From S-Tier to F-Tier: A Developer's Candid Ranking of Modern LLMs

A recent developer-centric analysis has produced a comprehensive tier list of contemporary Large Language Models (LLMs), evaluating their performance, cost-efficiency, and suitability for real-world software development tasks. Anthropic’s Fable 5 emerged as the sole S-tier model, lauded for its “next-generation” intelligence and code quality, albeit requiring careful management. OpenAI’s GPT-4o Sol secured an A-tier placement, celebrated as an “unbelievably intelligent” default model with high token efficiency and strong iOS capabilities, despite minor initial setup challenges. OpenAI Luna and Deepseek V4 Flash also achieved A-tier, recognized for their exceptional value, speed, and utility in non-code tasks like summarization, with V4 Flash gaining points for being open-weight despite its lack of vision.

Conversely, Google’s Gemini series (3.7 Flash, 3.1 Pro, 3.5 Flash) landed firmly in the F-tier, heavily criticized for “absurd incompetence” due to poor token efficiency, escalating costs, and unreliability, with Gemini 2.0 Flash noted as a significant regression from its S-tier predecessor. Deepseek V4 Pro was also relegated to F-tier for its high price combined with a surprising lack of vision capabilities. Anthropic’s Opus 5, initially appearing “incredible,” was ultimately placed in D-tier due to its “mimic behavior” where outputs seemed intelligent but failed in practical code merging. The evaluation highlighted crucial factors beyond raw per-token pricing, emphasizing token efficiency and essential features like vision support. Open-weight models such as Kimmy K3 (B-tier) were praised for pushing frontiers but cautioned that perceived cost advantages might be negated by hosting expenses and restrictive licensing terms. The analysis underscores a significant capability gap between top-tier models that can handle complex, autonomous tasks and lower-tier options requiring constant human oversight.