Picking a plan for Codex or Claude Code? Weigh SemiAnalysis's 5x value gap against cost per task

#What changed

SemiAnalysis published limit tests of the major AI subscriptions on October 5, 2026. It ran repeated calls that isolate one token type at a time, watched each plan's usage meter move, and priced the result at API list rates for an agentic workload. On the mid-tier models, GPT-6.1 Sol versus Claude Opus 5.5, its chart (shared on X by Peter Gostev) shows about $2,084 a month of API-equivalent value on ChatGPT Pro 200 against $11,726 on Claude Max 20x, with a similar gap of roughly 5.5x at the $100 and $20 tiers. On the flagship models, GPT-6 Astra and Fable 5.1, SemiAnalysis says the limits are quite similar, though Fable can only use half of a Claude plan.

Tibo (Thibault Sottiaux, who works on Codex at OpenAI) quoted a reply from Meet Pandya and called it the real story. That reply shows an Artificial Analysis chart pricing one Intelligence Index task at $0.72 for GPT-6.1 Sol (max) and $5.98 for Claude Opus 5.5 in the max configuration shown. The argument is that API-equivalent dollars count tokens, not finished work, and Sol's tokens are cheaper.

#Who is affected

Anyone choosing or renewing ChatGPT Plus, Pro 100, Pro 200 or Pro 500, or a Claude Pro or Max plan, mainly to run coding agents like Codex or Claude Code. It matters most if you regularly hit weekly limits. According to SemiAnalysis, Pro 200 plans bought before OpenAI's late-September cut keep the older, higher limits until October 29, while new purchases get the halved limits right away.

#What to do now

1. Price your own work, not the headline. For a week, count how many real tasks you finish on each plan before the weekly meter runs out. That number captures both the token allowance and how many tokens each model spends per task.
2. If you are on a grandfathered Pro 200 plan, run that count before October 29 so you can compare the old and new limits with your own numbers.
3. Recheck after any limit or price change. SemiAnalysis says it caught one provider running a small A/B test on limits, so allowances can shift without notice.

#What is not confirmed

OpenAI has not published its own plan-value numbers in response, and Tibo's post is an argument rather than a data rebuttal. The 5x figure is SemiAnalysis's measurement for its own September agentic workload mix, and its full dashboard is paywalled. SemiAnalysis also says it does not think Artificial Analysis's benchmark tasks represent real work, and Artificial Analysis lists Opus 5.5 at medium effort at $1.34 per task, far below the $5.98 max figure. Neither chart measures cost per finished task in real Codex or Claude Code use.