Don't Pick Models Off Headlines — Benchmark Them
· 3 hours ago
Days like today are exactly why you shouldn't pick models off headlines.
Leaks, rumors, and two-word quotes are fun — but the question is which model wins on the tasks you actually run.
Compare them on real tasks thevibefather.com/benchmarks
More from the feed
Announcement
MiniMax Token Plan — I’m in the Top 1.1% of Users
Top 1.1% on MiniMax Token Plan. Not a flex for flex’s sake — a receipt for how hard agentic coding burns tokens when you actually ship.
My MiniMax usage report (data through 21 Jul 10 to 00 UTC)
- 76 days on Token Plan
- 6.39B tokens total consumption
- Usage ranking. Top 1.1%
- 64 active days, peak day Apr 26, 2026 at 502.59M tokens
That is the whole pitch of Token Plan in one screenshot. One subscription unlocks the latest MiniMax models — frontier coding, 1M ultra-long context, native multimodality — with text, image, and speech sharing the same quota. When you live in long agents and fat repos, a shared quota beats three separate meters.
The multi-model reality (Codex + Claude Code too)
I do not run one religion. I run a harness.
Codex / OpenAI side of the ledger
- 12.1B lifetime tokens
- 1.3B peak tokens
- Longest chat 3h 51m
- Longest streak 31 days
- 645 chats, 510 total skills used, 27 skills explored
- Heaviest plugins.
@sites(171),@browser(140),$seo-scale-2026(70)
Claude Code side of the ledger
- 145 sessions
- 122,098 messages
- 173.8M total tokens
- 32 active days, longest streak 14d, current streak 6d
- Favorite model Opus 4.8
- Roughly 648× more tokens than Moby-Dick
Same operator. Three stacks. That is vibe coding in 2026 — portfolio seats, not monogamy.
10% off MiniMax Token Plan (invite)
MiniMax is running Invite Friends, Win Together
- Friends get 10% off plus Builder benefits when they subscribe
- Referrers earn 10% cashback plus community perks
If you want Token Plan with the discount
Get MiniMax Token Plan — 10% off via my link →
Disclosure, that is my referral link. You get the subscriber discount. I may earn cashback. The usage numbers above are from my own product dashboards — MiniMax, Codex, and Claude Code — not marketing mockups.
Announcement
Codex Hits 10M Weekly Users — Paid Usage Reset Today
10 million. That is the new weekly active user number for ChatGPT Work + Codex.
OpenAI’s own “agentic breakout” chart goes from about 2M on March 16 to 6M on July 12, then almost doubles to 10M on July 21 — roughly 2× growth in a week. That is not a soft curve. That is a spike.
And for paid users of Codex and ChatGPT Work, there is a practical gift with the milestone. Usage resets today, July 21. New day, fresh allowance. If you were sitting on the limit, this is the green light to open the tools again.
I care about this less as a vanity scoreboard and more as a signal that agentic coding seats are becoming mainstream volume, not a niche experiment. Capacity chases demand. Resets show OpenAI is still willing to unclog the pipes when the user count jumps.
Primary source Tibo on X — 10M + usage reset
Tip
Tip, Keep a Local Copy of the Weights You Depend On
Quick tip, given the news, if a Chinese open-weight model matters to your workflow, keep a local copy of the weights.
Hosted access can change — a policy, an API shutdown, a region block. Local weights can't be revoked. This is prudent, not panic.
Why this is suddenly relevant White House Weighs Executive Order on Chinese Open-Weight AI
