LLM Price History, Charts & Trends

Track how LLM inference pricing has evolved across providers. Historical token price trends, cost reduction curves, tier breakdowns, and forward-looking analysis — updated regularly.

Frontier prices fell ~60% in 12 months

GPT-4-class models dropped from $30/1M input tokens in mid-2023 to under $5 by mid-2025, driven by efficiency gains and competition from open-weight alternatives.

Budget tier now under $0.10/1M

Llama 3.1 8B and similar models are available for $0.05–$0.10/1M input tokens on Groq, Together AI, and Fireworks — 300× cheaper than frontier models from 2023.

Context windows grew 8× in 2 years

The average supported context window expanded from ~16K tokens in 2023 to 128K–1M tokens by 2025, enabling new long-document and multi-turn use cases.

Throughput up 5× on specialized hardware

Groq and Cerebras leveraged custom silicon to push throughput past 1,000 tokens/sec — a 5× improvement over GPU-based inference from 2023.

Historical TrendsJan 2023 – Jun 2026

Median input price per 1M tokens (USD) by tier — log scale

Live Pricing Table

— current rates across all providers

Best for:

0 of 0 models

Provider / Model	Tier	Input /1M	Output /1M	Cached /1M	TTFT	Throughput	Context	Intel.	Caps
Loading pricing data…

Prices are per 1M tokens in USD. Historical data is sourced from provider announcements and pricing pages. TTFT = median time-to-first-token. Throughput = output tokens/sec. Verify current prices before production use.

Build vs. Buy

Running your own model on rented GPU compute can be significantly cheaper at scale. A single H100 at ~$2.50/hr can serve ~500K tokens/min of Llama 3.3 70B — compare that against API costs for your expected volume. GPU pricing table and benchmarks page to estimate your self-hosted cost.

FAQ — LLM Pricing History & Trends

How much have LLM API prices dropped since 2023?

Frontier model input prices dropped from ~$60/1M tokens in early 2023 to under $1/1M by mid-2026 — a 98%+ reduction in 3.5 years. Budget-tier models are now available for $0.05/1M tokens, making LLM inference accessible for high-volume production workloads.

Why do LLM API prices keep falling?

Price reductions are driven by hardware efficiency gains (H100 → H200 → Blackwell), model distillation and quantization techniques, increased competition from open-weight models like Llama and Mistral, and economies of scale as inference infrastructure matures.

How is LLM API pricing calculated?

LLM APIs charge per token — typically split into input (prompt) tokens and output (completion) tokens. Prices are quoted per 1 million tokens in USD. A token is roughly 4 characters or ¾ of a word in English.

Which LLM API is cheapest in 2026?

Budget-tier models from Groq, Together AI, and Fireworks AI typically offer the lowest per-token prices. Use the tier filter above to compare budget options.

What is time-to-first-token (TTFT)?

TTFT is the latency between sending a request and receiving the first token of the response. Lower TTFT means faster perceived responsiveness, which matters for real-time chat applications. Groq and Cerebras typically lead on TTFT due to custom inference hardware.

What is the difference between frontier and budget LLM tiers?

Frontier models (GPT-4o, Claude 3.5 Sonnet, Gemini 2.5 Pro) offer the highest capability but cost $2–$15+ per 1M input tokens. Budget models (Llama 3.1 8B, Mistral 7B) cost $0.05–$0.20 per 1M tokens and are suitable for high-volume, lower-complexity tasks.

Do LLM providers offer prompt caching discounts?

Yes — OpenAI, Anthropic, and Google offer cached input pricing at 50–75% off standard input rates for repeated prompt prefixes. This is shown in the "Cached" column on this page where available.