Bash sandbox enforces filesystem and network boundaries. Cloud behavior differs. Codex cloud tasks run in isolated environments where setup scripts have internet access and agent internet access is off by default. Claude cloud sessions pre-approve file edits and do not offer Manual mode whether they use the default Anthropic-hosted environment or an eligible Team or Enterprise organization's optional self-hosted environment; other permission prompts appear in claude.ai. Anthropic-hosted sessions use isolated virtual machines. Self-hosting moves the session process and repository checkout to organization-operated infrastructure, while model inference and session control still use api.anthropic.com. Compare the exact local or cloud mode rather than treating either product name as one permission model.
Which produces a mergeable change with less supervision on your repository?
Do not answer this from one benchmark or demo. First decide whether you are comparing the products as sold—using each product's normal default model and mode—or the agent harnesses with the same model version where both offer it. From the same clean commit, give each a bounded task, identical repository instructions and acceptance checks, and the closest comparable permissions, time, and spend ceilings each permits; record every mismatch. Run at least three fresh attempts per setup. That can reveal instability, not establish a universal advantage. For every run, record the date, client version, model, mode, elapsed time, reported usage, and every human intervention; mark missing or non-comparable figures unknown. Then inspect the diff and judge requirements met, tests passed, regressions, and unnecessary edits. The result applies only to that repository, task, setup, and date.
Use the current price row above for subscription costs. Both products list a $20 monthly individual plan that includes coding-agent access, but neither promises a fixed repository-task allowance. OpenAI's newer rate card says Plus, Pro, and Business accounts plus most Enterprise-family accounts use token-based Codex rates, while its Codex pricing page still says only Business and new Enterprise accounts have migrated; check the account's Usage panel. Codex also shares one agentic usage and credit pool with other eligible ChatGPT agent features. Claude Code shares Claude's included usage pool. Some Codex Plus and Pro users can add credits, while Claude Pro subscribers can prepay usage credits for terminal use; API keys are separate pay-as-you-go paths. For a hard $20 ceiling, leave Codex credit auto-reload off, do not add credits or API billing, leave Claude extra-usage credits off, and do not attach a Console API key. Then test a representative week and record the exact client, model, task size, and reported usage rather than converting either plan into an unsupported task count.
How may prompts and private code context be used for model training?
For Plus and Pro, OpenAI may use Codex content to improve models. ChatGPT conversations and Codex tasks can be opted out through general data controls, while full-environment Codex training has a separate control in Codex Settings that the general controls do not change. Business, Enterprise, Edu, and API inputs and outputs are not used for training by default; eligible API organizations can opt in. Anthropic says Claude Code data from Free, Pro, and Max accounts may be used for training when the privacy setting is on, while Team, Enterprise, and API content is not used for training by default. Set the relevant account or workspace control before opening private code, and review retention separately rather than treating a training opt-out as a deletion policy.
