Codex vs Claude - an Honest Comparison

Product(s) reviewed

• Codeex / OpenAI Codex toolset
  • Includes CLI and IDE/desktop/cloud variants
  • The video focuses on the developer/CLI workflow
• Claude Code / Anthropic Claude toolset
  • Includes CLI and IDE/desktop variants
  • Also focused on the developer/CLI workflow

Key features compared

• The agentic “harness” (editor/tooling, context handling, permissions, file editing, running tests) is considered as important as the underlying model.
• Subproducts/features have converged:
  • Skills, hooks, project rules, sub-agents, MCPs
  • Differences are described as minor, so the main differentiator becomes model + harness + value.
• Notable asymmetry in model access by subscription:
  • Codeex $20+ plan includes Soul (flagship)
  • Claude Code $20 plan includes Sonnet 5 (middle tier)
  • Claude flagship (Opus 5) requires $100 plan+ (unless you pay extra usage)

Benchmarks / performance findings (major points)

• Company-published benchmarks are considered unreliable, because each vendor selects benchmarks that make its model look best (“marking their own homework”).
• Independent benchmark tie mentioned:
  • On “Tasteful Solver Rate” (Senior Suite benchmark, refreshed end of July)
  • Fable 5 / Opus 5 / GPT 5.6-Soul are tied at 34.7%
  • Noted as ~100 tasks, so the result isn’t decisive by itself.
• Practical conclusion from demos:
  • On correctness, both models tend to succeed.
  • Main differences are speed, depth, and verbosity/initiative.

Pricing & usage limits (numerical points)

API pricing (flagships roughly similar)

• GPT 5.6 (Soul)
  • $5 / 1M input tokens
  • $30 / 1M output
• Claude Opus 5
  • $5 / 1M input
  • $25 / 1M output (slightly cheaper on output)
• Fable 5
  • $10 / 1M input
  • $50 / 1M output
• Cheaper models big discrepancy
  • OpenAI Luna: $0.20 / 1M input, $1.20 / 1M output
  • Anthropic’s lineup is implied to be less competitive at that low price point

Subscription entry prices (similar entry points; value differs)

• Codeex
  • free tier
  • $8
  • $20+
  • Pro starts at $100
• Claude
  • free tier
  • $20/mo (or $17 annually)
  • $100 max, $200 max (higher multiplier usage)
• The value asymmetry highlighted:
  • Codeex $20 includes Soul
  • Claude $20 does not include Opus

Usage generosity (community estimates; hard to verify)

• Codeex $20+ plan (community estimate)
  • ~10–100 “Soul messages” per 5-hour window
  • Up to ~2,000 messages on Luna
  • Mention: OpenAI temporarily removed the 5-hour cap for paid plans (replaced with effectively weekly limits at time of filming)
• Claude Code Pro (estimate)
  • ~45 prompts per 5 hours
  • Limits tied to the same quota pool as Claude app/website usage
  • Mention: an additional weekly cap

Direct “value” claim from the reviewer’s experience

• For similar tasks, Codeex used ~3–4x less quota than Claude (reviewer’s estimate), making Codex a clear winner on value.

Demo results (what happened, not just claims)

The reviewer ran the same five tasks across multiple model tiers, timed them, and verified outputs.

1) Explain an unfamiliar open-source codebase

• All models produced decent analysis
• Stronger models gave better results
• Opus 5 took ~5x longer than Soul
  • Opus wrote ~635 lines of analysis
  • Plus more verification behavior (tests/scripts, more thoroughness)
• Behavior difference:
  • Claude/Anthropic described as taking more initiative and writing more code
  • Codeex described as more to-the-point

2) Build a complete habit tracker from scratch (FastAPI + SQL + HTML/JS)

• Codeex completed faster, but was less focused on verification
• Opus 5 generated more tests
  • Opus: 27 tests
  • Other models described as more bare-bones
• Optimization emphasis:
  • Claude version: logic verification
  • Codeex version: a working app, including better initial styling in the reviewer’s example

3) Refactor/split a ~5,040-line Python file without breaking anything

• 6 out of 7 models succeeded
• Haiku failed
  • It created separated code
  • But didn’t remove the original file, so the refactor didn’t fully complete
• Large runtime differences
  • Opus 5: ~36 minutes, described as ~5x longer than Soul
  • Reason: Opus wrote a complete “proof tool” / verification mechanism to ensure no regressions
• Theme:
  • Higher-end Claude models do more than requested when prompts aren’t highly specific (longer time, more code, more documentation/testing)

4) Fix a vague/complex planted bug report

• All models found and fixed the issue in the small codebase
• Soul produced the most concise fix (~2 lines)
• Claude models took ~3x longer but still achieved correctness

5) Review a PR with 3 planted bugs (and verify extra issues)

• All models found all 3 planted bugs
• They also found a 4th bug the reviewer didn’t intentionally add
• Claude/Opus took longer but provided a more complete/bulletproof solution
  • More thorough follow-up code and/or explanations

Pros / Cons (as concluded by the video)

Codeex (pros)

• Faster and more efficient on identical tasks (sometimes 4–7x faster cited in recommendation section)
• More generous subscription usage
  • Reviewer’s experience: 3–4x cheaper in quota for comparable tasks
• Flagship model access at $20 (Soul included)
• More to-the-point behavior (less extra code/verification when not required)

Codeex (cons)

• Less depth/verification compared to Opus-class Claude models
  • Sometimes fewer tests and less “proof tooling”

Claude Code (pros)

• More depth and verification
  • Example: 27 tests and “proof tool” behavior during refactor demo
• More initiative
  • Writes more, does extra work even if not explicitly asked
• Potentially better for depth-first tasks where robustness matters

Claude Code (cons)

• Slower and consumes more quota/time for comparable correctness
• Higher-tier access to flagship (Opus) costs more ($100 plan+), making value harder at entry pricing

Overall recommendation / verdict

• Choose Codeex if you want the fastest, most efficient workflow and better quota value—especially since $20 includes flagship Soul.
• Choose Claude Code if you want maximum depth, more tests, and more robustness/extra verification—at a higher time/quota cost and higher price tier for flagship.
• Best practical advice from the reviewer:
  • If affordable, use both (“different failure modes”):
    • one may be faster for a change
    • the other may better catch issues around it

Unique points mentioned (consolidated list)

1. Benchmarks are often vendor-biased; company charts can’t settle the winner.
2. Independent benchmark (“Tasteful Solver Rate”) shows a 3-way tie at 34.7% for Fable 5 / Opus 5 / GPT 5.6-Soul.
3. Harness quality (context/tools/permissions/file editing/running tests) heavily affects results.
4. Subscription availability is asymmetric (Codeex $20 includes flagship; Claude $20 does not).
5. Codeex is described as more open-source (Apache 2.0) vs Claude Code described as partially open source and not fully public.
6. Popularity (GitHub stars): Claude Code ~142K vs Codeex ~106K (as cited).
7. API pricing: Soul and Opus are close; Claude slightly cheaper on output for Opus; OpenAI’s cheapest tier (Luna) is much cheaper.
8. Community estimates suggest Codeex provides much higher prompt/message counts than Claude under subscription windows (with caveats).
9. OpenAI temporarily removed the 5-hour cap on paid plans (replaced with weekly limits at time of filming).
10. Reviewer’s efficiency claim: Codeex tasks were 3–4x cheaper in usage for similar work.
11. Demo 1: Opus wrote ~635 lines and took 5x longer for codebase analysis; Claude is more initiative-heavy.
12. Demo 2: Opus generated 27 tests; Codeex ran faster with simpler verification.
13. Demo 3: Refactor success 6/7 models; Haiku failed to fully complete refactor (didn’t remove original file); Opus took ~36 minutes (~5x longer) due to verification tooling.
14. Demo 4: All models fixed vague bug; Soul produced a 2-line fix; Claude took ~3x longer.
15. Demo 5: All models found 3 planted bugs and also a 4th unintended bug; Claude provided more thorough/bulletproof resolution.
16. Final guidance: differences are mostly speed + depth of explanation, not raw correctness.

Speaker views / roles

• Single main speaker/reviewer:
  • Conducted timing, verification, and demos
  • Provided pricing/usage estimates
  • Recommended:
    • Codeex for efficiency
    • Claude for depth
    • and both if possible
• No other speakers are present in the provided subtitles.