Claude vs Codex: what AI builders are saying

Get every new Claude and Codex video summarised in your inbox.

Agentic Autonomy and Loop Engineering

Several creators who cover both tools directly note that Claude Code and Codex share the foundational loop primitive — both support a for/goal command that tells the agent to keep working until an exit condition is met — but creators find meaningful differences in how each realises long-running autonomy in practice. Chris, in his loop engineering breakdown, treats Claude Code and Codex as near-equivalent starting points for loop-based workflows, noting that most AI coding agents including Claude Code, Codex, and Cursor support the same goal-driven loop structure. AI Jason similarly discusses both in the context of trigger types and loop contracts without strongly favouring one, describing the go command as available in both Claude Code and Codex.

Where creators do draw a distinction is in the practical experience of extended unattended runs. Matthew Berman reports building a Minecraft clone and a functional Excel clone using Codex loops running for five to seven days without interruption, describing GPT-5.6 inside Codex as taking the most direct path to a solution of any model he has used. Riley Brown adds that Codex's computer use runs in the background without taking over your screen, in contrast to Claude Code, making it far more practical for automated QA testing and agentic loops. Creators have not described equivalent multi-day unattended Claude Code sessions in comparable detail, though the Creator Magic channel documents overnight Claude-orchestrated build queues of thirty-plus items completing successfully.

Dan Shipper, who tested Codex with GPT-5.6 over thirty days, calls it more practical than Claude for everyday use and frames it as his primary operating system for all knowledge work. Against this, the Creator Magic host notes that Claude Opus was used as the orchestrator-planner inside Tank precisely because of its reasoning quality, with cheaper models handling execution — a pattern that implicitly positions Claude as the stronger planning intelligence even when Codex handles the agentic loop itself.

Raw Model Quality and Benchmark Standing

The most direct head-to-head quality comparisons come from creators covering the GPT-5.6 launch alongside Claude. Wes Roth cites benchmark data showing GPT-5.6 Soul scored 53.6% on Agents Last Exam at a cost of $763, versus Claude Fable 5 at 40.5% costing $2,300 — a better score at roughly one-third the price. Soul also achieved an 80 on the Artificial Analysis coding agent index, beating Fable 5 by 2.8 points while using less than half the output tokens. Riley Brown and Ross Mike characterise GPT-5.6 Soul as a class below Claude Fable at the time of their review, noting that GPT-6 is expected to be the true Claude competitor — a nuance that suggests the benchmark picture may be contested or model-version-sensitive.

Matthew Berman, in his cost-cutting analysis, observes that GPT-5.6 Soul uses far fewer tokens per task than Claude Fable, making its real cost-per-task significantly lower despite similar per-token list pricing. Jack Roberts, whose video directly tests Claude flagship models against cheaper alternatives, finds that Claude Fable 5 produced slightly better results than Claude Opus 4.8 in taste-sensitive work such as cold email and website design, but the difference was subtle enough that most people would not notice — a finding that indirectly contextualises where Claude's quality premium actually manifests.

One creator at Creator Magic raises a caveat relevant to both tools: frontier models including Claude produce wildly inconsistent results day-to-day due to server load and undisclosed prompt or quantisation changes, making local models more appealing for repeatable automated workflows. This consistency concern is attributed specifically to Claude in that source, without an equivalent observation being made about Codex's reliability across sessions.

Pricing and Token Efficiency

Token efficiency emerges as one of the most consistently cited practical differences between the two tools across creators who discuss both. Jack Roberts notes that Codex, powered by GPT-4.5 at the time of his review, uses approximately four times fewer output tokens per task than Claude, making it more cost-efficient for heavy users. Matthew Berman frames the same dynamic around GPT-5.6, observing that Soul takes the most direct path to a solution and uses far fewer tokens than Claude Fable to achieve equivalent results, even where per-token list prices are comparable. Wes Roth's benchmark data reinforces this, showing Codex with GPT-5.6 reaching a higher Agents Last Exam score than Claude Fable at less than one-third the cost.

Matthew Berman's dedicated cost-reduction video adds a structural observation: third-party tools like Cursor have built-in automatic model routing that shifts simpler sub-tasks to cheaper models even when a frontier model is selected, whereas first-party tools including both Codex and Claude Code have no incentive to do this routing on the user's behalf. This positions both Codex and Claude Code similarly as tools where the user must actively manage routing themselves. Jack Roberts and David Ondrej both describe routing Claude as the orchestrator while delegating execution to cheaper models such as Opus 4.8, DeepSeek, or GLM 5.2 — a strategy creators apply to Claude specifically, though the GLM 5.2 setup guide by Greg Isenberg describes an equivalent approach where Codex is used for review in a plan-with-Opus, execute-with-GLM, review-with-Codex pipeline.

On headline pricing, creators cite Claude Fable 5 at approximately $45 per two million input tokens and 500,000 output tokens, while GPT-5.6 Soul is noted at $5 per million input tokens and $30 per million output tokens at launch — though creators caution that real cost-per-task depends heavily on token efficiency rather than list price alone.

Computer Use and Background Automation

Codex's computer use implementation is highlighted by multiple creators as a distinctive capability, with Riley Brown drawing an explicit contrast to Claude Code. Riley notes that Codex's computer use runs in the background without taking over your screen, unlike Claude Code, which he describes as far more practical for automated QA testing and agentic loops. Matthew Berman provides concrete examples of this in action, demonstrating Codex browser control used to automate DNS migrations across Vercel, DigitalOcean, and GoDaddy, and to auto-scale a Supabase instance during a large data import — all without interrupting the user's primary workflow.

Codex also introduced a record-and-replay feature — described by Riley Brown as the future of computer use for knowledge work — that lets users record their screen for up to thirty minutes and convert that session into a reusable slash-command skill. Claude Code does not appear to have an equivalent screen-recording-to-skill feature in the corpus; instead, creators describe Claude's computer use as more screen-blocking and therefore less suited to background automation. Jack Roberts notes the Codex Chronicle feature, which captures screen context via screenshots so the model has live awareness of what the user is working on, and Dan Shipper corroborates this, describing Chronicle as giving Codex continuous context that improves response quality over time.

For Claude, creators describe workarounds rather than native background computer use. The Creator Magic host documents running Claude Code inside Tank via TMUX terminals and scheduling tasks through build queues, effectively achieving background operation through orchestration infrastructure rather than a built-in capability. A community tip shared in that channel advises specifically against using Claude 5 for token-heavy computer use or web scraping tasks, recommending cheaper models instead — a caveat creators do not apply to Codex in the same terms.

Skills, Integrations, and Platform Ecosystem

Both Claude and Codex have converged on a skills-based architecture — reusable markdown instruction files that give agents consistent, repeatable behaviour — and creators note that the two ecosystems are more interoperable than they might appear. Jack Roberts demonstrates a prompt-based workflow for migrating skills created in Claude directly into Codex, and the Brock Mesarich channel shows Claude Cowork skills stored as .md files being downloaded and uploaded into ChatGPT Work, the consumer-facing counterpart to Codex, with no reformatting required. Matt Wolfe observes that skills and plugins are largely universal across Claude Code, Codex, Cursor, and several other agents.

Where the ecosystems diverge is in native platform integrations. Codex has a native plugins and skills marketplace for one-click installation, whereas Claude's integrations are primarily delivered through MCP connectors — Zapier MCP, Firecrawl, Higgsfield, Clay — that require manual setup but extend Claude to over nine thousand applications. The Creator Magic host identifies a meaningful limitation: skills saved in Claude-specific directories are not accessible to other agents such as Grok or Codex without being resaved using an open protocol or shared directory structure, suggesting Claude's skill ecosystem is more siloed by default.

On the IDE and workspace integration side, Claude Code's Anthropic-native ecosystem includes Claude Tag, a Slack integration that gives teams a shared, persistent Claude instance per channel with ambient monitoring and external tool access. Codex's equivalent is ChatGPT Work, described by Brock Mesarich as offering the same core feature set — skills, sites, and scheduled tasks — with the addition of native image generation, which Claude Cowork lacks without an external connector. Creators in the corpus treat the two as roughly comparable non-technical agentic interfaces, though neither is described as clearly superior across all integration categories.