Claude + Codex under $200: who plans, who executes, who reviews?

Start with the $120 pair; $200 is a ceiling

Here “$20 Codex” means Codex included with ChatGPT Plus, and “$100 Claude” means the public monthly Max 5x tier. This example is separate from the author’s actual $125 Claude bill and does not inherit that account’s measured token totals. Price alone does not establish a capacity ratio.

My starting arrangement is Claude for daily discussion and planning, Sol for clearly specified implementation, and the other model for important reviews. Claude also executes routine work directly. Keep the executor flexible: a $20 Codex allowance should not be treated as unlimited background capacity.

On mobile, swipe the table horizontally. With a keyboard, focus the table region and use the arrow keys.

What the rankings support—and how they change the division

Checked on 12 October 2026, the AA Intelligence Index v4.3.2 ranks the three configurations differently from a simple Opus-over-Sonnet hierarchy. The rank below is the model page’s 227-model comparison class, not an unrestricted global ranking. Terminal-Bench 4.0 is the more relevant indicator for terminal-based execution.

• Daily Sonnet: its tested Max configuration leads this selection on both metrics. That supports taking Sonnet seriously beyond small tasks. It does not establish the score of Sonnet Medium; use the official daily-work/default-effort guidance for that starting setting, then test it on your work.
• Complex planning or review: Opus High scores above Sol High on this composite and terminal test, making Opus a reasonable escalation candidate. There is no separate architecture-planning or code-review ranking here, and Sonnet Max leads both. Do not reserve every difficult task exclusively for Opus.
• Bounded Sol execution: Sol High retains a 51.5% terminal pass rate at a lower measured API task cost in this test. This supports an efficiency-oriented trial, not a claim that Sol is the highest-quality executor. Difficult implementation can stay with Claude when completion quality matters more.
• Escalation inside Claude: try a higher effort on the current model or Opus on a concrete blocker. Sonnet Max is another candidate when available, but its higher score here comes with higher task spending. Do not make Max the default for every prompt.

On mobile, swipe the table horizontally. With a keyboard, focus the table region and use the arrow keys.

• Artificial Analysis: Sonnet 5.5 Max, Default Fallback
• Artificial Analysis: Opus 5.5 High, Default Fallback
• Artificial Analysis: GPT-6.1 Sol High
• Artificial Analysis: Terminal-Bench 4.0 methodology
• Artificial Analysis: Intelligence Index methodology

For $20 Codex + $100 Claude, use this starting division

The table above supports a capability-and-efficiency tradeoff, not a rigid planner/executor hierarchy. In particular, Sonnet Max leads the selected terminal results; Sol is assigned bounded execution as a workflow experiment, not because it ranks first. The daily Medium settings below were not the configurations tested in that ranking. Keep a successful current model working and use the versions your account actually offers.

On mobile, swipe the table horizontally. With a keyboard, focus the table region and use the arrow keys.

Reverse the pair: GPT $100 + Claude $20

Use ChatGPT Pro as the main workspace and Claude Pro as a focused second opinion. Keep ordinary planning, implementation and tests together in Codex/Work, instead of requiring a Claude consultation before every task. Claude Pro is not a lower-capability model: current Claude Code defaults to Opus 5.5 on Pro as well as Max; actual account and client availability still matters.

The practical plan difference is continuity. Current OpenAI documentation says Pro has no five-hour limit, while weekly limits may still apply. Claude Pro has five-hour and weekly limits shared with Claude Code. That supports reserving Claude for valuable handoffs, without asserting that a $100 plan has five times the capacity of a $20 plan.

• Rank-based rationale: Opus High has a higher AA score and Terminal-Bench result than Sol High in the table above, so a targeted Opus second opinion is a reasonable trial. Those scores do not prove that Opus finds more review defects; independent review is a workflow hypothesis.
• Sol remains the main executor because the selected test shows useful capability at lower average API task cost, combined with the Pro usage-window rules and your choice of workspace. It does not become the higher-ranked model when you swap subscription tiers; API task cost does not establish subscription capacity.
• Do not route every hard task to Opus automatically. Sonnet Max leads this selected benchmark; if available, it is another trial candidate. The Max score is not evidence for Sonnet Medium, and using Max on a smaller plan does not create extra allowance.

On mobile, swipe the table horizontally. With a keyboard, focus the table region and use the arrow keys.

• Compare Pro + Claude Pro for this reverse pair (US)
• OpenAI: current usage windows and shared allowance
• Claude: Pro and Claude Code shared limits
• Claude Code: current Pro model defaults

Example: add provider submissions without building it twice

• Opus plans: define the fields, review queue, duplicate handling, rate limits and acceptance cases. Ask for decisions and a short implementation brief, not a second speculative implementation.
• Sol executes: take one bounded slice, implement it and return the changed files, commands run, results and unresolved questions. Keep one writer in a worktree; parallel writers need isolated worktrees and an integration owner.
• Opus reviews the diff: focus on authorization, untrusted inputs, duplicate handling and failure states. Turn findings into specific issues; Sol fixes confirmed defects and reruns the affected checks.
• You accept the result in the real browser. For a one-line label change, skip the chain and let the current tool finish it.

Switch when the work calls for it, not on a fixed token ratio

Use actual remaining limits and reset times, plus accepted output, to decide. Do not allocate work as 20:100 from subscription prices or compare raw cached-token totals across different workloads. The two-failed-attempt rule is an editorial stopping habit, not a vendor limit.

On mobile, swipe the table horizontally. With a keyboard, focus the table region and use the arrow keys.

Add another subscription only for a missing role

If you use ChatGPT for research, files and deliverables throughout the day, consider making Pro the main workspace and keeping Claude Pro for a second opinion. Assign Sol implementation and Claude focused review. Work and Codex share usage, so research and coding must fit the same allowance.

Add Cursor for an IDE workflow you value: navigation, inline edits and reviewing changes. Do not buy it merely to run the same tasks again. Model availability and its Other Models pool need a separate check; Sol 6.1 access and a fixed token allowance are not assumed here.

Two public $100 plans consume the full $200 before tax. Upgrade only after recording which completed tasks were delayed by each limit. Claude chat and Claude Code share subscription limits; eligible Max API credits are separate from interactive Claude Code capacity.

• Compare the author’s specific Claude Code and Codex observations