AI Model Tier List for Agentic Workflows (April 2026)

Claude is S-tier, but Anthropic is bleeding trust

Claude earns the top spot because every competitor benchmarks against Claude Opus and Meta engineers are using it to build. The creator calls it "undisputed king" and even floats a god-tier above S. The catch: it is "the most expensive model," and Anthropic has been silently cutting plan allowances — Z.AI, Kimi, and MiniMax all tell you exact request counts, Anthropic does not. Worth staying on, but watch your bill.

GPT 5.4 is the real challenger, not Gemini

GPT 5.4 lands in S-tier as the ~95% Claude option. It matches Claude on Wild Claw Bench real-world tasks and wins on cost, which is why users are migrating. OpenAI killed Sora this month to refocus on core models and is also trying to beat Claude Cowork. Gemini 3.1 Flash lands in C — the team says "don't bother," noting it loses to Kimi, MiniMax, and even Step on Wild Claw Bench, and is squeezed on price by every Chinese lab.

MiniMax M2 is the best Chinese daily driver right now

MiniMax M2 sits at A-tier and is the most usable Chinese model: generous coding/token plan, full transparency on request counts, and 100 tokens/sec off-peak (which means US users always sit in off-peak). The catch: pure agentic performance is "lacking a little bit," and the cheap base/Flash variants are unusable — the V2 Flash is "basically like Haiku. Don't touch that."

Kimi, Step, Grok, and Xiaomi Mimo form the A/B tier

Kimi gets a rare A+ because several crypto orderflow platforms only accept its API keys, and the swarm ecosystem is real. Step is B-tier — cheap and fun, but "the Chinese are fighting their asses off." Grok is A-tier for everyday/X work, but the creator warns: if Grok keeps failing your code, "it's not you. It's Grok" — it ranks below Kimi, MiniMax, and Step on Wild Claw Bench. Xiaomi Mimo is the only Chinese model with a 1M context window and beat Gemini on agent tasks; pro is solid, non-pro is not.

DeepSeek is frozen and Z.AI is on probation

DeepSeek has shipped nothing since before Chinese New Year 2025 — you are "using 2025 tech in 2026" and it is not even listed on the Wild Claw benchmark. Their new bundled-token paper is genuinely novel research the West is not doing, so expect an S-tier jump on the next release. Z.AI 5.1 one-shotted a game two days before recording, which is impressive, but it has been "unstable over this weekend" with rate limits and congestion after the CEO publicly invited stress tests. Currently A-tier, pending re-test.

Trinity and Hermes are placeholders

The creator admits they "didn't do our homework" on Trinity (A-tier placeholder) and Hermes Agent (B-tier placeholder). The full living tier list lives at boxminingai.com and will be updated as April 2026 progresses.