LLM Pricing History (2023–2026)
In March 2023, GPT-4 launched at $30 / $60 per 1M tokens. Three years later, comparable-quality models (DeepSeek V4 Flash, Gemini 2.5 Flash) ship at $0.07–$0.30 per 1M tokens — a ~99% price drop. This page tracks every major model launch and price change, cross-referenced from provider pricing pages and Wayback Machine snapshots.
Data is free to cite (CC BY 4.0). For live pricing on every current model, see the cost calculator.
The 99% drop, in one chart
Blended cost (70% input + 30% output) of the cheapest frontier-class model available each quarter:
| Quarter | Cheapest frontier model | Blended $/1M |
|---|---|---|
| Q1 2023 | GPT-4 8K | $39.00 |
| Q4 2023 | GPT-4 Turbo | $16.00 |
| Q2 2024 | GPT-4o (launch) | $8.00 |
| Q3 2024 | Claude 3.5 Sonnet | $6.60 |
| Q1 2025 | DeepSeek V3 | $0.52 |
| Q3 2025 | Gemini 2.5 Flash | $0.96 |
| Q1 2026 | DeepSeek V4 Flash | $0.13 |
| Q2 2026 | DeepSeek V4 Flash | $0.13 |
Full launch & price history
Every major model launch, its initial list price, and its current price if still available. Prices in USD per 1M tokens.
| Model | Provider | Launched | Launch in/out | Today in/out | Price change |
|---|---|---|---|---|---|
| Llama 4 Scout 17B | Meta | Apr 2025 | $0.11 / $0.34 | $0.11 / $0.34 | flat |
| Gemini 3.1 Pro | Apr 2026 | $1.25 / $10.00 | $1.25 / $10.00 | flat | |
| Gemini 2.0 Flash | Dec 2024 | $0.15 / $0.60 | retired | — | |
| DeepSeek V3 | DeepSeek | Dec 2024 | $0.27 / $1.10 | retired | — |
| Gemini 1.5 Pro | Feb 2024 | $3.50 / $10.50 | $1.25 / $5.00 | −58% | |
| Mistral Large | Mistral | Feb 2024 | $8.00 / $24.00 | $2.00 / $6.00 | −75% |
| DeepSeek V4 Flash | DeepSeek | Feb 2026 | $0.07 / $0.27 | $0.07 / $0.27 | flat |
| Mistral Medium 3.5 | Mistral | Jan 2026 | $0.40 / $2.00 | $0.40 / $2.00 | flat |
| Claude 2 | Anthropic | Jul 2023 | $8.00 / $24.00 | retired | — |
| Llama 2 70B (Replicate) | Meta | Jul 2023 | $0.65 / $2.75 | retired | — |
| GPT-4o mini | OpenAI | Jul 2024 | $0.15 / $0.60 | $0.15 / $0.60 | flat |
| Llama 3.1 70B (Groq) | Meta | Jul 2024 | $0.59 / $0.79 | $0.59 / $0.79 | flat |
| Claude 3.5 Sonnet | Anthropic | Jun 2024 | $3.00 / $15.00 | retired | — |
| Claude Opus 4.8 | Anthropic | Jun 2026 | $5.00 / $25.00 | $5.00 / $25.00 | flat |
| GPT-3.5 Turbo | OpenAI | Mar 2023 | $2.00 / $2.00 | $0.50 / $1.50 | −60% |
| GPT-4 (8K) | OpenAI | Mar 2023 | $30.00 / $60.00 | retired | — |
| Claude 3 Opus | Anthropic | Mar 2024 | $15.00 / $75.00 | retired | — |
| Gemini 2.5 Pro | Mar 2025 | $1.25 / $10.00 | $1.25 / $10.00 | flat | |
| GPT-5.5 | OpenAI | Mar 2026 | $1.25 / $10.00 | $1.25 / $10.00 | flat |
| GPT-4o | OpenAI | May 2024 | $5.00 / $15.00 | $2.50 / $10.00 | −41% |
| DeepSeek V2 | DeepSeek | May 2024 | $0.14 / $0.28 | retired | — |
| GPT-4 Turbo | OpenAI | Nov 2023 | $10.00 / $30.00 | retired | — |
| Mistral 7B | Mistral | Sep 2023 | $0.25 / $0.25 | retired | — |
| Claude Sonnet 4.6 | Anthropic | Sep 2025 | $3.00 / $15.00 | $3.00 / $15.00 | flat |
Five lessons from three years of LLM pricing
1. Frontier price halves every ~6 months. The cheapest frontier-class model has dropped ~50% per 6-month window since GPT-4 launched. Budgeting AI spend on today's prices is a mistake; plan for prices to halve before your next planning cycle.
2. Output tokens are sticky, input tokens collapse. Input pricing has fallen 99% since 2023. Output pricing has only fallen ~80%. The structural reason: output is autoregressive, so it can't be batched the same way as prefill.
3. New tiers replace, they don't compete. OpenAI didn't lower GPT-4 prices to match Claude — they shipped GPT-4o at a lower starting price. Treat each generation as a new SKU and re-pick your model every release.
4. Open-weight pressure works. Llama 2 dragged Mistral and Anthropic prices down. DeepSeek V2/V3/V4 reset the floor again. Watch open releases; they're the leading indicator for hosted prices.
5. The cheap models eat the workloads. The fastest-growing line items aren't frontier inference — they're cheap-tier batch jobs (classification, extraction, embedding). The unit cost is rounding error; the volume is enormous.
Cite this data
Released under CC BY 4.0. Free to reuse in articles, talks, blog posts, and research — please link back to llmcalculator.net/llm-pricing-history.
LLMCalculator.net. (2026). LLM API pricing history (2023–2026) [Dataset]. https://llmcalculator.net/llm-pricing-history@misc{llmcalculator_pricing_history_2026,
author = {{LLMCalculator.net}},
title = {LLM API Pricing History (2023--2026)},
year = {2026},
url = {https://llmcalculator.net/llm-pricing-history},
note = {Dataset, CC BY 4.0}
}LLMCalculator.net (2026). "LLM API Pricing History 2023–2026." https://llmcalculator.net/llm-pricing-history (CC BY 4.0).Source: <a href="https://llmcalculator.net/llm-pricing-history">LLM Pricing History — LLMCalculator.net</a> (CC BY 4.0)