Since mid-July 2026, the models most teams run every day got much cheaper. The most capable ones did not. On a blended basis, OpenAI's main model line fell 64%, Google's Flash tier fell 56% on its introductory rate, and Claude Sonnet's standard price fell 33%. Budget models fell furthest: GPT-6 Luna costs 91% less than GPT-5.6 Luna did at launch. Claude Fable 5.1 and GPT-6 Astra still list at $10 input and $50 output per million tokens. If you pay a large AI bill, re-test your mid-tier model this month before you think about switching providers.
- OpenAI's GPT-6 Sol ($2 / $10) costs 64% less on a blended basis than GPT-5.5 did on July 14 ($5 / $30).
- Claude Sonnet 5's standard price went from a planned $3 / $15 to a permanent $2 / $10, a 33% blended drop.
- Gemini 3.8 Flash's introductory $0.75 / $3.75 runs through December 31, 2026. On the $1.50 / $7.50 standard rate, the Flash tier is only 11% cheaper than in July.
- Claude Opus 5.5, released September 22, is $4 / $20, 20% below Opus 5 and Opus 4.8.
- Frontier list prices held: Claude Fable 5.1 and GPT-6 Astra are both $10 / $50. Fable 5.1 did cut cache reads by 75%.
- All prices are standard API rates in USD per million tokens, input / output.
This index tracks what the major AI APIs cost and how those prices have changed since July. It is built from the data behind our AI cost calculator, which we have kept in sync with official provider pages since July 14, 2026. Every time a price changed, the old value stayed in our version history, so I can put a July price next to a September price and show the change.
The headline numbers
One number per model, so you can compare them fairly.
Comparing models on input price alone is misleading, because output tokens usually cost five to seven times more. So the index uses a blended price that assumes three input tokens for every output token, which is a common shape for chat and agent work. The formula is (3 × input + output) ÷ 4. If your workload writes more than it reads, your savings from the output cuts below will be bigger.
Blended price per million tokens, July 14 vs September 25, 2026
| Tier | July 14 model | July 14 | Sept 25 model | Sept 25 | Change |
|---|---|---|---|---|---|
| OpenAI main model | GPT-5.5 ($5 / $30) | $11.25 | GPT-6 Sol ($2 / $10) | $4.00 | -64% |
| Google Flash (intro rate) | Gemini 3.5 Flash ($1.50 / $9) | $3.38 | Gemini 3.8 Flash ($0.75 / $3.75) | $1.50 | -56% |
| Google Flash (standard rate) | Gemini 3.5 Flash ($1.50 / $9) | $3.38 | Gemini 3.8 Flash ($1.50 / $7.50) | $3.00 | -11% |
| Anthropic mid tier (standard rate) | Claude Sonnet 5 ($3 / $15) | $6.00 | Claude Sonnet 5 ($2 / $10) | $4.00 | -33% |
| Anthropic premium | Claude Opus 4.8 ($5 / $25) | $10.00 | Claude Opus 5.5 ($4 / $20) | $8.00 | -20% |
| Anthropic frontier | Claude Fable 5 ($10 / $50) | $20.00 | Claude Fable 5.1 ($10 / $50) | $20.00 | 0% |
Source: Spectrum AI Labs price index. Official list prices, USD. July 14 figures are from our first tracked snapshot. On July 14, Sonnet 5 was sold at an introductory $2 / $10 with $3 / $15 announced as its standard price.
OpenAI's budget line didn't exist in its current form in July, so I track it from GPT-5.6 Luna's launch price instead. Alibaba and xAI joined our tracking later, on August 17. Since then Alibaba's flagship got cheaper and xAI's price has not moved.
Other tracked lines
| Line | Earlier price | Blended | Now | Blended | Change |
|---|---|---|---|---|---|
| OpenAI budget | GPT-5.6 Luna at launch ($1 / $6) | $2.25 | GPT-6 Luna ($0.10 / $0.50) | $0.20 | -91% |
| OpenAI budget, after July 30 cut | GPT-5.6 Luna ($0.20 / $1.20) | $0.45 | GPT-6 Luna ($0.10 / $0.50) | $0.20 | -56% |
| Alibaba flagship (since Aug 17) | Qwen3.7-Max ($2.50 / $7.50) | $3.75 | Qwen3.8-Max ($2 / $6) | $3.00 | -20% |
| xAI (since Aug 17) | Grok 4.6 ($2 / $6) | $3.00 | Grok 4.6 ($2 / $6) | $3.00 | 0% |
Source: Spectrum AI Labs price index. Alibaba rates are international Model Studio list prices; regional prices differ.
Current API prices, September 25, 2026
Standard rates in USD per million tokens, input and output.
AI API prices by provider
| Model | Provider | Input | Output | Notes |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10 | $50 | Cache reads $0.25 |
| Claude Opus 5.5 | Anthropic | $4 | $20 | Released Sept 22; cache reads $0.20 |
| Claude Opus 5 | Anthropic | $5 | $25 | Previous generation |
| Claude Sonnet 5 | Anthropic | $2 | $10 | Made permanent Aug 10 |
| Claude Haiku 4.5 | Anthropic | $1 | $5 | |
| GPT-6 Astra | OpenAI | $10 | $50 | $20 / $75 over 272K tokens |
| GPT-6 Sol | OpenAI | $2 | $10 | Released Sept 22; $4 / $15 over 272K |
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | Released Sept 22; $0.20 / $0.75 over 272K |
| GPT-5.6 Sol | OpenAI | $4 | $20 | Promotional, at least through Nov 21; list $5 / $30 |
| GPT-5.6 Terra | OpenAI | $2 | $12 | |
| GPT-5.6 Luna | OpenAI | $0.20 | $1.20 | |
| Gemini 3.8 Flash | $0.75 | $3.75 | Intro rate through Dec 31, 2026; then $1.50 / $7.50 | |
| Gemini 3.1 Pro | $2 | $12 | Up to 200K tokens; $4 / $18 above | |
| Gemini 3.5 Flash | $1.50 | $9 | Previous generation | |
| Grok 4.6 | xAI | $2 | $6 | Under 200K tokens; $4 / $12 above |
| Qwen3.8-Max | Alibaba Cloud | $2 | $6 | International rate |
| GLM-5.3-Flash | Z.ai | $0.15 | $0.50 | Standard rate checked Sept 9; the launch promotion has ended |
Source: Official provider pricing pages, checked September 25, 2026. See Sources below.
Caching changes the math for agents that re-read the same context: Fable 5.1's cache reads cost 2.5% of its input price, and OpenAI's GPT-6 models charge 10% for cached input. Long prompts also cost more on several models. OpenAI bills the whole request at a higher rate once a prompt passes 272K tokens, and Gemini 3.1 Pro and Grok 4.6 switch to higher rates above 200K.
What changed between July and September
Every price move in the index, in order.
- June 30: Claude Sonnet 5 launches at an introductory $2 / $10, with $3 / $15 announced as its standard price.
- July 21: Gemini 3.6 Flash launches at $1.50 / $7.50.
- By July 25: GPT-5.6 Sol, Terra and Luna are listed at $5 / $30, $2.50 / $15 and $1 / $6.
- July 30: OpenAI cuts GPT-5.6 Terra to $2 / $12 and Luna to $0.20 / $1.20.
- August 10: Anthropic makes Sonnet 5's $2 / $10 permanent instead of raising it to $3 / $15.
- August 13: Gemini 3.7 Flash launches at an introductory $0.75 / $3.75.
- August: OpenAI cuts GPT-5.6 Sol to a promotional $4 / $20.
- September 1: Claude Fable 5.1 keeps $10 / $50 but cuts cache reads from $1 to $0.25.
- Early September: Gemini 3.8 Flash keeps 3.7's introductory $0.75 / $3.75. GPT-6 Astra launches at $10 / $50.
- September 22: OpenAI releases GPT-6 Sol at $2 / $10 and Luna at $0.10 / $0.50. Anthropic releases Claude Opus 5.5 at $4 / $20.
What this means for your bill
Re-test before you switch.
Not sure which AI model to use?
20 models · Personalized picks · 60 seconds
Providers are fighting hard over the models people run all day and holding their prices on the most capable ones. My guess is that the frontier price stays put until one lab has a clear quality lead to defend. That is an opinion, not something the data proves.
Start by re-testing your default model. If you picked a mid-tier model in July, the same provider probably sells a cheaper one of similar quality now. GPT-6 Sol at $2 / $10 costs exactly what Claude Sonnet 5 costs, and independent tests put it roughly level with GPT-5.6. Our GPT-6 Sol and Luna guide has the details.
When you budget, use standard rates rather than introductory ones. Gemini 3.8 Flash's $0.75 / $3.75 ends on December 31, 2026, and GPT-5.6 Sol's $4 / $20 is promotional. If you plan around those numbers, January will surprise you.
And look at caching before you downgrade. For agent workloads, a cheaper cache rate can save more than a cheaper model. It is why Fable 5.1 can cost less than Fable 5 at the same list price. The Fable 5.1 migration guide explains how.
For which model to use for which task, rather than which is cheapest, see the task-by-task model guide. To estimate your own monthly cost, use the AI cost calculator.
Prices still in flux
Not in the index yet.
Not in this month's table
DeepSeek changed its rate card in September, and we could not confirm the new rates from DeepSeek's own pages in time for this update, so DeepSeek is out of this month's index. MiniMax and Sakana AI also published new models this month, but their pricing either does not fit a simple input and output rate or could not be confirmed. I will add them once the official rate cards are stable. Until then, check each provider's pricing page directly.
How we track this
Official pages only, with every old price kept.
Each price comes from the provider's own pricing or model page, never from a reseller or aggregator. The history comes from our cost calculator's price data, which is version-controlled, so every past value here traces back to a dated snapshot. Tracking started on July 14, 2026, and Alibaba and xAI were added on August 17. In the change log, dates are the provider's effective dates where the provider published one.
The index lists standard pay-as-you-go API prices for prompts under each provider's long-context threshold. It leaves out batch, flex, priority, regional and enterprise rates. The blended price is (3 × input + output) ÷ 4 per million tokens, and your own ratio may differ.
This page keeps the same URL. I update it when a major provider changes a price, and check it at least once a month.
Sources
- Anthropic: API pricing
- Anthropic: Claude Opus 5.5
- Anthropic: Claude Fable 5.1 and Mythos 5.1
- OpenAI: API pricing
- OpenAI: Introducing GPT-6 Sol and Luna
- OpenAI: GPT-5.6 Sol model page
- Google: Gemini API pricing
- Google Cloud: Vertex AI generative AI pricing
- xAI: Grok 4.6 model page
- Alibaba Cloud: Model Studio pricing
- DeepSeek: API pricing
FAQ
What is the cheapest AI API model in September 2026?
GPT-6 Luna is the cheapest model in our index at $0.10 per million input tokens and $0.50 per million output tokens. Z.ai's GLM-5.3-Flash is close behind at $0.15 and $0.50, then GPT-5.6 Luna at $0.20 and $1.20. Google's cheapest current option is Gemini 3.8 Flash at an introductory $0.75 and $3.75 through December 31, 2026. DeepSeek is not in the index this month because its September rates could not be confirmed.
Are AI API prices going down?
In the middle and budget tiers, yes. Between July 14 and September 25, 2026, blended list prices fell 64% for OpenAI's main model line, 56% for Google's Flash tier on its introductory rate (11% on the standard rate), and 33% for Claude Sonnet 5's standard price, although Sonnet 5 was already sold at the lower price in July. The most capable models did not get cheaper: Claude Fable 5.1 and GPT-6 Astra both list at $10 input and $50 output per million tokens.
How much does Claude cost per million tokens?
As of September 25, 2026: Claude Fable 5.1 is $10 input and $50 output, Claude Opus 5.5 is $4 and $20, Claude Opus 5 is $5 and $25, Claude Sonnet 5 is $2 and $10, and Claude Haiku 4.5 is $1 and $5 per million tokens. Cached input costs less on every model.
What is a blended token price?
It is a single number that combines input and output prices using a fixed ratio, so models can be compared on one scale. We use three input tokens for every output token, a common pattern for chat and agent workloads. The formula is three times the input price plus the output price, divided by four.
How often is this price index updated?
We update it when a major provider changes a price or releases a new model, and check it at least once a month. Each update keeps the same URL and records the check date at the top of the page. The historical figures come from our cost calculator's dated snapshots since July 14, 2026.
Founder of Spectrum AI Labs — testing AI tools and models, and writing up what actually ships.
More about Paras →Stay ahead of the AI curve
We test new AI tools every week and share honest results. Join our newsletter.
