Codex vs Claude Code, measured side by side (plus Antigravity CLI)

As of 2026-11-11 13:38 UTC · the same fixed task every hour for every agent, from one EU server, on the plans we have (Claude Max subscription, ChatGPT Free plan, Google AI subscription) · only figures measured the same way for all

As of 2026-11-11 13:38 UTC, median time to answer the same fixed task, end to end, over the last 24 hours — Claude Code with Claude Opus 5.5: 15.2 s (24 runs); Claude Code with Claude Sonnet 5.5: 10.4 s (2 runs, too few to compare); Codex CLI with GPT-6 Luna: 22.2 s (3 runs, too few to compare); Antigravity CLI with Gemini 3.8 Flash: 100.6 s (24 runs); Antigravity CLI with GPT-OSS 120B: 16.9 s (2 runs, too few to compare). Claude Code vs Codex CLI over the last 7 days: too few runs to call it yet (167 successful runs of Claude Code with Claude Opus 5.5 and 3 of Codex CLI with GPT-6 Luna in 7 days). Failed calls in the last 7 days — Claude Code with Claude Opus 5.5: 0 of 167; Claude Code with Claude Sonnet 5.5: 0 of 2; Codex CLI with GPT-6 Luna: 0 of 3; Antigravity CLI with Gemini 3.8 Flash: 1 of 168; Antigravity CLI with GPT-OSS 120B: 0 of 2. On the 20 hard questions both answered in the last 7 days, Claude Code with Claude Opus 5.5 got 62% right and Codex CLI with GPT-6 Luna 45% — no significant difference (+17.1 points ± 11.9, z = 1.44).

Each agent runs on one account on one plan — Claude Code on the Claude Max subscription, Codex CLI on the ChatGPT Free plan and Antigravity CLI on the Google AI subscription — so plan limits, routing and default models are part of what is measured; another plan or model can be faster or slower. Per-agent pages: Is Claude Code down or slower today? · Is Codex down or slower today? · Is Antigravity CLI down or slower today? · all coding agents.

Which one answers faster?

Median time from launching the CLI to the complete answer for the same fixed task, by our own clock.

"Too few": not enough successful runs yet to compare with other agents (the rule is under the pair table). "Failed": calls that errored, timed out, were rate-limited or could not count as a clean measurement.

Faster or not, last 7 days

Rule: only end-to-end time on the same fixed task (from launching the CLI to the complete answer, median of successful calls; comparable once it has >= 12 calls in 24 h or >= 48 in 7 days; two agents differ when their 7-day medians are >= 15% apart and a Mann-Whitney test gives p < 0.01), the share of failed calls, and hard-panel accuracy of each pair on the questions both answered in the last 7 days (paired per-question difference, item-clustered SE; significant when |delta| >= 5 points and |z| >= 2.58 with >= 20 common questions). Tokens/s, thinking and time to first token are not compared across agents: Claude Code reports API time, Codex and Antigravity CLI turn time, and the tokenizers differ.

Which one is more accurate on the same questions?

Each pair is compared only on the hard questions both answered in the last 7 days, question by question. The 78-question panel was selected at the edge of Claude Opus 5.5's ability (questions it solves only sometimes; 62% in the calibration run), so it is not a neutral ranking of models: read it as "on these questions, in the same week". The questions are not published.

What we do not compare, and why

Tokens per second, thinking tokens and time to first token are not compared between agents: Claude Code reports API time while Codex and Antigravity CLI report turn time, Antigravity CLI does not stream thinking, and the tokenizers differ, so the same number means different things. Each agent's own page judges them against that agent's own last 7 days. See Is Claude Code down or slower today? · Is Codex down or slower today? · Is Antigravity CLI down or slower today?.

Questions

Is Codex faster than Claude Code?

On the same fixed task over the last 7 days: too few runs to call it yet (167 successful runs of Claude Code with Claude Opus 5.5 and 3 of Codex CLI with GPT-6 Luna in 7 days). Over the last 24 hours Claude Code with Claude Opus 5.5 took a median 15.2 s (24 runs) and Codex CLI with GPT-6 Luna 22.2 s (3 runs, too few to compare). Claude Code runs on the Claude Max subscription, Codex on the ChatGPT Free plan.

Is Claude Code or Codex more accurate?

On the 20 hard questions both answered in the last 7 days, Claude Code with Claude Opus 5.5 got 62% right and Codex CLI with GPT-6 Luna 45% — no significant difference (+17.1 points ± 11.9, z = 1.44). The 78-question panel was selected at the edge of Claude Opus 5.5's ability (questions it solves only sometimes), so it is not a neutral ranking of models.

Which coding agent fails more often?

Failed calls on the same hourly task in the last 7 days — Claude Code with Claude Opus 5.5: 0 of 167; Claude Code with Claude Sonnet 5.5: 0 of 2; Codex CLI with GPT-6 Luna: 0 of 3; Antigravity CLI with Gemini 3.8 Flash: 1 of 168; Antigravity CLI with GPT-OSS 120B: 0 of 2. A call counts as failed when the CLI errors out, times out or is rate-limited, or cannot count as a clean measurement.

Why not compare tokens per second or thinking tokens?

Tokens per second, thinking tokens and time to first token are not compared between agents: Claude Code reports API time while Codex and Antigravity CLI report turn time, Antigravity CLI does not stream thinking, and the tokenizers differ, so the same number means different things. Each agent's own page judges them against that agent's own last 7 days.

Which models and plans are compared?

Claude Code (Claude Opus 5.5), Codex CLI (GPT-6 Luna), Antigravity CLI (Gemini 3.8 Flash), Claude Code (Claude Sonnet 5.5) and Antigravity CLI (GPT-OSS 120B); Claude Code on the Claude Max subscription, Codex CLI on the ChatGPT Free plan and Antigravity CLI on the Google AI subscription. Each runs on one account from one EU server; plan limits, routing and default models are part of what is measured.

How it is measured

Task

fixed modular-exponentiation task with one exact answer (same prompt every hour); every hour at :37 UTC, one fixed task per agent.

Time

End to end, by our own clock: from launching the CLI on our server to the complete answer, including the CLI's own start-up; the median of successful calls.

Failed calls

The CLI errored out, timed out or was rate-limited, or the call cannot count as a clean measurement (a fallback model, a retry, a refusal, an attempted tool call, extra context loaded by the CLI). a failed call is 'not measured', never a wrong answer.

Models

Claude Code with Claude Opus 5.5 (effort high); Codex CLI with GPT-6 Luna (effort high); Antigravity CLI with Gemini 3.8 Flash (effort high); Claude Code with Claude Sonnet 5.5 (effort high); Antigravity CLI with GPT-OSS 120B (effort medium).

Where

one server in the EU (Hetzner, Germany); a single account per vendor.

Data

/api/coding-agents.json (JSON, CC BY 4.0) · markdown · site methodology