Claude Code, Codex, Cursor and Hermes AgentAI coding agents compared on one map
Claude Code or Codex, Cursor or Hermes Agent: where the agent runs, what it sees, which model it thinks with and what it costs — five harnesses side by side and which one to pick for the task. With links to the docs.
and alternatives: OpenCode, Aider, Cline, GooseA harness of the same type with any model, via an API billed per token or locally on your own computer
AI code editorCursorAI-native IDE, models from several vendors
AI assistant in the browserChatGPT, Claudecode runs in a sandbox, no project access
The snapshot tag means the information is current as of September 2026 and may become outdated within a week.
Which harness for which task
The choice starts with the task. The two circles are model families, Claude and GPT: a product’s place on the map shows which models it works with. Claude Code and Codex are each tied to their own family, Cursor and chat sit in the overlap, and Hermes Agent fully covers Cursor and chat, partly overlaps Claude Code and Codex and reaches beyond both circles, connecting any model, including local ones.
notes · snapshot September 2026
Task
Group
One-off question without the repository
chat: ChatGPT or Claude
Change inside an open project, several files
agent in the editor
Task end to end without step-by-step involvement
Claude Code or Codex
Work in one project in a familiar editor with models from several vendors
Cursor
A custom harness on any model, billed per token
Hermes Agent
Prompt for production use via API with large inputs
data in XML tags
Common questions
Claude Code or Codex
They match in project access and actions and differ in the model (Claude and GPT) and in individual features that change from release to release. The choice is set by the model and plan the team already has.
Cursor or Claude Code
Cursor works inside an editor with an open project and lets you choose between models from different vendors. Claude Code takes a task from start to finish in the terminal, but only on Claude models.
Hermes Agent or Claude Code
Hermes Agent is an open-source harness: it connects any model, including local ones, and you pay per token. Claude Code is tied to the Claude family and a subscription. Choose Hermes for models from several vendors or your own model.
Claude Code alternatives
Of those covered here: Codex (the same design, GPT models), Cursor (an editor, models from several vendors) and Hermes Agent. OpenCode, Aider, Cline and Goose are built in a similar way.
Takeaway
In this exampleA task sets two requirements for a harness: whether it needs the project and which model family fits. Vendor agents work with one family, Cursor and chat offer a choice between Claude and GPT, and open-source harnesses take any model, including local ones.
In generalFirst pick the group by the task, then the product within the group. Claude Code and Codex differ in the model and in individual harness features that change from release to release, yet conceptually they are built the same way. So what decides is the model and plan the team already has.
Next stepDescribe the task in one line and settle three conditions: does it need the project, does it need a human at every step, and does it need another vendor’s model or a local model. Pick the group by those conditions.
What stays the same
Claude Code and Codex are competing products whose capabilities converge with every release. The gap between them is small, so this article relies on what remains constant.
Stable
model family
how a harness is built
ways of working with an agent
Changes every week
features and modes
default autonomy settings
positioning
models, plans and limits
Takeaway
In this exampleEight Claude Code releases shipped between 17 and 25 September; four Codex releases and the new GPT-6 models shipped between 22 and 26 September.
In generalA feature comparison of Claude Code and Codex goes out of date faster than anyone can make sense of it. Three things remain stable: the model family, how the harness is built and the ways of working with an agent.
Next stepWhen choosing between Claude Code and Codex, base the decision on the model and plan already available to the team.
What a harness is
A model answers with text. The harness is everything around it: skills, tools, environment, the human in the loop. Every AI product has one, from a terminal agent to a browser chat. What differs is the composition. The "Via API" column uses Hermes Agent as an example; there are many open-source harnesses, such as OpenCode, Aider, Cline and Goose, and they are built the same way.
snapshot · September 2026
Via APIHermes Agent · public API
EnterpriseClaude Code · Codex · private API
CloudClaude Code · Codex
ChatChatGPT · Claude
Human
along / at the end
along / at the end
along / at the end
every step
Environment
your computer
your computer
vendor VM
sandbox
Tools
connectors
Skills
Model
any · API at $/token or local
Claude / GPT · private API, plan
Claude / GPT · private API, plan
Claude / GPT, free / plan
differs from enterpriseyespartly
Chat and agent
How Claude Code differs from Claude in chat
The model family is the same; the harness differs. Chat answers in a conversation and does not see your project. Claude Code reads code, edits files and runs commands in the project: in the terminal, an editor, the desktop app or a browser. Docs
How Codex differs from ChatGPT
The same rule applies: the GPT models are shared, the environment differs. ChatGPT holds a conversation without project access; Codex works with code and runs commands.
Takeaway
In this exampleThe four types differ in two properties: where code runs and how model access works. Open-source and enterprise run on your computer and differ only in model access, which is where the cost difference comes from.
In generalLocal types fit work with files and secrets on your computer, the cloud fits long and parallel tasks, chat fits tasks without project access. Constraints narrow the choice: an enterprise plan is not available in every jurisdiction.
Next stepDecide whether the task needs files on your computer, whether the agent should work without you, and which model you can use. The first points to a local type, the second to the cloud, neither to chat.
What they see and do
snapshot · September 2026
Hermes Agent
$$$
Claude Code and Codex
$$
Cursor
$$
Chat: ChatGPT, Claude
$
$cheaper$$$pricier · exact prices in the tooltipEdits project filesRuns commands on your computerRuns the “edit → test → edit” loop on its ownWorks in an isolated environment
What is what
What is Codex
OpenAI’s agent for working with code: it takes a task, edits files and runs commands. It works in the terminal, in an editor extension and in the cloud, on GPT models. Project access is what sets it apart from the ChatGPT chat. Docs
What is Cursor
A code editor with a built-in agent: the agent completes tasks on its own, runs terminal commands and edits code. It works in an open project and supports models from several vendors. Docs
What is Hermes Agent
An open-source agent from Nous Research. It runs on your computer, connects to different model providers, including Claude and open models, and can work with a local model. Docs
Takeaway
In this exampleClaude Code and Codex match in project access and actions; they differ in the model and in individual harness features while staying conceptually the same.
In generalCursor stands out with its built-in editor and multi-vendor models, chat with its lack of project access. Between Claude Code and Codex, the model and plan determine the choice.
Next stepIf the task requires project files, chat is not suitable. Then choose by the model already available to the team.