All promotions
GuidesDashboardMCP

The price of the frontier

Take every model that set a new intelligence record when it shipped — 13 of them among 111 scored models — and ask what each costs today. The answer swings wildly: the record has been held at $0.21/1M (Phi 4 (DeepInfra)) and at $60/1M. Being the best and being expensive are not the same thing.

Jul 20242025Jul 20252026Jul 2026$0.2$0.5$1$3$10$30$60Mixtral 8x22B Instruct (Mistral) — new record: score 122 on 2024-04-17 · today $8/1MGPT-4o mini — new record: score 126 on 2024-07-18 · today $0.75/1MQwen 2.5 72B (DeepInfra) — new record: score 129 on 2024-09-19 · today $0.76/1MPhi 4 (DeepInfra) — new record: score 130 on 2024-12-11 · today $0.21/1MDeepSeek Chat (DeepInfra) — new record: score 146 on 2024-12-26 · today $1.21/1MClaude Sonnet 4.5 (Anthropic) — new record: score 147 on 2025-09-29 · today $18/1MGemini 3 Flash Preview (Google) — new record: score 152 on 2025-12-17 · today $3.5/1MGemini 3.1 Pro Preview (Google) — new record: score 155 on 2026-02-19 · today $14/1MGPT-5.3 Codex (OpenAI) — new record: score 157 on 2026-02-24 · today $15.75/1MClaude Fable 5 (Anthropic) — new record: score 163 on 2026-06-09 · today $60/1MClaude Opus 5 (Anthropic) — new record: score 164 on 2026-07-24 · today $30/1MClaude Fable 5.1 (Anthropic) — new record: score 166 on 2026-09-01 · today $60/1MGPT-6 Astra (OpenAI) — new record: score 169 on 2026-09-03 · today $60/1M

Each dot: a model the day it set a new Epoch Capabilities Index record among scored models in this index. Horizontal line: its reign. Y axis: today's cheapest blended price (log) — not the launch price. Hover for details, click for the model's page.

This is the frontier through time: who has held the record, and what that record costs now. The site's other frontier is a snapshot of the present — the efficient frontier asks which models, at today's prices, are worth buying at all.

Every record-holder

Record setModelLabScoreToday /1M
2024-04-17Mixtral 8x22B Instruct (Mistral)Mistral122$8
2024-07-18GPT-4o miniOpenAI126$0.75
2024-09-19Qwen 2.5 72B (DeepInfra)Alibaba129$0.76
2024-12-11Phi 4 (DeepInfra)Other130$0.21
2024-12-26DeepSeek Chat (DeepInfra)DeepSeek146$1.21
2025-09-29Claude Sonnet 4.5 (Anthropic)Anthropic147$18
2025-12-17Gemini 3 Flash Preview (Google)Google152$3.5
2026-02-19Gemini 3.1 Pro Preview (Google)Google155$14
2026-02-24GPT-5.3 Codex (OpenAI)OpenAI157$15.75
2026-06-09Claude Fable 5 (Anthropic)Anthropic163$60
2026-07-24Claude Opus 5 (Anthropic)Anthropic164$30
2026-09-01Claude Fable 5.1 (Anthropic)Anthropic166$60
2026-09-03GPT-6 Astra (OpenAI)OpenAI169$60

Intelligence = Epoch Capabilities Index by Epoch AI (CC BY 4.0), current scores — the frontier is computed among the 111 scored models of this index, so early records predating it are missing. Verified 2026-10-23. See also the falling price of intelligence and the age discount ·methodology.

FAQ

What does the best available LLM cost?

Right now the intelligence record among scored models in this index is held by GPT-6 Astra (OpenAI) (score 169), at $60/1M blended. But the frontier hasn't always been expensive: Phi 4 (DeepInfra) held the record at just $0.21/1M.

Is the price of the frontier going up or down?

Both, in waves. 13 models have successively held the intelligence record since 2024, and the record-holder's price has swung between $0.21 and $60/1M. Open-weight labs periodically crash the frontier's price; premium flagships reset it upward.

Are these launch prices?

No — each dot shows the model's cheapest price today, since historical launch prices aren't tracked for most of these models. Read it as: what each era's record-holder costs now. Prices mostly fall after launch, so early frontier models were dearer at the time than shown here.