---
title: "LLM API Pricing History: Cost per Million Tokens"
date: 2026-10-07
url: https://modevc.com/charts/llm-api-pricing-history/
keywords: ["llm pricing comparison", "ai inference cost comparison", "llm price per million tokens"]
publisher: Mode VC
---

# LLM API Pricing History: Cost per Million Tokens

Frontier input prices fell 56% a year at OpenAI and 40% a year at Anthropic's top tier. Every competitor gets the same price cut.

## Chart: OpenAI's headline model fell from $30 to $2 per million input tokens since Jun 2023: 93% in total, 56% a year.

List price per million input tokens at launch, US$, log scale. Points are launches; lines join them.

| Series | Date | Value | Note |
|---|---|---|---|
| OpenAI headline | 2023-06-13 | $30 | GPT-4 |
| OpenAI headline | 2023-11-06 | $10 | GPT-4 Turbo |
| OpenAI headline | 2024-05-13 | $5 | GPT-4o |
| OpenAI headline | 2026-09-29 | $2 | GPT-6.1 Sol |
| Anthropic Opus tier | 2024-03-04 | $15 | Claude 3 Opus |
| Anthropic Opus tier | 2025-11-24 | $5 | Claude Opus 4.5 |
| Anthropic Opus tier | 2026-10-07 | $4 | Claude Opus 5.5 |
| Anthropic Sonnet tier | 2024-03-04 | $3 | Claude 3 Sonnet |
| Anthropic Sonnet tier | 2026-09-28 | $2 | Claude Sonnet 5.5 |

Note: Opus 5.5 is the 7 Oct list price; its launch-day price was not captured. "Headline model" is our editorial grouping, not a vendor tier. Input prices only; output prices, caching discounts and tokens used per task also move cost. Annual rate = compound rate between first and last point.

## The signal

- **Launches:** [GPT-6.1 Sol](https://openai.com/index/introducing-gpt-6-1-sol/) (29 Sep): "Near-Astra agentic coding and computer use at one-fifth the API price." (OpenAI's claim).
- **Pricing:** Anthropic's [pricing page](https://platform.claude.com/docs/en/about-claude/pricing) lists Claude Opus 5.5 at $4 per million input tokens (captured 7 Oct).
- **Launches:** [Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) (28 Sep): "Faster and cheaper than Sonnet 5 on most work, with better agentic coding scores." (Anthropic's claim).

## The Mode read

GPT-6.1 Sol launched on 29 September. OpenAI says it does near-frontier agent work at one-fifth of the earlier API price. The same week, Anthropic listed Claude Opus 5.5 at $4 per million input tokens.

Join the launch prices and the slope is steep. OpenAI's headline model went from **$30** (GPT-4, Jun 2023) to **$2** (GPT-6.1 Sol): 93% in total, 56% a year compounded. Anthropic's top tier went from $15 to $4, 40% a year.

For an AI-native company, model spend is the biggest cost of goods. When it falls by about half each year, gross margin grows without any work. But every competitor gets the same cut. The margin stays only with the company whose customers would not leave if it kept its price while its costs fell.

> Cheaper inputs help everyone. Only distribution lets you keep the savings.

**What we would back:** Businesses priced on outcomes or seats, not tokens, so falling model costs turn into margin.

## Sources

- [OpenAI and Anthropic launch posts and pricing pages](https://www.anthropic.com/news/claude-3-family)
- [Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing)
- [GPT-6.1 Sol announcement](https://openai.com/index/introducing-gpt-6-1-sol/)