Why Color Coding the Quality Column
A 70-row model table with a numeric quality column is technically readable but slow to scan. The eye is much faster at parsing color buckets than two-digit numbers. That's why the Quality column on TokenRate's calculator uses four color bands: purple (80+, flagship), sky blue (65–79, balanced), emerald green (50–64, mid-tier), and zinc grey (under 50, budget). Open the calculator, let your eyes go down the column, and within seconds you know the quality landscape — without reading any individual numbers. This post documents what each color means and how to use it for fast model triage.
Purple (80+): Flagship-Tier Quality
Purple badges mark frontier-quality models — GPT-5 (78–80), Claude Opus 4 (80), OpenAI o3 (82), Grok 4 (78), Claude Sonnet 4.7 (80). These are the models you reach for when the task is genuinely hard or when output quality directly drives revenue. Purple correlates closely with Arena AI Elo 1480+ and Artificial Analysis Intelligence Index 78+. The catch: purple-tier models cost 10–30x more than the cheapest credible alternatives. Don't route 100% of traffic to a purple model — use it as your escalation tier in a multi-model router. For specific picks, see top quality LLMs 75+.
Sky Blue (65–79): Balanced Production Default
Sky badges mark the production sweet spot — Claude Sonnet 4.7 (80, just barely upgraded to purple), GPT-5 mini (68), Gemini 2.5 Pro (76), DeepSeek R1 (73), Claude Haiku 4.5 (65), DeepSeek V3 (65), Gemini 2.5 Flash (66), o3-mini (72). This is where 70–80% of production traffic should land: quality good enough for chat, RAG, structured outputs, and most agentic workflows, at prices 5–30x cheaper than purple-tier. Sky is the right tier filter for most teams' baseline routing — the Filter panel calls this 'Good (50+)' but practically you want 65+ for user-facing work. See flagship balanced fast reasoning LLM tiers for the full taxonomy.
Emerald (50–64): Mid-Tier Workhorses
Emerald badges mark mid-tier models — Llama 4 Maverick (62), Llama 4 Scout (55), Gemini 2.5 Flash-Lite (55), Mistral Large (51), GPT-4o mini (51), Qwen 2.5 Coder 32B (58), Codestral (50). These are the volume-throughput tier: not flagship-quality, but reliably above the threshold where outputs become user-visible failures on routine tasks. Use emerald for classification, summarization, intent detection, light extraction, embeddings reranking, and any high-volume work where 'good enough' beats 'perfect but 10x more expensive'. For specific picks at this tier, see best LLMs under $1 per million tokens and underrated bargain LLMs Qwen Mistral Llama.
Zinc (Under 50): Budget Tier, Use Carefully
Zinc badges mark models below the 50 quality threshold — Mistral Small (42), Llama 3.1 8B (30), Llama 3.2 1B (15), Mistral Nemo (38), older Llama 3 variants. These are usable but error-prone — instruction-following gets shaky, structured outputs need extra validation, hallucination rate rises. Right use cases: synthetic data generation (sampling errors average out), pre-filter triage stages, embedding-adjacent classification, on-device inference where size matters more than quality. Wrong use cases: anything user-facing, anything where errors aren't caught downstream. The Quality preset 'Rated only' on the Filter panel leaves zinc-tier models in the list; 'Good (50+)' hides them. For the trap to avoid, see why the cheapest LLM isn't always the best value.