Claude Code vs Codex (2026): benchmarks, cost per task, and a fair trial

Which should you choose?

Neither wins outright in 2026, and the honest answer has three parts. On public agentic-coding leaderboards the two harnesses trade places by a few points, and which model your plan entitles you to matters more than which CLI you type. On measured token efficiency Codex is currently the cheaper route to a comparable result. On the one independent head-to-head that has been published, Claude Code holds a large lead. Every one of those readings was taken on a date, against specific model versions, and all three move.

So the practical advice has not changed: start with the agent you already pay for, run the same real task in your own repository, and count your own review time. The numbers below tell you what to expect and where to look. They do not decide anything for your codebase — the scorecard further down does.

What the leaderboards actually rank

Terminal-Bench scores an agent-model system — the harness and the model together — on 89 technical tasks completed through a command-line interface, spanning software engineering, machine learning, security, data processing, scientific computing, and system administration. That framing is the useful part: the same model scores differently under different CLIs, and the same CLI scores differently with a different model. A headline that names only a model is describing half the system.

Three things to take from that table. The top entry is a Codex CLI run and the second is a Claude Code run, about four points apart — a smaller gap than the distance between two models inside the same family. The same model under a different harness moves by more than three points, which is the whole argument for testing the CLI rather than the model. And Claude Opus 5, released on 24 July 2026, does not appear on the board at all: Anthropic reported Frontier-Bench v0.1, CursorBench 3.2, ARC-AGI 3, OSWorld 2.0, GDPval-AA v2 and Zapier AutomationBench at launch instead. A missing entry is a missing measurement, not a low score.

Which model you can run at all is set by your plan. In Claude Code the default model resolves to Opus 5 on Max, Team Premium, Enterprise and the Anthropic API, and to Sonnet 5 on Pro and Team Standard. The Sonnet 5 row above sits about nine points below the Fable 5 row under the identical harness — the largest single effect in this comparison that you control with money rather than with prompting. Max vs Pro covers that decision on its own.

Token efficiency and cost per task

A leaderboard says who finished. It does not say what finishing cost. Artificial Analysis publishes both, running the same Intelligence Index against each model and reporting the output tokens it consumed and the bill it generated.

Read across the row rather than down a column. Opus 5 finishes four index points ahead while emitting roughly 1.6 times the output tokens and costing roughly 2.1 times as much to complete the same evaluation. That ratio is what people mean anecdotally by "Codex is cheaper for the same job", and it is why a subscription that feels generous on one side can feel tight on the other for identical work.

Both vendors are chasing that ratio directly, which is another reason the numbers expire. Anthropic says Opus 5 reaches comparable results on its trading benchmarks using roughly a seventh of the reasoning tokens of Opus 4.8 and under half the latency, with a fast mode at around 2.5 times the default speed. Artificial Analysis notes that GPT-5.6 Sol uses fewer output tokens than most models of comparable intelligence, at about 15,000 tokens per Intelligence Index task against GPT-5.5’s 16,000.

The one independent head-to-head worth citing

Vendor comparisons of these two are everywhere; controlled ones are rare. The clearest published example is SocSci-Repro-Bench, a June 2026 academic benchmark of 221 tasks drawn from 54 papers across political science, sociology, psychology and communication, which ran both agents against the same replication materials and scored whether the published result came back.

Read it for the shape, not the score. A 31-point task-level gap is far too large to be noise, and it is a genuine like-for-like run. It is also already stale: it used Opus 4.6 and GPT-5.3, two model generations behind what both CLIs ship as defaults today, and re-running statistical analyses in R and Stata is one workload, not yours.

The vendor-run comparisons point the same way on quality and the opposite way on cost. Codegen’s June 2026 write-up reports a blind review by more than 500 developers in which Claude Code’s output was rated cleaner and more idiomatic 67% of the time, against 25% for Codex and 8% tied, alongside SWE-bench Verified within a tenth of a point for both (88.6% and 88.7%) and a wider Claude lead on SWE-bench Pro (69.2% against 58.6%). It is a vendor blog rather than a controlled study: treat the direction as consistent with the academic result and the exact percentages as unverified.

What each side costs in 2026

Both sides meter on a rolling five-hour window with a weekly limit on top. Anthropic resets the weekly window at a fixed time assigned to your account and counts claude.ai, Claude Code and Cowork against the same allowance. OpenAI refreshes Codex limits weekly and shares one allowance between local and cloud chats. Neither publishes a task count, so any "N prompts per window" figure you find elsewhere is an estimate, not a term.

If you are weighing API billing against a subscription, Anthropic’s own guidance for enterprise deployments is the useful anchor: around $13 per developer per active day, $150–250 per developer per month, staying under $30 per active day for 90% of users. That band straddles Max 20x, which is why a subscription is usually the cheaper path for one person coding daily. Agent teams change the arithmetic — they use approximately 7x the tokens of a standard session when teammates run in plan mode.

Capability differences that survive a model change

Model scores move every few weeks. The shape of each CLI moves much more slowly, and it is what you will still be living with after the next release.

Both give you a permission boundary; they express it differently. Claude Code has permission modes, a plan mode you cycle to with Shift+Tab, and /rewind checkpoints that restore both the conversation and the files. Codex reads sandbox_mode and approval_policy from ~/.codex/config.toml, ships named permission profiles — :read-only, :workspace, :danger-full-access — and exposes /permissions inside a session. The Codex CLI workflow guide has that configuration in full, and the sandbox and permission guide covers the Claude side.

The practical differences to check

Configuration is part of the comparison. A run with unrestricted access, network tools and a large context is a different experiment from a run with constrained permissions and fewer tools. Keep access comparable and record unavoidable differences. Use the official documentation linked below for the current settings available in your installed version.

Choose three tasks that resemble your actual work

• A bug fix with an observable failing case: the result must remove the failure and keep related behavior intact.
• A small feature with a clear boundary: record the expected interface, accepted inputs, and one invalid input.
• A review task against a frozen diff: assess whether findings are specific, reproducible, and relevant.

Avoid a first evaluation that includes a deployment, a major migration, or a rewrite without acceptance criteria. A reversible local change gives you a result you can inspect. If you cannot describe success without referring to the agent’s explanation, make the task smaller before comparing tools.

Prepare a fair comparison

Start both trials from the same committed state, in separate worktrees. Use fresh branch names and directories for each task. Keep one written brief, the same test command, and the same allowed external tools. Record the CLI version, selected model, effort level, date and settings. Do not guess the model from a product name — on a Claude subscription the default differs by plan, and the Codex default differs by tier.

git status --short
git worktree add -b eval/claude ../eval-claude HEAD
git worktree add -b eval/codex ../eval-codex HEAD
claude --version
codex --version

Give the same task to each agent independently. Do not paste one agent’s solution into the other trial. If one run needs clarification, record the intervention and give equivalent information to the other. Follow the parallel worktree setup for launch, environment isolation, review and cleanup.

Fix [reproducible issue] in this checkout.
Expected behavior: [observable acceptance criteria].
Scope: [files or component]. Preserve unrelated behavior.
Run [verification command] and report what it actually proves.
If you cannot verify a requirement, identify the missing evidence.
Do not commit or deploy. Return a reviewable diff and concise explanation.

Use a scorecard that includes your review time

A faster response can still produce a slower accepted change if you spend longer fixing it — which is exactly what the token-efficiency table above cannot tell you. Conversely, one failed attempt does not establish that a tool is unsuitable. Repeat with the three task types, and keep the raw observations before turning them into a preference.

Treat this as a small operational trial, not a scientific leaderboard. Task difficulty, caching, permission prompts, model changes and your familiarity with each CLI all affect the result. A useful decision can be conditional: use one for a familiar service and the other for independent review.

When using both is worth the overhead

A mixed workflow makes sense when tasks are independent or roles are sequential. Let one agent implement a parser fix, then ask the other to review the committed diff against the original failing case. Ask the reviewer to reproduce its findings and point at the affected lines. Review the findings yourself before changing working code.

Running both on the same files at the same time adds coordination without producing an independent result. If both are already available, use worktrees. If buying a second subscription is the decision, first record how often your existing tool actually blocks useful work and what a fallback would recover — the AI subscription audit is that exercise written down.

CmdBrief can keep both CLIs, their attention states and the diff in one Mac workspace, alongside Grok CLI, Kimi CLI, OpenCode and Cursor CLI. It does not change the models’ answers or grant extra provider capacity. Compare the Mac GUI options if session navigation is the problem, and usage limits if interruptions are.

The closing advice is the same as the opening advice, now with the numbers behind it. Leaderboards separate these two by a few points, token accounting separates them by a factor of two on cost, and the one controlled study separates them by thirty. None of those is your repository. Give both the same commit, the same brief and the same verification command, count the review time, and re-run it when the next model lands — because every figure on this page has a date attached to it.

• Re-check the default model on your plan first: it changed on both sides during 2026 and it moves the result more than the CLI does.
• Re-check the benchmark version, not just the score. Terminal-Bench 2.1 repaired 28 of the 89 tasks inherited from 2.0, so a 2.0 number and a 2.1 number are not the same measurement.
• Re-check the price per million tokens before assuming a cost ratio still holds; both vendors shipped cheaper tiers this year.
• Re-run your own three tasks after any model change you actually notice, and keep the previous run’s scorecard so you are comparing to something.
• Re-check what your existing tool blocked before buying the second subscription, rather than after.

Primary sources

Provider terms and product behavior can change. Confirm plan, pricing, model, and permission details in the current official documentation.

• 01Terminal-Bench 2.1 leaderboardsnorkel.ai
• 02Anthropic: introducing Claude Opus 5anthropic.com
• 03Artificial Analysis: Claude Opus 5artificialanalysis.ai
• 04Artificial Analysis: GPT-5.6 Solartificialanalysis.ai
• 05Artificial Analysis: GPT-5.6 has landedartificialanalysis.ai
• 06AI Coding Agents Can Reproduce Social Science Findings (arXiv 2606.11447)arxiv.org
• 07Codegen: Claude Code vs Codex (vendor comparison)codegen.com
• 08Claude pricingclaude.com
• 09ChatGPT and Codex pricinglearn.chatgpt.com
• 10OpenAI API pricingdevelopers.openai.com
• 11Claude Code: model configurationcode.claude.com
• 12Claude Code: manage costs effectivelycode.claude.com
• 13Claude Code CLI referencecode.claude.com
• 14Codex CLI documentationlearn.chatgpt.com
• 15Git worktree documentationgit-scm.com