Claude Code or Codex in September 2026: Pricing, Models and Fit

Run Claude Code and Codex next to each other on one refactor, same branch and same task, and no tidy winner emerges. The tools approach the problem from opposite ends and take different fights. Timing complicates things in a way pricing pages leave out: in September 2026 both vendors released new frontier models within three days. Anthropic put out Claude Fable 5.1 and Claude Mythos 5.1 on September 1, then OpenAI replied with GPT-6 Astra on September 3 and made it the default in the Codex CLI a day later. A comparison written before that week describes yesterday's models.

Two different products

Claude Code is Anthropic's coding agent for the terminal. It works inside the local environment, with reach into files, git history and the shell. The developer reviews a plan, the agent executes and a human remains in the loop at the permission level they picked.

Codex is OpenAI's agentic coding product and spans more surfaces: CLI, VS Code extension, web app, iOS app and desktop app, all linked through a single ChatGPT account. Much of its work is asynchronous and happens in cloud sandboxes. A developer assigns a task, goes off to do something else and reviews the resulting diff afterward.

That split, local and synchronous against cloud and asynchronous, is the real fork in the road. The rest of the comparison flows from it.

Pricing

Headline prices look nearly copy-pasted, at $20, $100 and $200 on the consumer tiers. The symmetry is deliberate, and it makes sticker price a poor tiebreaker.

Details are where they split. Codex has a real free tier and an $8 Go plan, neither of which Claude Code matches, since Claude Code requires a paid Anthropic plan or your own API key. Against that, OpenAI's Plus limits are tight enough that a developer on the OpenAI forum described credits vanishing within hours during peak use, until OpenAI altered the reset schedule. Each platform meters usage over a rolling five-hour window plus a weekly cap and keeps exact absolute numbers private. Any article quoting precise weekly hours is estimating.

There's another wrinkle. In April 2026 Codex replaced per-message pricing with token-based credits, so Plus and Pro allowances now count tokens used and not messages sent. OpenAI claims roughly fourfold better token efficiency. Independent tests on the model pair of that period confirmed the 4x figure for raw tokens, but that doesn't mean finished work costs a quarter as much, since the two tools may take different numbers of steps to reach a solution.

Benchmarks, with caveats

The current picture is messy. Both companies shipped new models in the same week, and independent evaluation is still catching up.

By Anthropic's published numbers, Claude Fable 5.1 reaches 55.8% on Terminal-Bench 4.0, and on SWE-bench Pro it posts 81.2. GPT-5.6 Sol manages 37.3% on that terminal test. Artificial Analysis ranks Fable 5.1 above Astra on its Coding Agent Index (70 versus 67) and on its wider Intelligence Index (66 versus 61).

Astra isn't beaten across the board. On the ARC-AGI-3 harness it scores a strong 62.7%, compared with 30.2% for Claude Opus 5. Reviewers who sent the same prompts through both models saw Astra pull ahead on computer use, long-context retrieval and math, and saw Fable 5.1 produce visibly longer, better-defended answers for analysis and writing.

Put briefly, Fable 5.1 owns the stronger, better-verified lead where agentic coding and terminal work are concerned. Astra does better at math, form-filling agent jobs and short high-volume tasks. No complete head-to-head on a single shared suite exists from either company, and Anthropic has shared no full reasoning or long-context results for Mythos 5.1, just one Terminal-Bench row. Broader Mythos comparisons found elsewhere call for caution.

Real usage paints a different picture than launch benchmarks. One much-quoted comparison covered the earlier generation, pairing Codex with GPT-5.3 against Claude Code with Opus 4.6. Claude Code burned about 1.4 times as many tokens on matched tasks, which translates to around 23% more cost per equivalent finished job. That gap hasn't been independently measured again for the Astra and Fable 5.1 pairing.

How each one feels day to day

Regardless of the model underneath, the tools feel different in daily use.

Claude Code stays close to the work. It runs in the terminal and sees the real file system, git history and shell state. MCP server support lets teams wire it into whatever internal tools they already run. Sessions can run long and stay coherent, and permission control is finer-grained, so there's more say over what it may touch unasked. Blind developer reviews cited in several 2026 comparisons favor it by roughly two to one on genuinely hard tasks: complex refactors, tricky bug hunts and anything where a wrong answer costs real rework.

Codex leans toward delegation. A task is submitted, a cloud sandbox starts and the developer moves on until a diff comes back. The async design suits routine, high-volume, low-supervision jobs like overnight batches, form filling and PRs across numerous small repos. It also combines a global AGENTS.md with per-directory project files, reading from the repository root down to the working directory so nearer files override broader ones, which is more structured project configuration than Claude Code offers today. More surfaces are covered too, including a real mobile app, a desktop app and a web interface.

Data residency for EU teams

For teams in the EU, this section may weigh more than any benchmark. Codex gets EU data residency through ChatGPT Enterprise, or through the API once residency is set explicitly. On Anthropic's own consumer and team plans Claude Code offers no EU residency whatsoever. The sole path runs through AWS Bedrock in Frankfurt or Vertex AI in the EU, with your own API key and not Anthropic directly. For a US-only team that's irrelevant. Under EU data rules it can settle the decision before anyone opens a benchmark chart.

When each one fits

Claude Code tends to suit:

• Difficult, high-stakes work. Multi-file refactors and stubborn bugs, where an incorrect answer costs more than the time saved. Blind reviews keep favoring it for these.
• Close coupling to the local environment. Direct shell, file system and git access with no cloud round trip.
• Custom tool integration. Its MCP support is more mature for teams with internal tools the agent should talk to.
• Extended coherent sessions. Finer permission control makes long autonomous runs safer to leave alone.

Codex tends to suit:

• Cheap entry. A genuine free tier and an $8 Go plan beat Claude Code's requirement to pay up front.
• Async, high-volume work. Overnight batch tasks, PR generation and routine jobs across many repos.
• Surface coverage. CLI, VS Code, web, iOS and desktop under one account, against Claude Code's terminal-first focus.
• Safer sandboxing by default. Cloud sandboxes limit the blast radius in a way local execution doesn't out of the box.
• EU residency. A real option via Enterprise, which Claude Code lacks on its own infrastructure.

Common questions

Which costs less? At the entry level, Codex wins, as a free tier plus $8 and $20 options sit below Claude Code's $20 floor. With heavy team use both land in about the same $100 to $250 per developer per month band. One often-cited test saw Claude Code cost around 23% more per equivalent task through heavier token use, but that number predates the newest models from both vendors.

Which model is smarter at the moment? Neither, outright. Fable 5.1 sits ahead on the disclosed coding and terminal benchmarks and on Artificial Analysis's combined indices, whereas Astra sits ahead on math, computer-use agent work and the ARC-AGI-3 harness. Choose by the task and not by a single leaderboard figure.

Can a team use both? Yes, and plenty already do: Codex for overnight, low-supervision batches and Claude Code for interactive daytime work on harder problems. Nothing in either architecture stops them running side by side.

Does Claude Code offer EU data residency? Not through Anthropic's own plans. It must be routed via AWS Bedrock Frankfurt or Vertex AI EU with a personal API key, whereas Codex gets EU residency through ChatGPT Enterprise or straight from the API.

Can this week's benchmarks be trusted? Read them as a snapshot. Both vendors launched frontier models three days apart in early September 2026, outside evaluation of GPT-6 Astra is still being filled in, and a number of outlets pointed out that wide API access only began days after launch.

Picking one

After two weeks of using both, the choice hinges on what kind of week it is. During a gnarly refactor, when mistakes are costly, Claude Code's tighter local loop and stronger results on hard coding tasks come out on top. With a stack of routine, low-risk jobs to clear overnight, Codex's async cloud delegation is the better fit.

Neither vendor will let the other sit still for long. If budget is the one variable under control, begin with Codex's free tier or the $8 Go plan and check whether it covers the daily grind. If the hardest problems are what really eat time, give Claude Code's $20 Pro tier a trial before stepping up to the $100-plus plans from either company.