Codex vs Claude, Cloaking Comeback, Anthropic Pricing — AI Daily Jun 24
556 messages · 88 active members
Overview
Topics
Builders converged on Codex as the better value play — @jarvisballer reports running it for days without hitting limits, while Claude Code burns tokens faster and stops dead at quota. Workflow consensus: PRD in Claude, build in Codex, then cross-review. @thewildzeno notes neither model catches every issue in its own code, so adversarial review between them is critical. Bonus: Codex CLI includes gpt-image-2 in your sub.
Long debate on whether Claude's $200 sub is subsidized or whether API pricing carries massive margin. @tidemid alleged Chinese resellers (yunwu.ai) buy Claude/OpenAI OAuth plans with stolen cards and sell inference traces to labs like GLM at 70-90% discounts, framing it as a key reason training costs stay 'reasonable' for cheaper Chinese models.
@basant shared that Codex improved his 8-year-old JS cloaker with 20+ signal detection, extending Taboola client site lifespans from days to 3+ weeks — and praised Codex for executing gray-hat work without Claude's moralizing. Separately, builders reverse-engineered viral animated eggplant/banana doctor ads, concluding no cloaking is needed since FB's AI only sees produce. @arielletolome shipped her own 'Officer Ariel' parody series.
@taszacharis asked about always-on agents for sourcing marketing and creative hires. The group shared a battle-tested funnel: Typeform with knockout questions → ClickUp auto-sort → LinkedIn check → videoask → test. @WhiskeyATX's team filtered 400 applicants to a shortlist without touching a resume.
@iamgalba shared a full pipeline: metapi.io for Meta ad scraping, Bright Data/Apify for VOC mining on Amazon/Reddit/Trustpilot, then auto-repurposing winners into search/image/video and matching presell pages. Parallel thread: @wizardwu advocates test in ABO, scale in CBO, but split CBOs by angle since mixing TAMs causes FB to overspend on smaller pockets and spike CAC.
Key Takeaways
- Codex at $20/mo is dominating Claude Code on value — rarely hits limits, runs 11+ hour sessions, includes gpt-image-2; use both adversarially since neither catches every bug in its own code
- Chinese token resellers (yunwu.ai) are arbitraging stolen Claude/OpenAI plans at 70-90% off — likely a source of training traces for cheaper Chinese models
- FB ad AI doesn't recognize implied nudity in animated produce ads — eggplant/banana yapper ads run uncloaked; cloaking is still alive for Taboola affiliate traffic via Codex-improved JS
- Knockout questions in Typeform + ClickUp auto-sort can filter 400 applicants to a shortlist with zero resume reading
- For Meta scaling, split CBOs by angle — mixing angles with different TAMs causes FB to overspend on smaller pockets and spike CAC; VOC-mined ad copy reads as real customer language
Hot Threads
Codex vs Claude Code value, cloaking comeback for Taboola, and downgrading Claude to $20/mo
Why Anthropic charges so much: R&D, margins, and Chinese resellers training on leaked traces
Viral eggplant/banana animated doctor ads and whether they need cloaking on FB