Intelligence & cost · updated Oct 2, 2026
The same intelligence keeps getting cheaper.
Every level of AI capability starts with one lab, one closed model, and a high price. Then the level fills up with competitors and downloadable models until it costs almost nothing. GPT-4-level intelligence is now 650× cheaper than at its debut.
Pick a level of intelligence
Intelligence 125 first arrived in Mar 2023 with GPT-4 (OpenAI), at $37.50 per million tokens. Today 130 models from 14 labs reach it, 56 of them open-weight. The cheapest, gpt-oss-20b, costs $0.058: 650× cheaper.
First reached
Mar 2023
GPT-4, OpenAI · launch price $37.50
Available today
130 models
from 14 labs · 56 open-weight · 179 have ever reached it
Cheapest today
$0.058
gpt-oss-20b · 650× cheaper than at debut
Open-weight lag
14.8 months
Qwen2-72B (Alibaba) was the first downloadable model here
| Model | Lab | Weights | ECI | Cheapest $ / M tokens | Via |
|---|---|---|---|---|---|
| gpt-oss-20b2025-08 | OpenAI · US | open | 137.8 | $0.058 | DeepInfra |
| DeepSeek-R1-Distill-Qwen-14B2025-01 | DeepSeek · China | open | 135.4 | $0.070 | Nscale |
| gpt-oss-120b2025-08 | OpenAI · US | open | 139.9 | $0.070 | DeepInfra |
| Phi-42024-12 | Microsoft Research · US | open | 130.4 | $0.087 | DeepInfra |
| Qwen2.5-32B2024-09 | Alibaba · China | open | 128.5 | $0.095 | Nebius |
| Gemma 3 27B2025-03 | Google DeepMind · US | open | 130.0 | $0.10 | DeepInfra |
| DeepSeek V4 Flash 07312026-07 | DeepSeek · China | open | 154.5 | $0.10 | DeepInfra |
| Qwen3.5-9B2026-02 | Alibaba · China | open | 139.5 | $0.11 | DeepInfra |
| DeepSeek-V4-Flash2026-04 | DeepSeek · China | open | 146.1 | $0.11 | DeepInfra |
| Qwen3-32B2025-04 | Alibaba · China | open | 138.5 | $0.12 | OVHcloud |
| Qwen3-14B2025-04 | Alibaba · China | open | 138.2 | $0.12 | Nebius |
| Gemini 2.5 Flash-Lite (Jun 2025)2025-06 | Google DeepMind · US | closed | 133.9 | $0.13 | Oracle Cloud |
Showing the 12 cheapest of 130 available models.
Source: Capability levels use Epoch AI's Capabilities Index (ECI), which stitches many benchmarks into one score anchored at Claude 3.5 Sonnet = 130 and GPT-5 = 150. A model reaches a level when its ECI estimate is at or above it. Prices are providers' list prices per million tokens, blended 3:1 input to output. · Epoch Capabilities Index · LiteLLM price map
Launch price versus today, at every level
Each row is one level of capability. The orange dot is what it cost when the level was new; the blue dot is the cheapest way to buy it today.
The frontier itself stays expensive
What collapses is the price of last year's best. The most capable models today still cost dollars per million tokens:
- Claude Opus 5.5 · ECI 167.3$8
- GPT-6 Astra · ECI 166.5$20
- Claude Sonnet 5.5 · ECI 165.2$4
- Claude Fable 5.1 · ECI 164.8$20
- Claude Opus 5 · ECI 162.9$10
- GPT-5.5 Pro · ECI 162.4$67.50
How to read this, and its limits
Per token, not per task. Reasoning models spend more tokens per answer, so a per-token price understates their cost per job. Epoch AI's per-question analysis, which accounts for this, still finds the price of a fixed level of performance falling about 13 times a year. The plunging price of thought.
A benchmark score is not a job. Matching a model's capability index does not mean matching it on every task you care about. Treat levels as broad classes, and read the confidence intervals in Epoch's data before splitting hairs.
Launch prices. For models that debuted before this tracker started, launch prices come from archived copies of each lab's own pricing page or launch post. Since then, the sync records each model's first price itself. When a premium tier reached a level first, the debut price is the cheapest option within that level's first month.
Who you buy from matters. The cheapest listing is often an inference host serving an open-weight model, sometimes in a reduced-precision version. List price is not always what large customers pay.