Theo's AI Model Tier List Ranking

Theo, creator of T3 Code, released a tier list ranking 15 current AI models from S to F based on four evaluation dimensions. - S tier: Claude Fable 5 (Anthropic), described as the best model ever made. - A tier: GPT-5.6 Sol (OpenAI, top of A, near S), GPT-5.6 Luna, DeepSeek V4 Flash; Sol is his daily default, the latter two are high-value workhorses. - B tier: Kimi K3 (Moonshot AI), first open-weight model to complete complex tasks end-to-end. - C tier: GPT-5.6 Terra, priced awkwardly with no clear use case. - D tier: GLM 5.3, Composer 2.5, Grok 4.6, Muse Spark, Opus 5, Sonnet 5; each has highlights but no practical reason to choose them. - F tier: Gemini 3.7 Flash, Gemini 3.1 Pro, DeepSeek V4 Pro; pricing/efficiency imbalance or flagship models lacking vision capability. - Evaluation dimensions: token efficiency (cost per task, not per token), real task cost (anchored to Cursor Bench, Artificial Analysis data), vision capability (image input support mandatory in 2026), workflow fit (ability to deliver independently without human supervision—the S-tier differentiator).
Key insights
Token Efficiency Over Per-Token Pricing
Theo argues that per-million-token pricing is meaningless; the real metric is total tokens consumed to complete a task. A model with a low per-token price but high token consumption can be more expensive overall. This shifts the cost evaluation from unit economics to task-level efficiency.
Vision Capability Is a Hard Requirement in 2026
Theo heavily penalizes models that do not support image input, as he frequently pastes screenshots into his workflow. In his framework, any model lacking vision in 2026 receives a severe downgrade, reflecting the growing importance of multimodal interaction in real-world usage.
Workflow Fit Separates S-Tier from the Rest
The defining characteristic of an S-tier model is its ability to deliver work independently without human supervision. Claude Fable 5 is the only model that meets this bar, making it the sole S-tier. This criterion emphasizes autonomous task completion over raw benchmark scores.
Gemini Models Placed in F-Tier Due to Pricing and Capability Gaps
Gemini 3.7 Flash, Gemini 3.1 Pro, and DeepSeek V4 Pro all landed in F-tier. Reasons include imbalanced pricing/efficiency and, for flagship models, the absence of vision capabilities. This placement underscores Theo's strict standards on cost-effectiveness and multimodal support.