On April 16, 2026, OpenAI shipped Codex Desktop v26.415 under the title "Codex for (almost) everything," adding background computer use, multi-agent execution, image generation, persistent memory, 90-plus plugins, and self-scheduling automations to its flagship coding tool — and immediately sparked debate about whether the update makes Anthropic's Claude Code obsolete.
A feature-by-feature review of both products' current documentation suggests the answer is no. But the comparison is close enough to matter. Our read: the launch video framing did more work than the features themselves, but the underlying convergence is real and worth taking seriously.
The "AI Employee" Framing
The post-launch commentary has been shaped largely by a widely shared video declaring "Sam Altman just made Claude look like a toy." The video positions both Codex and Claude Code not as autocomplete tools but as AI employees — agents capable of owning tasks across days, across tools, and across machines.
That framing raises the bar considerably. The question is no longer whether a tool can suggest a function; it is whether it can produce something a working developer would accept from a junior engineer. Evaluated against that standard, both tools have converged on the same structural shape: desktop applications that control the user's machine, run parallel agents, remember preferences across sessions, connect to third-party services, and schedule their own future work. The meaningful differences are now in implementation trade-offs rather than feature presence.
What the April 16 Release Actually Added
OpenAI's official announcement lists six headline capabilities delivered with the Codex Desktop v26.415 release.
Background Computer Use. Codex can drive macOS GUI applications using its own cursor, running in isolated processes that operate asynchronously from the main application. The practical effect is that Computer Use can run in the background without disrupting the user's foreground keyboard and mouse. The officialdevelopers.openai.com/codex/app/computer-use documentation specifies the agent can operate apps "across anything you have running." The feature launched on macOS and on Windows via a native sandbox using PowerShell, without requiring WSL. It was not yet available in the EU or UK at the time of the release.
Subagent and multi-agent workflows. Codex can spawn specialized agents in parallel using isolated git worktrees, then merge results. The release included a new spawn_agents_on_csv capability, agent nicknames, and a revised picker interface designed to make managing concurrent agents less cumbersome.
Image generation via gpt-image-1.5. Inline image generation for mockups, frontend designs, and game assets ships as part of the ChatGPT subscription with no API key required.
Persistent memory, in preview. Codex now retains preferences, recurring workflows, technology stack conventions, project guidelines, and known pitfalls across sessions. The data is stored locally under ~/.codex/memories/.
A 90-plus plugin marketplace. Integrations covering Slack, Notion, Google Workspace, GitHub, GitLab, Atlassian Rovo, CircleCI, Render, and others are available alongside skills and MCP servers, all installable from a centralized marketplace.
Thread automations. Codex can schedule future work, wake itself up to continue long-running tasks, and resume the same thread across days or weeks.
Claude Code's Parallel Feature Set
Anthropic's release pace through Q1 2026 covered the same ground on most dimensions. According to public Claude Code documentation and the Anthropic blog, the following shipped during the same period.
Computer Use for Claude Code arrived on April 4, 2026 as a research preview for Pro and Max subscribers. Per Anthropic's official announcement on Dispatch and Computer Use, Claude can "point, click, and navigate what's on your screen to perform the task itself. It can open files, use the browser, and run dev tools automatically — with no setup required." The feature runs on macOS and Windows.
Dispatch allows users to assign tasks to Claude from a phone while Claude Code uses the desktop in their absence. Subagents are a first-class concept with custom system prompts, scoped tool access, and parallel execution; pressing Ctrl+B backgrounds a subagent without interrupting the current session.
Routines are durable cloud-scheduled agents running on Anthropic infrastructure, with no requirement for the user's machine to remain powered on. Desktop scheduled tasks provide on-machine cron-style automation that survives restarts. The /loop bundled skill adds session-scoped scheduling with cron expressions, dynamic interval selection, and a seven-day expiry.
For persistent context, Claude Code offers three distinct layers: CLAUDE.md project files, which are version-controlled and team-shareable; Auto Memory, which captures preferences and patterns at the user level; and Skills, which store reusable procedures. MCP connectors numbered 50-plus as of February 2026, with interactive surfaces inside the chat interface for Slack, Figma, Canva, and Asana.
Computer Control: Background vs. Safety-First
Codex's Computer Use is designed for background operation. Agents run in isolated processes and communicate asynchronously with the main app, meaning Codex can drive a separate window while the user keeps working in the foreground without interference. That decoupled architecture is the primary technical differentiator from Anthropic's approach.
Claude Code's Computer Use takes a more conservative posture by design. Anthropic's documentation frames it as a fallback mechanism: Claude prioritizes MCP connectors and command-line tools first, falling back to GUI control only when no direct integration exists. The implementation includes in-model scans for prompt-injection attempts embedded in page content encountered during browsing or file reading.
The trade-off is real. Codex's architecture permits more aggressive continuous background automation; Claude Code's architecture reduces the attack surface for malicious content but constrains how aggressively it will take over the foreground. Worth noting: for solo developers running supervised sessions this gap is largely theoretical — but for teams deploying agents to run overnight jobs unattended, the prompt-injection surface becomes a genuine architecture decision, not a footnote.
Parallel Agents: Batch Management vs. Developer Control
Both platforms support parallel agent execution with different ergonomic priorities. Codex's spawn_agents_on_csv capability and agent nickname system are oriented toward batch task management, giving users human-readable handles to monitor multiple agents simultaneously. Git worktree isolation ensures parallel agents do not interfere with each other's file state.
Claude Code's subagent model exposes custom system prompts and scoped tool access, giving developers direct control over what each agent can see and do. The Ctrl+B shortcut for backgrounding agents is designed to minimize friction in workflows where the developer wants to keep working in the primary session while delegating a parallel task without losing the current conversational thread.
Image Generation: Codex's Uncontested Win
This is the clearest asymmetry in the comparison. Codex ships inline image generation via gpt-image-1.5, integrated directly into the subscription with no additional setup. Users can generate mockups, interface designs, and graphic assets without leaving the coding environment.
Claude Code does not offer integrated image generation. Anthropic's product is oriented toward code, text, and tool-use tasks. Users who need image synthesis alongside coding assistance must switch tools or route through an MCP connector to an external image service.
Persistent Memory: Flat File vs. Layered System
Both platforms offer cross-session memory with meaningfully different architectures. Codex stores memories in a flat file at ~/.codex/memories/, a single-layer approach that is easy to inspect and manually edit.
Claude Code uses three layers with different scopes. CLAUDE.md files provide project-scoped, version-controlled context that teams can share through a repository. Auto Memory captures preferences and patterns at the user level a
Written by Hiram Clark, Editor — vybecoding.ai
Published on May 1, 2026