Top AI Models to CHOOSE (Intelligence Comparison)

Anthropic Opus 4.6 and GPT-5.3 Codex are the only two true frontier coding models in February, and the creator is burning ~$200/day on Opus.

The US frontier: Opus vs. GPT-5.3 Codex

Opus 4.6 and OpenAI's GPT-5.3 Codex are called "the absolute coding juggernaut" and "the coding juggernaut" respectively. The creator splits them by personality: Opus is "chatty, sometimes quirky" and walks beginners through problems, while GPT "can get done, but won't really talk to you that much." Both sit at "$5 per million tokens," making them "absolutely the most expensive" options. The creator admits the spend "sounds like a lot, but it's worth it if you're building something big like a dashboard" or fixing a major bug.

Gemini and Grok fall behind

The creator hasn't seriously tested Gemini's new thinking model and notes "people just seem to be ignoring Gemini." Grok is "okay, but coding not so much" — kept in the rotation only because it ships with X Premium and has live X/Twitter data. Llama via Ollama delivers "around 80% of what the top models are providing" for free on local hardware, which the creator rates as "open weight, not bad, but also at the same time, not Frontier."

Chinese models: 95% of the capability, real inconsistency

Minimax M2.5, Kimi, and GLM 5 are "dirt cheap" and "can do almost 95% of what US models are doing," with the gap now "just a few weeks of progress." The Minimax coding plan is "$20 a month" with 300 prompts every 5 hours, which the creator calls "more than good enough" for a "normal person." Kimi's "Aegis swarm" (agent swarm beta) is the most interesting feature; the creator pays ~$40/mo on the Allegretto plan but says he'd "only do this plan for one month."

The catch: inconsistency on critical jobs

Chinese models are "very inconsistent." The creator won't use them when "giving it critical passwords" or shipping anything where a bug gets your X account suspended. The rule of thumb: "if you're just making something for yourself, a small tool for yourself, absolutely just go wild with Minimax" — but anything user-facing goes to Opus 4.6 or GPT.

Practical setup

OpenRouter is recommended for testing every model on one bill. For private data, run Ollama locally instead of paying a subscription.

Don't lock in

"I would never really pay annually for any of these." The creator rotates monthly to chase the next release.

What's next

DeepSeek "was the disruptor, but I don't really see them showing up that much." A successor is rumored in the next few days; if it lands, this ranking changes.