In 30 Sekunden
In a short published March 28, 2026, Matthew Berman gives a rapid, subjective tier ranking of AI products and model families. He puts Claude in S, ChatGPT in A, MiniMax, Gemini and Qwen in B, Grok and DeepSeek in C, Llama and Mistral in D, and separately assigns Llama 4 an F. His judgments are personal preferences rather than standardized comparisons of specific model versions. His prediction that Qwen would stop releasing models did not come true: Alibaba subsequently announced later Qwen releases.
- 1Claude is the favoriteBerman rates Grok C but says Claude is his current favorite and puts it in S.
- 2Open weights and trade-offsHe praises MiniMax’s model-weight releases and rates it B, while ChatGPT earns an A for breadth of features.
- 3Low marks for Llama and MistralHe says he rarely uses Llama and finds Mistral less compelling for his own needs, assigning both D.
- 4Qwen predictionHe likes Qwen but lowers it from A to B over concern about its future; subsequent releases disproved his prediction.
- 5Llama 4 controversyHe criticizes the family’s release and benchmark controversy. Meta released Scout and Maverick, while its announced Behemoth model was still in training at launch.
Das Wichtigste
- Berman’s March 2026 list puts Claude in S tier and ChatGPT in A tier based on his own usage.
- MiniMax, Gemini and Qwen receive B ratings, with Berman praising available model weights and Gemini’s pricing.
- He assigns Grok and DeepSeek C, Llama and Mistral D, and separately criticizes Llama 4.
- His prediction of no further Qwen models was incorrect; Alibaba later announced Qwen3.8-Max.
- A tier list of brands, apps and model families is not a controlled benchmark of particular model versions.
Den bearbeiteten Text nach Kapiteln lesen
Grok, Claude and MiniMax
Berman gives Grok a C, Claude an S and MiniMax a B. His rankings reflect his own experience and appreciation for available weights.
Asked to rank Grok, I give it a C: I do not consider it a great model apart from its usefulness for searching Twitter. Claude, by contrast, is my current favorite. I enjoy using it across tasks and put it in S tier. These are my impressions in March 2026, not results from a controlled head-to-head test.
MiniMax is a great model that I have used a little. I appreciate Chinese companies making capable models available and give extra credit for released weights. I place MiniMax in B because, in my view, it is not quite at the frontier. Open weights and fully open-source releases are not necessarily the same, and access or licensing varies by model.
ChatGPT, Llama, Mistral, Gemini and DeepSeek
ChatGPT earns an A for features. Berman ranks the other families between B and D based on his personal usage and impressions.
ChatGPT gets an A: it has a broad set of features, so it is hard to fault on any single dimension, even if I do not consider it best in class at one particular thing. I put Llama in D because I never use it and question its relevance to my own work. Mistral initially impressed me and has released models openly, but it is not meeting my needs, so I also give it D.
Gemini is a good workhorse with attractive pricing, though I do not use it very much, so it gets a B. I am less impressed by the DeepSeek models available when I am speaking and put them in C. These grades group products or model families rather than specifying identical versions, tasks or testing conditions.
Qwen and Llama 4
Berman praises Qwen but worries about its future, then criticizes Llama 4 and the benchmark controversy. Later Qwen releases make his prediction outdated.
Qwen is my favorite among the open-weight Chinese model families. I would give it an A, but after departures and disruption at the team I worry that few or no further models will appear, so I lower it to B. That was a prediction, not an established fact: Alibaba later announced further models, including Qwen3.8-Max in August 2026.
Asked about other models, I bring up Llama 4. Meta released Scout and Maverick in April 2025 and described the larger Behemoth as still in training. I criticize the Maverick benchmark controversy, including claims about response style, and the absence of the larger release, then give Llama 4 an F. The controversy involved an experimental Maverick version used in LM Arena that differed from the downloadable release; accusations about why it scored well should not be treated as settled proof.
Mehr aus diesem Kanal
34:21EN
Matthew Berman· 27. Okt. 2026Matthew Berman Examines Anthropic’s IPO Filing, AI Growth and GovernanceMatthew Berman, Sam Altman · Forward FutureThe presenter reviews the S-1’s financial metrics, compute strategy and voting structure while questioning the tension between rapid AI development and safety.
0:26ENMatthew Berman· 25. Okt. 2026Sam Altman on Human Values and Problem-Solving in an AI WorldSam Altman · OpenAIIn a short interview excerpt, Altman considers the enjoyment of solving hard problems and the human role in interpreting results.
0:36ENMatthew Berman· 25. Okt. 2026Kimi K3.1 and a Rubik’s Cube Solver DemonstrationMatthew Berman · Forward FutureMatthew Berman shows a scrambled virtual cube being solved and gives the Kimi K3.1 test a favorable assessment.
0:21ENMatthew Berman· 24. Okt. 2026An AI-Generated Kimi Benchmark Chart Used Invented ScoresMatthew Berman · Forward FutureMatthew Berman asks for a graph of Kimi K3.1 Pro benchmarks, then discovers that the displayed figures are illustrative placeholders.
25:50ENMatthew Berman· 24. Okt. 2026Kimi K3.1 Pro Tested: Strong 3D Demos, Uneven Results and Competitive PricingMatthew Berman · Forward FutureMatthew Berman tests the open-weight model on coding, 3D rendering and visual explanations, then compares benchmark results and prices.
32:20ENMatthew Berman· 23. Okt. 2026Who Is Local AI For? The Costs, Trade-offs and a Hybrid FutureMatthew Berman · Forward FutureMatthew Berman weighs data control and independence against hardware cost, setup complexity and inference speed.