LLM Prices 2023–2026: How Far Inference Costs Have Fallen

GPT-4 launched in March 2023 at $30 per million input tokens and $60 per million output tokens. By October 2026, OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 (prompts up to 100K tokens) cost $0.10 input and $0.50 output. That is 300 times less per input token and 120 times less per output token. Epoch AI holds capability fixed and measured a faster fall. On a set of PhD-level science questions, the price of GPT-4-level performance dropped 40x per year.

The top tier did not follow. GPT-6 Astra and Claude Fable 5.1 both cost $10 input and $50 output, close to what GPT-4 charged for output in 2023. Some prices have gone up. Two more increases have fixed dates, and a third follows when a promotional price ends.

Every figure in this article is a USD list price per 1M tokens on the standard tier. Each was taken at launch or on the date of a price cut, with no batch, cache or volume discounts. Current figures are as of October 2026.

Output price per million tokens, 2023 to 2026

Each step on the vertical axis is ten times the one below it. Four things stand out.

OpenAI's flagship output price fell from $60 to $8 with GPT-4.1 in April 2025. It then climbed back to $30 with GPT-5.5 and to $50 with GPT-6 Astra. OpenAI now sells the middle slot as a separate model, GPT-6.1 Sol, at $2 / $10.

Anthropic's mid tier went from $32.68 (Claude 2, July 2023) to $15 (Claude 3 Sonnet, March 2024). It then held at $3 / $15 for more than two years, from Claude 3 Sonnet to Sonnet 4.6. Claude Sonnet 5 cut it to $2 / $10 in June 2026.

Google's Flash tier was cheapest in August 2024, at $0.30 output after a 71% cut. Every generation since has cost more than that. Gemini 3.8 Flash costs $3.75 output today and is scheduled to cost $7.50 from 1 January 2027.

The cheapest models fell from $2 (GPT-3.5 Turbo, March 2023) through $1.25 (Claude 3 Haiku, March 2024) to $0.60 (GPT-4o mini, July 2024). Since then the floor has moved between $0.27 and $0.50 output. Open models on third-party hosts go lower. On Groq, gpt-oss-20B costs $0.075 / $0.30.

Together, the lines show list prices falling fast until mid-2024 and then flattening. After that, vendors mostly put new models into existing price slots rather than cutting the slot itself.

The price tables

OpenAI flagship

The 32K-context version of GPT-4 cost twice as much, $60 / $120.

Anthropic mid tier

Anthropic had no separate mid tier in 2023. Claude 2 was its main model, and Claude Instant ($1.63 / $5.51) was the cheap one. The July 2023 PDF carries no date inside. The month comes from its file name and from the Claude 2 launch coverage.

Google Flash tier

Gemini 1.5 Flash prices are for prompts up to 128K tokens. Longer prompts paid double. The Gemini 3.5 Flash launch price is according to CloudZero, a third-party cost blog.

The cheapest models

DeepSeek priced V3 in yuan, and the dollar figures are TechNode's conversions. The promotional rate ended on 8 February 2025. After that the standard rate was about $0.27 / $1.09. DeepSeek-R1 is a reasoning model. It is listed here because VentureBeat compared it at launch with OpenAI o1 at $15 / $60.

What stayed expensive

Every year from 2023 to 2026, the top of OpenAI's or Anthropic's list cost $50 or more per million output tokens. GPT-5.5 Pro costs far more. Google's top models stayed lower. Gemini 4 Argon is $10 output, rising to $20 on 1 December 2026.

Sources: GPT-4 from the OpenAI forum thread above, Claude 3 Opus from Simon Willison, Claude Opus 4 from Neowin, GPT-5.5 Pro from the OpenAI help centre, the 2026 flagships from the vendor pricing pages in Sources.

Individual model lines do get cheaper. Anthropic's Opus went from $15 / $75 (Opus 4, May 2025) to $5 / $25 (Opus 4.5) and $4 / $20 (Opus 5.5, September 2026). Above Opus, the company added Claude Fable 5 (June 2026) and Fable 5.1 (September 2026) at $10 / $50. The top slot itself has not become cheap.

Prices also rise. Five changes from 2024 to 2027 go upward:

TechCrunch called the Haiku change "a 4x hike". The two Gemini increases double the price of models that are on sale today. Introductory prices are a pattern to plan for. A model chosen during its launch window may cost twice as much a quarter later.

How fast the decline is

Two studies put a rate on it.

Epoch AI published its analysis on 12 March 2025 (Cottier, Snodin, Owen, Adamczewski). It found that "the price to achieve GPT-4's performance on a set of PhD-level science questions fell by 40x per year". Across the milestones they tracked, the rates ranged "from 9x to 900x per year". They add a caveat: "The fastest price drops in that range have occurred in the past year, so it's less clear that those will persist."

Guido Appenzeller of a16z named the trend "LLMflation" on 12 November 2024. He wrote: "For an LLM of equivalent performance, the cost is decreasing by 10x every year." His example, "What cost $60/million tokens in 2021 costs $.06/million tokens today", is a fall of about 1,000x in three years. Sam Altman gave a similar figure in February 2025. His post Three Observations says: "The cost to use a given level of AI falls about 10x every 12 months."

Both studies hold capability fixed and look for the cheapest model that reaches it. A vendor's price list does something else. It shows what the model in a given slot costs. The tables above give these multiples:

The per-year figure for Luna is 120 raised to the power 1/3.6. The 40x-per-year rate and the 6x total do not contradict each other. One measures a fixed capability. The other tracks a price slot that kept getting a better model. A team that keeps the same model slot sees the second number. A team that moves each job to the cheapest model that passes its tests gets closer to the first.

Input fell faster than output

The ratio between output and input price has widened on OpenAI's flagship line:

Today's mid and small tiers sit at 5 (Sonnet 5.5 and GPT-6.1 Sol at $2 / $10, Luna and Haiku 5.5 at $0.10 / $0.50). A workload that reads long documents and writes short answers gained more from the last three years than one that writes long answers. Reasoning-heavy and generation-heavy jobs should be priced on the output column first.

What it means for budgets

Plan for model swaps. The large savings in this history came from moving a job down a tier. Work that needed GPT-4 at $60 output in 2023 could be tested in 2026 on mid-tier models at $10 or small ones at $0.50. Slot prices moved slowly, so the saving arrives only when someone re-runs the evaluation and makes the switch. The model routing guide covers how to send each request to the cheapest model that handles it.

Take one job at three points in the tables. It uses 2,000 input tokens and 500 output tokens per call, one million calls a month, at list prices.

The same volume on GPT-6 Astra ($10 / $50) would cost $45,000 a month. A job that stayed on the flagship slot since 2023 saw its bill halve. A job that moved down to the mid tier saw it fall tenfold. One that a small model can handle costs 1/200 of the 2023 figure.

Keep the architecture vendor-neutral. Google's Flash tier went from $0.30 to $3.75 output between August 2024 and October 2026, a 12.5-fold rise. OpenAI's flagship rose from $8 to $50. A team locked to one provider's SDK, prompt format and tool schema pays whatever its next generation costs. A thin client layer and an evaluation set per feature make a switch a configuration change.

Re-price the bill every quarter. Between 30 July and 13 October 2026, OpenAI cut GPT-5.6 Luna by 80% and then cut GPT-5.6 Sol. Anthropic and OpenAI released mid-tier models at $2 / $10 (Claude Sonnet 5.5, GPT-6.1 Sol). Google launched Gemini 4 Argon at an introductory price that doubles on 1 December. The LLM API pricing comparison lists the October 2026 figures with a source per row.

Expect usage per product to rise faster than the price per call falls. Agents make many calls per task and re-send a growing context on each one. Context windows of about 1M tokens are now standard on current Claude and GPT-6 models and on Gemini 4 Argon. A cheaper token does not guarantee a smaller invoice. The hidden token tax of agents shows where those extra tokens come from.

The study and the price cut, first-hand

What to do next

List every feature in production with its model, tier and launch price, and mark any that sit on a flagship or an introductory price.

For each flagship feature, run your evaluation set against the current mid-tier and small models of two vendors.

Put 1 December 2026 and 1 January 2027 in the calendar if you use Gemini 4 Argon or Gemini 3.6 or 3.8 Flash.

Divero offers a free audit, and the first conversation is free. Write to [email protected] if you want this re-pricing checked against your bill.

Sources

• Epoch AI: LLM inference prices have fallen rapidly but unequally across tasks
• a16z: Welcome to LLMflation
• Claude API pricing
• Gemini API pricing
• OpenAI GPT-6 Astra model page
• OpenAI GPT-5 nano model page
• Anthropic model pricing, July 2023 (PDF)
• Google Developers Blog: Gemini 1.5 Flash price cut
• Simon Willison: o3 price drop
• Simon Willison: Claude 3 launch prices