Output pricing for tracked models in 2026 spans from $0.28/M tokens (DeepSeek V4 Flash) to $50/M (Claude Fable 5) — a 178× spread. Here is how the market actually segments, using the data behind our price-performance leaderboard.
The three price tiers
- Utility tier (< $1/M out): DeepSeek V4 Flash ($0.28), Hunyuan Hy3 ($0.56), DeepSeek V4 Pro and MiMo V2.5 Pro ($0.87). Fast enough and smart enough for classification, extraction and high-volume agents.
- Workhorse tier ($1–10/M out): GPT-5.6 Luna ($1.2), Kimi K3 and the mid-tier Gemini models. This is where most production workloads land.
- Frontier tier ($25–50/M out): Claude Opus 5 ($25), GPT-5.6 Sol ($30), Claude Fable 5 ($50). You pay for the last 10–15 benchmark points.
Price per intelligence point
Divide output price by Podium Score and the picture changes: Kimi K3 delivers 83.3 points at a workhorse price, while the frontier tier charges 10–30× more for single-digit score gains. That is exactly what the value leaderboard ranks — intelligence per dollar, blended with API pricing.
Input tokens matter for agents
Agentic workloads consume 10–50× more input than output tokens. Models like GPT-5.6 Luna ($0.2/M in) and DeepSeek V4 Flash ($0.14/M in) are disproportionately cheap for long-context agents; check long-context rankings before choosing.
Bottom line
There is no single "cheap" answer — there is a cheap tier for each task type. Compare live prices for any four models on the comparison page.