---
title: "AI Inference Cost Statistics 2026: Price per Token Trends"
date: 2026-11-10
author: "Aaron Weiss"
featured_image: "https://tricitylocalnews.com/wp-content/uploads/2026/10/ai-inference-cost-statistics-price-per-token-trends-v2.jpg"
categories:
  - name: "AI"
    url: "/ai.md"
tags:
  - name: "Statistics"
    url: "/tag/statistics.md"
---

# AI Inference Cost Statistics 2026: Price per Token Trends

OpenAI’s API Standard pricing puts the AI inference cost of GPT-6 Astra at **$10** per million input tokens and **$50** per million output tokens. Andreessen Horowitz (a16z) estimated in November 2024, in a study of large language model (LLM) prices, that for an LLM of equivalent performance, the cost is decreasing by **10x** every year. Gartner forecasts that, within AI-optimized infrastructure as a service (IaaS) spending, in 2026, global spending on inference (**$23.3 billion**) will surpass that of training (**$19 billion**).

Anthropic says Claude 4.7 and later models use a newer tokenizer that produces approximately **30%** more tokens for the same text, so a lower price per token does not always mean a lower bill per task. The figures below cover vendor list prices posted in October 2026, decline rates from Epoch AI, a16z and Stanford’s AI Index, and forecasts from Gartner and Deloitte. Cost-per-task and workload savings figures come from the vendors’ own tests and are labeled that way.

## Key Takeaways

- Epoch AI found that, across six benchmarks, the rate of decline varies dramatically depending on the performance milestone, ranging from **9x** to **900x** per year, in an analysis published in March 2025.
- Stanford’s 2025 AI Index reports that the inference cost for a system performing at the level of GPT-3.5 dropped over **280-fold** between November 2022 and October 2024.
- Anthropic lists Claude Opus 4 output at **$75/MTok** and Claude Opus 5.5 output at **$20/MTok**, where MTok means one million tokens.
- OpenAI reports that, in its own evaluation, at maximum effort, GPT-6.1 Sol costs **$5.47** per task on average, compared with **$23.21** for Opus 5.5 and **$23.80** for Astra.
- In Gartner’s forecast for AI-optimized infrastructure as a service (IaaS), the share of AI-optimized IaaS spending that supports inference is set to reach **59%** in 2027.
- Apollo’s chief economist reports that the price of a single token has fallen more than **90%** since 2023.

## Editor’s Choice

- GPT-6 Astra, OpenAI Standard tier, short context: **$10.00** input and **$50.00** output per million tokens.
- GPT-6 Luna, OpenAI Standard tier, short context: **$0.10** input and **$0.50** output per million tokens.
- Claude Haiku 5.5, prompts up to 100,000 tokens: **$0.10/MTok** for prompts up to 100,000 tokens, with output at **$0.50/MTok**.
- Gemini 3.8 Flash on Google Cloud, introductory price: **$0.75**/**$3.75** per million tokens input/output through December 31, 2026.
- GPT-6 Astra input, Ultrafast mode, short context: **$60.00** per million tokens.
- NVIDIA Data Center revenue, second quarter of fiscal 2027: **$89.0 billion**, up **117%** from a year ago.

## AI Inference Cost per Million Tokens by Model

- OpenAI’s mid-tier model, GPT-6.1 Sol, lists at **$2.00** input and **$10.00** output per million tokens on the Standard tier for short context.
- Long-context requests to GPT-6 Astra cost **$20.00** input and **$75.00** output per million tokens, double the short-context input rate.
- Anthropic lists Claude Fable 5.1 at **$10/MTok** input and **$50/MTok** output, the same headline rates as GPT-6 Astra.
- Claude Sonnet 5.5 lists at **$2/MTok** input and **$10/MTok** output, level with GPT-6.1 Sol.
- Claude Haiku 5.5 rises to **$0.50/MTok** for prompts over 100,000 tokens, with output at **$2.50/MTok**.
- Output tokens list at **5x** the input price for GPT-6 Astra, and the same 5-to-1 ratio holds on every Standard short-context row OpenAI shows.
- Google Cloud’s introductory Gemini 3.8 Flash rate of **$0.75**/**$3.75** per million tokens sits between the OpenAI and Anthropic small and mid-tier models.

| Provider | Model | Input ($ per million tokens) | Output ($ per million tokens) |
|---|---|---|---|
| OpenAI | GPT-6 Astra | 10.00 | 50.00 |
| Anthropic | Claude Fable 5.1 | 10 | 50 |
| Anthropic | Claude Opus 5.5 | 4 | 20 |
| OpenAI | GPT-6.1 Sol | 2.00 | 10.00 |
| Anthropic | Claude Sonnet 5.5 | 2 | 10 |
| Google Cloud | Gemini 3.8 Flash (introductory, through December 31, 2026) | 0.75 | 3.75 |
| OpenAI | GPT-6 Luna | 0.10 | 0.50 |
| Anthropic | Claude Haiku 5.5 (prompts up to 100,000 tokens) | 0.10 | 0.50 |

*Source: OpenAI API pricing, Anthropic Claude pricing and Google Cloud Vertex AI pricing pages, as of October 2026*

### About This Data

- **17** sources: **8** vendor pricing pages and model announcements from OpenAI, Anthropic and Google Cloud, **1** NVIDIA earnings release filed with the SEC, **5** research publications from Epoch AI, a16z and Stanford’s AI Index, **2** analyst forecasts from Gartner and Deloitte, and **1** economist note from Apollo.
- Source dates run from **2023** OpenAI announcements to pricing pages read in **October 2026**.
- Selection: vendor, company filing, research and named-forecaster publishers only; market-research sellers excluded.
- Figures are reviewed on a rolling basis and updated when sources publish new editions.

### How much does AI inference cost per million tokens?

Short-context list prices on OpenAI’s Standard tier run from **$0.10** input for GPT-6 Luna to **$10.00** input for GPT-6 Astra per million tokens. Anthropic lists input prices for Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5 from **$0.10** per million tokens for Claude Haiku 5.5 prompts up to 100,000 tokens to **$10** for Claude Fable 5.1. Output tokens cost more than input on every row.

## How Fast AI Inference Costs Are Falling

- On the Massive Multitask Language Understanding (MMLU) benchmark, a16z found that when GPT-3 became publicly accessible in November 2021, it was the only model that was able to achieve an MMLU of 42, at a cost of **$60** per million tokens.
- By the time of a16z’s writing, the cheapest model to achieve the same score was a Llama 3.2 model from model-as-a-service provider Together.ai, at **$0.06** per million tokens.
- a16z summarized that the cost of LLM inference has dropped by a factor of **1,000** in **3** years.
- For the higher MMLU score of 83, the price for models at this level has come down by about a factor of **62** since GPT-4 came out in March 2023.
- Epoch AI found the price to achieve GPT-4’s performance on a set of PhD-level science questions fell by **40x** per year.
- Epoch AI cautions that the fastest price drops in that range have occurred in the past year, so it’s less clear that those will persist.
- Stanford’s 2025 AI Index reports that the inference cost for a system performing at the level of GPT-3.5 dropped over **280-fold** between November 2022 and October 2024, driven by increasingly capable small models.
- At the hardware level, Stanford’s 2025 AI Index reports costs have declined by **30%** annually, while energy efficiency, which sets how much [data center electricity](/data-center-energy-consumption-statistics/) each token uses, has improved by **40%** each year.

 PRICE DECLINE · Source: Epoch AI, March 2025; Andreessen Horowitz, November 2024   PRICE DECLINE    Epoch AI · 2025 LLM Price Decline Rates by Study (x per Year)         Epoch AI, fastest milestone  900 Epoch AI, GPT-4 level on PhD-level science questions  40 a16z, LLM of equivalent performance  10 Epoch AI, slowest milestone  9  0 200 400 600 800 1K   SOURCE Epoch AI, March 2025; Andreessen Horowitz, November 2024    View the data| Study and milestone | Price decline (x per year) |
|---|---|
| Epoch AI, fastest milestone | 900 |
| Epoch AI, GPT-4 level on PhD-level science questions | 40 |
| a16z, LLM of equivalent performance | 10 |
| Epoch AI, slowest milestone | 9 |





> **Key finding:** Epoch AI’s analysis found that the rate of decline varies dramatically depending on the performance milestone, ranging from **9x** to **900x** per year. The spread is wide because each benchmark threshold is met by a different set of cheap models, and Epoch’s own caveat is that the fastest rates come from the most recent year.

### How fast is the cost of AI inference falling?

LLM inference cost for a fixed capability falls fast on a16z’s measure: For an LLM of equivalent performance, the cost is decreasing by **10x** every year. a16z took the average of input and output prices where they differed, so its rate is a blended figure.

## Recent Developments

- **September 29, 2026:** OpenAI updated its GPT-6 Astra page to point to OpenAI’s latest model, GPT-6.1 Sol, whose standard API prices are **$2** per million input tokens, **$0.10** per million cached input tokens, and **$10** per million output tokens.
- **September 22, 2026:** OpenAI said it was “expanding our GPT-6 family with GPT-6 Sol and GPT-6 Luna”, the release that put GPT-6 Luna on OpenAI’s price card.
- **August 26, 2026:** NVIDIA reported revenue for the second quarter ended July 26, 2026, of **$96.2 billion**, up **18%** from the previous quarter and up **106%** from a year ago.
- **August 10, 2026:** Gartner projected that worldwide AI-optimized infrastructure as a service (IaaS) spending is projected to grow **96%** through 2026, reaching **$42 billion**.
- **July 30, 2026:** OpenAI’s pricing page notes that Priority processing was renamed Fast mode on July 30, 2026.

## Cost per Token vs Cost per Task

- Anthropic says Claude 4.7 and later models use a newer tokenizer that produces approximately **30%** more tokens for the same text, and that the exact increase depends on the content and workload shape.
- Anthropic states that for Claude Opus 5.5, input and output tokens are **$4** and **$20** per million, **20%** less than Opus 5.
- Anthropic also says “our tests show that at default settings it will cost **40%** less than Opus 5 on typical workloads”, twice the cut in the per-token price.
- OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices.
- On OpenAI’s per-task measure, GPT-6 Astra costs **4.4x** as much per task as GPT-6.1 Sol, slightly less than the **5x** gap in their per-token list prices.
- OpenAI describes GPT-6.1 Sol as delivering substantial scientific capability at over **75%** lower cost than either model.

 AVERAGE COST PER TASK · Source: OpenAI, Introducing GPT-6.1 Sol, September 2026 (vendor-run evaluation)   AVERAGE COST PER TASK    OpenAI · 2026 Cost per Task at Maximum Effort ($)         GPT-6 Astra  $23.80 Claude Opus 5.5  $23.21 GPT-6.1 Sol  $5.47  0 5 10 15 20 25   SOURCE OpenAI, Introducing GPT-6.1 Sol, September 2026 (vendor-run evaluation)    View the data| Model | Average cost per task ($) |
|---|---|
| GPT-6 Astra | 23.80 |
| Claude Opus 5.5 | 23.21 |
| GPT-6.1 Sol | 5.47 |





> **Worth noting:** OpenAI’s figures of **$5.47** per task for GPT-6.1 Sol and **$23.21** for Opus 5.5 come from OpenAI’s own evaluation of scientific work at maximum effort. A rival’s model was run by the vendor selling the cheaper one, so the gap is a vendor claim until an independent test repeats it.

### What is the AI inference cost per token versus per task?

Per-token price is the list rate; per-task cost multiplies it by how many tokens a model uses. Anthropic says Opus 5.5 input and output tokens are **$4** and **$20** per million, **20%** less than Opus 5, yet its tests put typical workload cost **40%** lower. Token count per task moves the bill as much as the rate does.

## Batch, Flex, Fast and Ultrafast Pricing

- OpenAI’s [API pricing page](https://developers.openai.com/api/docs/pricing) shows the Batch tier listing GPT-6 Astra at **$5.00** input and **$25.00** output per million tokens for short context, half the Standard rate.
- OpenAI says Fast mode delivers up to **2x** the speed of Standard processing at **2x** the Standard price.
- Ultrafast, the top mode on the page, lists GPT-6 Astra at **$60.00** input and **$300.00** output per million tokens.
- From Batch to Ultrafast, GPT-6 Astra input carries a **12x** price spread for the same model.
- The lowest-priced model is cheaper still on Batch, with GPT-6 Luna at **$0.05** input and **$0.25** output per million tokens.
- Regional processing (data residency) endpoints are charged a **10%** uplift for models released on or after March 5, 2026, and Federal Risk and Authorization Management Program (FedRAMP) endpoints carry the same uplift.
- Anthropic says Fast mode for Opus 5.5 offers up to **2.5x** speed and costs **$8** per million input tokens and **$40** per million output tokens.

 Input vs Output · Source: OpenAI API pricing page, as of October 2026      OpenAI API · 2026 GPT-6 Astra Price by Mode ($ per Million Tokens) Input vs Output    Input, short context  Output, short context          300 240 180 120 60 0   5  25  5  25  10  50  20  100  60  300   Batch Flex Standard Fast Ultrafast  Input, short context · 5 Output, short context · 25 Input, short context · 5 Output, short context · 25 Input, short context · 10 Output, short context · 50 Input, short context · 20 Output, short context · 100 Input, short context · 60 Output, short context · 300   SOURCE OpenAI API pricing page, as of October 2026    View the data| Processing mode | Input, short context ($ per million tokens) | Output, short context ($ per million tokens) |
|---|---|---|
| Batch | 5.00 | 25.00 |
| Flex | 5.00 | 25.00 |
| Standard | 10.00 | 50.00 |
| Fast | 20.00 | 100.00 |
| Ultrafast | 60.00 | 300.00 |







## Cached Input Token Pricing

- OpenAI says GPT-6.1 Sol cached input costs just **$0.10** per million tokens, **95%** less than standard input pricing and **50%** less than GPT-6 Sol’s cached input pricing.
- On the Standard tier, GPT-6 Astra cached input is **$1.00** against **$10.00** input per million tokens for short context.
- GPT-6 Luna cached input lists at **$0.01** per million tokens, the lowest cached rate on OpenAI’s Standard short-context table.
- Batch pricing halves the cached rate again, with GPT-6 Astra cached input at **$0.50** per million tokens.
- Anthropic lists Opus 5.5 cache reads at **$0.20** per million tokens, **60%** less than Opus 5.
- Anthropic’s launch table shows Opus 5 cache reads at **$0.50** and cache writes at **$6.25** per million tokens.

 Standard input vs Cached input or cache read · Source: OpenAI API pricing page and Anthropic, Introducing Claude Opus 5.5, as of October 2026      OpenAI API · 2026 Standard vs Cached Input Price ($ per Million Tokens) Standard input vs Cached input or cache read    Standard input  Cached input or cache read          10 8 6 4 2 0   10  1  5  0.5  4  0.2  2  0.1  0.1  0.01   GPT-6 Astra Claude Opus 5 Claude Opus 5.5 GPT-6.1 Sol GPT-6 Luna  Standard input · 10 Cached input or cache read · 1 Standard input · 5 Cached input or cache read · 0.5 Standard input · 4 Cached input or cache read · 0.2 Standard input · 2 Cached input or cache read · 0.1 Standard input · 0.1 Cached input or cache read · 0.01   SOURCE OpenAI API pricing page and Anthropic, Introducing Claude Opus 5.5, as of October 2026    View the data| Model | Standard input ($ per million tokens) | Cached input or cache read ($ per million tokens) |
|---|---|---|
| GPT-6 Astra | 10.00 | 1.00 |
| Claude Opus 5 | 5 | 0.50 |
| Claude Opus 5.5 | 4 | 0.20 |
| GPT-6.1 Sol | 2.00 | 0.10 |
| GPT-6 Luna | 0.10 | 0.01 |







> **Why it matters:** Anthropic says cache reads (which make up the majority of agentic and coding work costs) are **$0.20** per million tokens, **60%** less than Opus 5. For agent workloads that resend the same context, the cache-read rate can matter more to the monthly bill than the headline input price does.

## LLM Price per Million Tokens Over Time

- OpenAI said GPT-4 Turbo is offered at a **3x** lower price for input tokens and a **2x** lower price for output tokens compared to GPT-4.
- OpenAI priced GPT-4o mini at **15 cents** per million input tokens and **60** cents per million output tokens, which it called more than **60%** cheaper than GPT-3.5 Turbo.
- OpenAI reported that mini scores **82%** on MMLU at that price.
- OpenAI called GPT-4o mini an order of magnitude more affordable than previous frontier models.
- Each row is stated in the source’s own terms, so some are absolute prices, and some are ratios to an older model.

| Model | Price as the source states it | Source |
|---|---|---|
| GPT-3, public access in November 2021 | $60 per million tokens | a16z |
| OpenAI o1 | $60 per million output tokens | a16z |
| GPT-4 Turbo | 3x cheaper input and 2x cheaper output than GPT-4 | OpenAI |
| GPT-4o mini | 15 cents input and 60 cents output per million tokens | OpenAI |
| Llama 3.2 model on Together.ai | $0.06 per million tokens | a16z |
| Claude Opus 4 | $15 input and $75 output per million tokens | Anthropic |
| Claude Haiku 3.5 | $0.80 input and $4 output per million tokens | Anthropic |
| GPT-6 Luna | $0.10 input and $0.50 output per million tokens | OpenAI |

*Source: Andreessen Horowitz, November 2024; OpenAI announcements and API pricing page; Anthropic Claude pricing page, as of October 2026*

## Claude API Price by Model Generation

- Anthropic’s [Claude pricing documentation](https://platform.claude.com/docs/en/about-claude/pricing) still lists Claude Opus 4.1 at **$15/MTok** input, the same input rate as Opus 4.
- Opus 5 input stood at **$5** per million tokens before Opus 5.5 cut it to **$4**.
- From Opus 4’s **$75** to Opus 5.5’s **$20** per million output tokens, the output price fell **73%**.

 OUTPUT PRICE · Source: Anthropic Claude pricing page and Introducing Claude Opus 5.5, as of October 2026   OUTPUT PRICE    Anthropic Claude · 2026 Claude Opus Output Price ($ per Million Tokens)         Claude Opus 4  75 Claude Opus 4.1  75 Claude Opus 5  25 Claude Opus 5.5  20  0 16 32 48 64 80   SOURCE Anthropic Claude pricing page and Introducing Claude Opus 5.5, as of October 2026    View the data| Model | Output price ($ per million tokens) |
|---|---|
| Claude Opus 4 | 75 |
| Claude Opus 4.1 | 75 |
| Claude Opus 5 | 25 |
| Claude Opus 5.5 | 20 |





The Opus line fell by nearly three-quarters, and the mid-tier moved down more gently:

- Claude Sonnet 4, Sonnet 4.5 and Sonnet 4.6 all list at **$3/MTok** input and **$15/MTok** output.
- Claude Sonnet 5.5 lists at **$2/MTok** input and **$10/MTok** output.
- The small-model line did not fall in a straight line: Claude Haiku 4.5 lists at **$1/MTok** input and **$5/MTok** output, against **$0.80** and **$4** for Claude Haiku 3.5.

 OUTPUT PRICE · Source: Anthropic Claude pricing page, as of October 2026   OUTPUT PRICE    Anthropic Claude · 2026 Claude Haiku Output Price ($ per Million Tokens)         Claude Haiku 3.5  4 Claude Haiku 4.5  5 Claude Haiku 5.5 (prompts up to 100,000 tokens)  0.50  0 1 2 3 4 5   SOURCE Anthropic Claude pricing page, as of October 2026    View the data| Model | Output price ($ per million tokens) |
|---|---|
| Claude Haiku 3.5 | 4 |
| Claude Haiku 4.5 | 5 |
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | 0.50 |





## AI Inference vs Training Spending

- Gartner’s [August 2026 IaaS forecast](https://www.gartner.com/en/newsroom/press-releases/2026-08-10-gartner-forecasts-worldwide-artificial-intelligence-optimized-iaas-spending-to-grow-96-percent-in-2026) states that in 2026, global spending on inference (**$23.3 billion**) will surpass that of training (**$19 billion**).
- Gartner adds that the growing share of inference workloads is expected to reshape cloud investment priorities.
- Deloitte predicts that inference, the process of running AI models, will account for roughly two-thirds of AI compute in 2026.
- Deloitte also predicts the market for inference-optimized chips will grow to over **$50 billion** in 2026.
- In Deloitte’s view, combined training/inference AI chips in data centers will continue to dominate the **$200 billion** AI chip market.

| Forecaster | Measure | Figure | Year |
|---|---|---|---|
| Gartner | Global spending on inference | $23.3 billion | 2026 |
| Gartner | Global spending on training | $19 billion | 2026 |
| Gartner | Share of AI-optimized IaaS spending supporting inference | 55% | 2026 |
| Gartner | Share of AI-optimized IaaS spending supporting inference | 59% | 2027 |
| Deloitte | Share of AI compute used for inference | roughly two-thirds | 2026 |
| Deloitte | Inference-optimized chip market | over $50 billion | 2026 |
| Deloitte | AI chip market | $200 billion | 2026 |

*Source: Gartner, August 2026; Deloitte 2026 TMT Predictions*

### Is inference more expensive than training?

Yes, on Gartner’s forecast: In 2026, global spending on inference (**$23.3 billion**) will surpass that of training (**$19 billion**). Deloitte expects inference will account for roughly two-thirds of AI compute in 2026. The two firms measure different things, cloud spending and compute share, so their figures are not one combined estimate.

## AI Infrastructure Spending Behind Inference

- Gartner’s table puts Total AI-optimized IaaS spending at **$42,276 million** (**$42.3 billion**) in 2026, the headline figure in the chart below.
- That series grows **180.0%** in 2025, **96.4%** in 2026 and **56.5%** in 2027.
- AI-optimized IaaS makes up **14.7%** of Gartner’s Total IaaS forecast for 2026, the wider [cloud computing market](/cloud-computing-statistics/) segment that reaches **$287,347 million** (**$287.3 billion**), up **29.3%**.
- NVIDIA reported Data Center revenue of **$89.0 billion** for the quarter ended July 26, 2026, up **117%** from a year ago, as demand for [data center hardware](/data-center-statistics/) kept rising.
- Stanford’s 2026 AI Index reports global AI compute capacity grew **3.3x** per year since 2022, reaching **17.1 million** H100-equivalents, with NVIDIA accounting for over **60%** of total compute.

 AI-OPTIMIZED IAAS SPENDING · Source: Gartner, Spending on Infrastructure as a Service, Worldwide, August 2026   AI-OPTIMIZED IAAS SPENDING    Gartner · 2026 Worldwide AI-Optimized IaaS Spending ($ Millions)          80K 60K 40K 20K 0   $21,529M 2025  $42,276M 2026  $66,143M 2027    SOURCE Gartner, Spending on Infrastructure as a Service, Worldwide, August 2026    View the data| Year | AI-optimized IaaS spending ($ millions) |
|---|---|
| 2025 | 21,529 |
| 2026 | 42,276 |
| 2027 | 66,143 |





Jensen Huang, NVIDIA’s founder and CEO, tied that hardware spending to token output: “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.” Stanford’s 2026 AI Index says the build-out is driven by [hyperscaler data center expansion](/ai-capex-statistics/) and sustained demand for frontier model training and inference, and puts NVIDIA ahead on [GPU and accelerator share](/gpu-market-share-statistics/) with over **60%** of total compute.

## Where AI Token Prices Are Rising

- Google Cloud says that for Gemini 3.8 Flash, Gemini 3.7 Flash, Gemini 3.6 Flash, and CodeMender, using these models, standard pricing of **$1.5**/**$7.5** per million tokens input/output will apply starting January 1, 2027.
- Against the introductory rate, that is a **2x** increase on both input and output.
- In a November 2024 analysis, a16z wrote that “OpenAI’s leading model today, o1, has the same cost per output token as GPT-3 had at launch (**$60** per million)”, three years after GPT-3 opened to the public.
- a16z read that as providers willing to concede the low end of the market and instead focus their efforts on the highest-quality tier.

| Model | Earlier price | Later price | Source |
|---|---|---|---|
| Gemini 3.8 Flash, input/output per million tokens | $0.75 / $3.75 through December 31, 2026 | $1.5 / $7.5 from January 1, 2027 | Google Cloud |
| OpenAI output, per million tokens | GPT-3 at launch: $60 | o1: $60 | a16z |
| Claude Haiku input, per million tokens | Haiku 3.5: $0.80 | Haiku 4.5: $1 | Anthropic |
| Claude Haiku output, per million tokens | Haiku 3.5: $4 | Haiku 4.5: $5 | Anthropic |

*Source: Google Cloud Vertex AI pricing page and Anthropic Claude pricing page, as of October 2026; Andreessen Horowitz, November 2024*

### Are AI API prices going up?

Some are. Google Cloud’s pricing page says that for Gemini 3.8 Flash, starting January 1, 2027, standard pricing of **$1.5**/**$7.5** per million tokens input/output will apply, up from **$0.75**/**$3.75**. Falling prices for a fixed capability sit beside flat or rising list prices for some named models.

## Cheaper Tokens, Bigger AI Bills

- Apollo reports that the Silicon Data Token Expenditure Index, which tracks total spending on large language model usage, has roughly doubled since late 2025.
- Over a longer window, Apollo notes, the price of a single token has fallen more than **90%** since 2023.
- Stanford’s 2026 AI Index charts annual compute spending by OpenAI and Anthropic, rising to **$16.3 billion** and **$6.8 billion**, respectively, in 2025.
- The AI Index splits that spending into R&D, inference, and unattributed costs.
- The same chapter notes Google reporting more than **$150 billion** in annual capex in 2025.

| Measure | Figure | Period | Source |
|---|---|---|---|
| Price of a single token | fell more than 90% | since 2023 | Apollo |
| Silicon Data Token Expenditure Index | roughly doubled | since late 2025 | Apollo |
| OpenAI annual compute spending | $16.3 billion | 2025 | Stanford AI Index 2026 |
| Anthropic annual compute spending | $6.8 billion | 2025 | Stanford AI Index 2026 |
| Google annual capex | more than $150 billion | 2025 | Stanford AI Index 2026 |
| Inference spending on AI-optimized IaaS, forecast | $23.3 billion | 2026 | Gartner |

*Source: Apollo, The Daily Spark, June 2026; Stanford HAI, 2026 AI Index Report; Gartner, August 2026*

> Apollo’s chief economist Torsten Slok calls it Jevons paradox: “As tokens get cheaper, companies don’t spend less but instead run more AI agents, automate more workflows and generate more code, pushing aggregate expenditure higher even as the unit cost of intelligence collapses.”

## Conclusion

The AI inference cost of a top model still reads **$10** per million input tokens and **$50** per million output tokens for GPT-6 Astra on OpenAI’s Standard tier, while a16z estimated that for an LLM of equivalent performance, the cost is decreasing by **10x** every year. Both hold at once: a fixed capability keeps getting cheaper, while the top row of each vendor’s price card stays far above its cheapest one.

The figures to watch next are the scheduled Gemini Flash price change, the next round of vendor price cards, and whether Gartner’s inference spending forecast holds as agent workloads grow.