Codex 4.8 vs Claude Code, MiniMax M3, Grok Imagine — AI Daily May 31
411 messages · 73 active members
Overview
Topics
Multiple builders reported abandoning Claude Code for implementation work, citing Codex 4.8's superior spec-following, fewer follow-up questions, and surgical fixes on large codebases at xhigh/Extra effort. Claude Code remains preferred for brainstorming and creative work, while Opus 4.8 drew mixed reviews — smart self-directed workflows but heavy token consumption, with several builders staying on 4.7.
MiniMax M3 reportedly beats GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro and surpasses Opus 4.7 on SVG-Bench. At 1T parameters even a $150K GB300 DGX workstation likely needs pairing, so cloud inference remains the practical path. Benchmark integrity skepticism (via a T3 video) tempered the hype.
@victorbrsss quoted $0.03/sec effective on Grok Imagine while @navuud and @mlsmdm cited xAI's official $0.08/sec for grok-imagine-video-1.5-preview — settings drive the spread, and it's still cheaper than Seedance at comparable quality. The viral sub-$0.30 talking-head model was traced to PrunaAI's p-video-animate on Replicate during a 70% launch discount ($0.009/sec at 720p) that expired May 31.
@expadz called Max Fam's course nearly a scam (5 outdated workflow videos in 6 months); consensus from @startropics and @KtargetMedia: Franky Shaw stays current, and a free YouTube tutorial beat most paid options. Builders also shared experiments generating native ads via Codex (celebrity angles, stimulus check creatives) and debated leaning into the AI-looking-creator bullying narrative for engagement.
@arielletolome is scraping FB Ad Library with a local RTX 6000 Blackwell running gemma4 (14-day ETA — @iannagy questioned the timeline given his 8-10 hr runs). Kieran outlined a self-improving creative loop (media buyer → creative strategist with RMBC prompting). Hermes-style orchestration runs parallel agents on separate git worktrees with single-writer-per-file rules, plus a Playwright UI critique prompt that scores pages on hierarchy/spacing/contrast.
Key Takeaways
- Codex 4.8 on xhigh/Extra is the new default for autonomous implementation; Claude Code is being relegated to brainstorming and creative work.
- MiniMax M3 claims SOTA on SWE-Bench Pro and SVG-Bench but at 1T params self-hosting is impractical — cloud is the move, and weight your own evals over leaderboards.
- Grok Imagine runs $0.08/sec official ($0.03/sec effective with settings) and beats Seedance on cost; the viral $0.30 talking-head was PrunaAI's p-video-animate during a now-expired 70% discount.
- Franky Shaw's AI UGC content stays current while Max Fam leans on stale creator-submitted workflows — a free YouTube tutorial beat most paid courses, and quality > quantity wins at scale.
- Run parallel agents on separate git worktrees with one writer per file path, then use an orchestrator to merge and verify; pair with a Playwright UI critique loop that scores pages until no metric drops below 4/5.
Hot Threads
Which AI UGC ad course is actually worth buying? Franky vs Max Fam vs xRoas breakdown
Anyone really pleased with Codex 4.8? What effort level are you running?
Mystery sub-$0.30 AI talking-head video model — which one is it?