LLM API pricing — what actually changed, 2024-01 – 2026-06

GPT-4 launched with $0.03 per 1k input tokens ($30/1M) and $0.06 per 1k output tokens ($60/1M) for the 8k context window model. The 32k context variant was priced at $0.06 / $0.12 per 1k. These were the highest list prices OpenAI published for a text model at the time.

GPT-4 Turbo (gpt-4-1106-preview) launched at $0.01 / $0.03 per 1k tokens ($10 input / $30 output per 1M). The existing GPT-4 model kept its old price. GPT-3.5-Turbo was simultaneously cut from $0.0015 / $0.002 per 1k to $0.001 / $0.002 per 1k. Announced at OpenAI DevDay November 6, 2023.

GPT-4o (gpt-4o) launched at $5 per 1M input tokens and $15 per 1M output tokens — half the input price of GPT-4 Turbo at the same quality tier. Announced at OpenAI Spring Update on May 13, 2024.

GPT-4o mini launched as the new low-cost frontier model at $0.15 per 1M input and $0.60 per 1M output — roughly 30× cheaper than GPT-4o on input. OpenAI simultaneously deprecated GPT-3.5-Turbo for new users. GPT-4o mini became the go-to cheapest OpenAI model.

OpenAI released o3-mini as an affordable reasoning model. The o1 series pricing was also adjusted. Reasoning models introduced a new pricing tier above frontier chat models.

Claude 3 introduced three tiers: Haiku (fastest, cheapest) at $0.25 input / $1.25 output per 1M tokens; Sonnet (balanced) at $3 / $15; Opus (most capable) at $15 / $75. Haiku became one of the cheapest quality models in the market at launch.

Claude 3.5 Sonnet launched at the same price point as Claude 3 Sonnet ($3 / $15 per 1M tokens) but with substantially better benchmarks — outperforming Claude 3 Opus on most tasks while costing 5x less. The price didn’t change; the value per dollar did.

An updated Claude 3.5 Sonnet (claude-3-5-sonnet-20241022) was released alongside Claude 3.5 Haiku at $0.80 per 1M input / $4 per 1M output — a ~3x increase over Claude 3 Haiku's $0.25/$1.25 but with substantially higher capability. Claude 3 Haiku remained available at original price.

Claude 3.7 Sonnet introduced a hybrid reasoning mode (extended thinking) at the same $3 / $15 per 1M price point as previous Sonnet models. Extended thinking tokens billed at output token rate.

Anthropic announced the Claude 4 family. NOTE: exact pricing and model lineup at announcement need verification against official Anthropic pricing page and Wayback captures from May 2025.

Google announced Gemini 1.0 Pro for general availability via the Gemini API (formerly PaLM API). Pay-as-you-go pricing introduced. Gemini 1.0 Pro was priced below GPT-4 at launch. NOTE: verify exact prices against Wayback captures. (Reported price — the Web Archive has no capture of this provider's pricing page this early, so there is no dated snapshot to cite; included for the trend, not as archive-verified.)

Gemini 1.5 Pro launched with 1M token context window. Gemini 1.5 Flash was introduced as a lightweight, fast, cheaper model at $0.35 per 1M input / $1.05 per 1M output (prompts up to 128k tokens) — establishing the Flash sub-brand as Google’s cheap-tier offering. 1.5 Pro was $3.50 / $10.50 (up to 128k). Long-context surcharges applied above thresholds. (Reported price — the Web Archive has no capture of this provider's pricing page this early, so there is no dated snapshot to cite; included for the trend, not as archive-verified.)

Gemini 2.0 Flash Experimental was released for free during the experimental/preview period. Paid pricing was not yet published; developers could use it at no cost through the Gemini API free tier. (Reported price — the Web Archive has no capture of this provider's pricing page this early, so there is no dated snapshot to cite; included for the trend, not as archive-verified.)

Gemini 2.5 Pro and Gemini 2.5 Flash launched. Flash continued as the low-cost option in the 2.5 lineup. NOTE: exact prices at launch need verification against official Google pricing page and Wayback captures from March 2025.

DeepSeek-R1 (reasoning model) and DeepSeek-V3 launched as API products. Prices were dramatically lower than Western competitors: DeepSeek-V3 at approximately $0.27 per 1M input / $1.10 per 1M output; DeepSeek-R1 at approximately $0.55 per 1M input / $2.19 per 1M output. Compared to GPT-4o at $5/$15, this represented roughly 10–18x cheaper input pricing. The launch caused significant industry discussion and competitor pricing pressure.

xAI launched the Grok API for external developers. Grok 2 was priced competitively against GPT-4o. NOTE: exact prices at launch need verification against official xAI docs and Wayback captures from November 2024. (Reported price — the Web Archive has no capture of this provider's pricing page this early, so there is no dated snapshot to cite; included for the trend, not as archive-verified.)

Mistral launched its commercial API with Mistral Large (then called Mistral Large 2) at approximately $8 per 1M input / $24 per 1M output. Mistral Medium and Small were also available at lower prices. Exact prices at launch need Wayback verification. (Reported price — the Web Archive has no capture of this provider's pricing page this early, so there is no dated snapshot to cite; included for the trend, not as archive-verified.)

Mistral released the Mixtral series (mixture-of-experts architecture) which had different pricing from the dense Mistral Large models. Mixtral was positioned as a high-performance, more affordable alternative. NOTE: exact prices need Wayback verification. (Reported price — the Web Archive has no capture of this provider's pricing page this early, so there is no dated snapshot to cite; included for the trend, not as archive-verified.)

WizardCost verification (2026-06-20) found the current 'Mistral Large' model is Mistral Large 3 at $0.50 per 1M input / $1.50 per 1M output — a dramatic reduction from Mistral Large 2's $2/$6. Our data was one generation behind; updated in models.json on 2026-06-20.