Codex CLI vs Claude Code (May 2026): a cross-cut benchmark comparison — SWE-bench, Terminal-Bench, cost

"Which is stronger — Codex CLI or Claude Code? I want a cross-cut comparison on 2026-May official benchmarks, with cost and task fit included, so I can pick the right one in the field." — We answer this using the Terminal-Bench 2.0 official leaderboard, Aider Polyglot, SWE-bench Verified aggregates, plus Anthropic / OpenAI official docs and GitHub releases — through the lens of embedding these agents in AI sales automation.

As of May 2026, the latest versions are Codex CLI 0.130.0 (released 2026-05-08) and Claude Code 2.1.143 (released 2026-05-15). Their backing flagship models are GPT-5.5 (rolled out 2026-04-23) and Claude Opus 4.7 (released 2026-04-16). Benchmark rankings flip depending on which axis you measure, so rather than a simple winner, read these as "task-by-task fit".

Sources: OpenAI Codex official changelog / Claude Code official changelog / Anthropic Newsroom / Terminal-Bench official leaderboard (tbench.ai) / Aider official docs / Claude / OpenAI official pricing pages. We only cite official information at publish time; third-party aggregates are explicitly labeled.

1. What Codex CLI and Claude Code are — May 2026 snapshot

Codex CLI (OpenAI)

• Latest: 0.130.0 stable (2026-05-08). 0.131.0-alpha line in progress.
• Default model: GPT-5.5 (rolled out 2026-04-23). The Codex-tuned GPT-5.3-Codex is selectable.
• Strengths: JSON-RPC 2.0 app-server; codex remote-control for full headless control from external processes; built-in image generation (gpt-image-2).
• Subscriptions: ChatGPT Plus / Pro / Business / Enterprise include Codex (Pro $100 = 5x Plus limits).
• Package: @openai/codex on npm, launched via codex.

Claude Code (Anthropic)

• Latest: 2.1.143 (2026-05-15).
• Default model: Claude Opus 4.7 (2026-04-16). Sonnet 4.6 / Haiku 4.5 selectable. Supports xhigh effort level and Fast mode.
• Strengths: First-class subagents (claude agents), /goal for completion-condition loops, /ultrareview, Plugin / Skill / MCP. 1M-token context standard.
• Subscriptions: Claude Pro / Max / Team include Claude Code. Direct API also supported.
• Package: @anthropic-ai/claude-code on npm, launched via claude.

2. Cross-cut benchmarks — Terminal-Bench, SWE-bench, Aider

Terminal-Bench 2.0 — real terminal tasks

tbench.ai's official leaderboard shows the top 10 as of 2026-05-15:

SWE-bench Verified — real GitHub issues

SWE-bench Verified is the heavyweight benchmark using real GitHub issues. OpenAI paused self-reporting in Feb 2026 over contamination concerns, so current numbers come from third-party trackers like Epoch AI.

• GPT-5.5: 88.7% (OpenAI self-reported, released 2026-04-23)
• GPT-5.3-Codex: 85.0%
• Claude Opus 4.7: ~82% (third-party aggregate)
• Claude Code (Opus 4.6 base agent): 80.9%

Anthropic explicitly states in the Opus 4.7 announcement that "excluding any problems that show signs of memorization, Opus 4.7's margin of improvement over Opus 4.6 holds", signaling transparency on contamination. We treat these numbers as "baselines with ±several percent error".

Aider Polyglot — multi-language code editing

Aider's official leaderboard evaluates against 225 Exercism problems across C++ / Go / Java / JavaScript / Python / Rust.

• gpt-5 (high): 88.0% correct, $29.08 cost (Rank 1)
• gpt-5 (medium): 86.7% correct, $17.69 cost (Rank 2)
• o3-pro (high): 84.9% correct, $146.32 cost (Rank 3)

3. CLI / subagent / plugin feature deltas

Where Claude Code wins

• Long-context investigation: 1M-token context plus subagent isolation suits large-repo overview tasks.
• Completion-condition loops: /goal drives "until all tests pass" or "until lint = 0" in one command.
• MCP / Plugin / Skill: Three mature extension hooks make it easy to inject internal knowledge.

Where Codex CLI wins

• Programmable execution: codex remote-control + JSON-RPC 2.0 enables full control from external scripts — perfect for CI / batch.
• Image generation: image_generation built in. The figures in this article are generated with it.
• Token efficiency: Third-party benchmarks report "about 1/4 the token consumption of Claude Code on the same task" (source: morphllm aggregates, reproducibility requires verification).

4. Token efficiency, context windows, API pricing

Sources: Anthropic Pricing Docs / OpenAI Pricing. Prices in USD; subject to FX moves and official revisions.

Try Sales Claw — the loop CLI agents sit behind.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub

5. AI sales automation fit — Sales Claw perspective

Sales Claw is an OSS tool designed to reduce mis-send and TOS-violation risk through policy control, pre-send automated checks, sales-NG detection, halting on CAPTCHA detection, send-frequency limits, audit log preservation, and auto-stop conditions. When embedding CLI agents into the loop, Codex CLI and Claude Code are complementary, not exclusive. In our own internal verification (May 2026, 100-company sample on the Sales Claw repo), the most stable setup paired Claude Code's /goal loop driving approach guardrails violations to zero with Codex CLI's image_generation producing OG cards in parallel. See our walkthrough on bundling claude agents and codex remote-control into one parallel headless platform for details.

Where Claude Code fits

• Form body generation: 1M context can hold company info, past send history, and approach guardrails in one shot.
• Diff-text checking: /goal "loop until no approach-guardrail violations"
• Large-repo overview: claude agents parallelizes workers across Sales Claw source.
• MCP integrations: Plug in internal knowledge servers (Notion / Slack / Postgres).

Where Codex CLI fits

• Image / OG generation: Built-in image_generation for blog covers and dynamic OG (the figures in this article come from it).
• Nightly batch headless parallelism: codex remote-control + JSON-RPC for N-way parallel from an external scheduler.
• Terminal command sequencing: Codex submits 82.0% on Terminal-Bench 2.0; strong on sh / pwsh workflows.
• Cheap classification / extraction: GPT-5.3-Codex ($1.75 / $14) for short tasks.

Hybrid example

# Run Sales Claw with Claude Code + Codex CLI side by side
# Phase A: Claude Code generates copy (long context + /goal loop)
claude agents \
  --add-dir ./company-data \
  --mcp-config ./mcp/sales-claw.json \
  --permission-mode plan \
  --model claude-opus-4-7 \
  --task "Generate guardrail-compliant copy for 100 companies"

# Phase B: Codex CLI handles OG image + batch verification
codex remote-control --port 7777 &
node scripts/dispatch-og-generation.cjs --port 7777 --count 100

6. Cost estimate and assumptions

Assumptions

• Volume: 10,000 companies / month (Sales Claw standard scale)
• Triage model: Claude Haiku 4.5 ($1 / $5 per MTok)
• Body model: Claude Sonnet 4.6 ($3 / $15) — comparison: GPT-5.3-Codex ($1.75 / $14)
• FX: 1 USD = 150 JPY
• Exclusions: CAPTCHA ~8% / sales-NG ~12% / no form ~15%
• Average tokens per company: ~4,000 in / ~800 out
• Cache hit ratio: 60% (company info / guardrails reused)
• Variance: ±30%

Under identical conditions, GPT-5.3-Codex (input $1.75 / output $14) comes to about ¥12,500 / month. Delta vs Sonnet 4.6 is ~¥3,900 / month. Measure on real data before adopting — the quality vs cost tradeoff is workload-specific.

7. Risk and safety design for no-human-review operation

Sales Claw doesn't ship CLI agent output directly — it applies pre-send automated checks, sales-NG detection, halting on CAPTCHA detection, send-frequency limits, audit log preservation, and auto-stop conditions to reduce risk structurally (see Figure 5 in section 5 for the full flow).

Legal / compliance

• Anti-spam (JP): Sender-info 4-requirement auto-fill (preferences.complianceFooter: true)
• TOS compliance: Pages marked "no sales solicitation" auto-skipped
• No CAPTCHA bypass: Halt with awaiting_approval, audit logged
• Send-frequency limit: Suppress repeated sends to the same domain
• Opt-out line: "If you would prefer not to receive..." auto-inserted

Prevent CLI agent runaway with auto-stop AND-conditions

Residual risk

• Missed-detection of new CAPTCHA schemes (mis-send possible until Sales Claw catches up)
• Lag on TOS-revision response (manual update of compliance metadata required)
• Vertical regulations (BFSI etc. require separate review)
• LLM hallucinations (wrong company info / wrong contact-person names)
• CLI-level bugs (Both Claude Code and Codex CLI ship weekly; behavior can drift)

8. Pre-production checklist + closing thoughts

Closing

As of May 2026, Codex CLI 0.130.0 and Claude Code 2.1.143 are best read as "pick by task; combine where it fits" rather than "one CLI to rule them all". Terminal-Bench 2.0 puts Claude-Opus-4.7-based vix at 90.2% (rank 1) and Codex CLI itself at 82.0% (rank 7); SWE-bench Verified aggregates put GPT-5.5 at 88.7% and Claude Opus 4.7 around 82%; Aider Polyglot puts GPT-5 (high) at 88.0%. There is no single winner — rankings reshuffle per benchmark.

For embedded-product use like AI sales automation, "measure 100 companies of your own list / copy rules / target sites" is a far more trustworthy selection signal than benchmark ranks. Sales Claw doesn't ship CLI output as-is — it reduces risk structurally with pre-send checks, sales-NG detection, CAPTCHA halts, audit log, and auto-stop conditions.

Next step: slice 100 companies from your list, run both Claude Code and Codex CLI for body generation, then compare pre-send-check pass rates and quality side by side. Start at the Sales Claw quick start. The OSS source is also free to download.

This is a convenience English translation of the Japanese-language original. In case of any discrepancy, the Japanese version is authoritative.

When benchmarks don't agree, measure on your own data. Sales Claw closes the loop.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub