LLM Price History Tracker — API $/token over time

Every price on this page is a published, dated list figure taken from the provider's own pricing page or its launch/price-change announcement — never an estimate. Prices are stored in USD per 1,000,000 tokens exactly as published (where a provider quotes $/1K, it is normalised by ×1000, with no rounding of the source figure). The dataset is a curated reference: one representative row per model family per price change, text generation only. Batch, cached-input, fine-tuning and region-specific SKUs are excluded, and that limit is stated so coverage isn't overstated.

The cost time-machine is deterministic arithmetic. For a price point, the monthly cost of your workload is:

cost = inputTokensₘ × inputPer1M + outputTokensₘ × outputPer1M

Your token volumes are entered in millions and prices are per million, so the multiply is direct. This is cross-checked internally against the per-token identity — raw tokens × ($/1M ÷ 1,000,000) — which must return the same figure to the cent, so the engine is self-verifying.

To build the four dated snapshots, the tool picks — for your chosen tier — the representative model whose effective date is the latest on or before each boundary (end of 2023, 2024, 2025, and today), then costs your workload at each. Savings against a baseline year b are (cost_b − cost_today) ÷ cost_b, measured from the earliest priced snapshot; if both token inputs are zero every cost is zero and savings show as “—”.

The Δ vs previous column compares each row to the most recent earlier price in the same model family: Δ = (new − old) ÷ old, computed separately for input and output. A negative Δ (green) is a price cut; a positive Δ (red) is a rise. The blended metric weights input and output at 3:1 — a typical chat/RAG ratio — and is used only for the chart and table display, never for your actual time-machine cost. The optional LKR display multiplies USD by a fixed, cited snapshot of Rs 305/USD; it is not a live rate. Because list prices change by provider announcement at any time, the whole page is stamped with a LAST_VERIFIED date of 2026-07-17 and every row links to its source.

Each price row links to its own source. The primary pricing pages and dated announcement archives used to pin figures are:

• OpenAI — API pricing
• Anthropic — API pricing
• Google — Gemini API pricing
• xAI — Grok models & pricing
• Together AI — hosted open-weights pricing (Llama, etc.)
• DeepSeek — official API pricing
• Central Bank of Sri Lanka — indicative exchange rates (for LKR display)

Prices were last cross-checked against these sources on 2026-07-17. As of that date the fastest-falling family tracked here is Google Gemini Flash, down 64.29% on combined input+output since its first listed price. API prices change by provider announcement at any time — always confirm against the linked source before committing to a budget.