JetBrains AI vs OpenAI Codex (2026)

Upgraded from P: invocation now spans IDE chat, tool window, standalone CLI in any terminal, CI/CD via GitHub Actions and GitLab, ACP clients, and remote or async runs, which is breadth rather than a single entry point.

Upgraded from P: six invocation surfaces plus scheduled tasks, GitHub Action, non-interactive exec mode, and third party triggers from Slack, Linear, GitHub and GitLab. The June basis predated most of this.

Upgraded from P: grounding runs on the IDE's semantic analysis engine rather than text embeddings, giving import graphs, class hierarchies and call chains, plus persistent AGENTS.md instructions and MCP connected external sources.

AGENTS.md is durable project scoped instruction context rather than agent memory; sessions are managed and resumable but no cross session memory store is documented.

Upgraded from P: a named Memories customization surface plus Computer History, alongside persistent projects, resumable sessions, compaction and AGENTS.md. The June basis predated the Memories feature.

Held at P: .aiignore, AGENTS.md and per command approval are real user level controls, but no organisation level identity or governance surface (SSO, RBAC, audit logging, admin policy) was retrieved for the AI product, and no security attestation specific to it. The axis conjunction is not met on the evidence found.

Progress reporting, reviewable diffs and session status are documented, but no persistent execution trace or audit log across sessions; the debugger session state is inspectable at runtime rather than retained as a record.

Upgraded from P. OpenTelemetry export with a documented event catalogue covering tool decisions, tool results and approval outcomes, plus a separate Compliance API and audit events surface and an Analytics API, is a retained per action record rather than usage reporting. Same standard applied to cursor and github-copilot, both of which held at F on documented audit surfaces. Recorded honestly: telemetry is opt in and off by default, and prompt content is redacted unless explicitly enabled.

Local model serving via Ollama, LiteLLM or LMStudio keeps inference on the developer's own machine, which is real deployment control at the individual level; held at P because no organisation level self host, VPC or residency option was retrieved.

Upgraded from P: local execution on the developer's own machine under an OS enforced sandbox, dev container support, Amazon Bedrock as a model provider path, and Private Link, IP allowlisting and mutual TLS for enterprise network control. Recorded honestly: cloud tasks run in OpenAI managed containers with no customer VPC or self hosted cloud option documented.

Modes and slash commands are behaviour presets rather than a library of prebuilt agents or installable packs; the MCP registry is a server catalogue, credited to Ext.

Upgraded from P: skills and plugins are both first class installable surfaces with build guides, enterprise plugin management and skill controls, and Codex accepts Claude Code plugin layouts. That is a distribution surface, which is the axis.

Corrected from N. The prior basis read OpenAI-only model, which describes the provider rather than the axis, the same error I corrected on anthropic-claude-code and google-jules. The model is selectable per session in config.toml, a Models page documents the choices, workspace model availability is an admin control, and Amazon Bedrock is documented as a deployment and model provider path. Held at P not F because all models are OpenAI's with no third party or BYOK option, narrower than activepieces or cline.

Debug mode is the strongest evidence on this axis in the lane so far: the agent inspects live runtime state through the real debugger rather than reasoning from logs, which is genuine verification of its own work.

Downgraded from F, ninth application of the lane wide axis rule. Code review, the Security plugin and CI scanning all act on the customer's code. The closest thing to agent evaluation is Auto-review, which judges individual agent actions against a risk policy at runtime, but that is a runtime guardrail already credited on HITL and not a harness for testing agent behaviour. Note the platform's Evals product is an API surface for the customer's own applications, not a harness for evaluating Codex itself, so it does not count here; the same fact should not do work on two records.

Corrected from N. Browser and Computer use are listed as documented Capabilities in Codex's own navigation with dedicated pages, alongside a browser extension and Appshots, and the security page names browser and Computer Use activity as a distinct traffic surface requiring its own feature controls. The June basis missed this entirely.