Book your free 30-day diagnosticprices read 7 – 23 Sept 2026
read 7 – 23 Sept 2026 · 36 prices
This weekTokensComputeWhat moved

Cloud and AI prices,
read from the source.

Every list price that a cloud or AI bill is built from, read off the provider’s own page and stamped with when. What moved, who moved it, and what it was before.

22 models · 6 GPU prices · 5 data platform prices · read 7 – 23 September 2026
Latest move1 Sept
-75%Anthropic cut the cached input price on Claude Fable 5.1.
$1.00→$0.25per 1M cached input tokensfirst-party · anthropic.com

What a million tokens costs, by model.

The dearest output token on the board is GPT-6 Astra at $50.00 and the cheapest is Llama 4 Scout at $0.59, a spread of 85 times.

ModelInput / 1MOutput / 1MCached inputContext90 days
OA
GPT-6 Astra
OpenAI · flagship
new 3 Sept
$10.00$50.00$1.001.05Mnewnew 3 Sept
A
Claude Fable 5.1
Anthropic · flagship
cached $1.00 → $0.25
$10.00$50.00$0.25200k-75%cached $1.00 → $0.25
G
Gemini 3.1 Pro Preview
Google · flagship
$2.00$12.00$0.201.05M0%
M
Mistral Large 3
Mistral
$0.50$1.50$0.05262k0%
DS
DeepSeek V4-Pro
DeepSeek
output $0.87 → $1.98
$0.66$1.98$0.0221M+128%output $0.87 → $1.98
X
Grok 4.6
xAI
new 12 Aug
$2.00$6.00$0.50500knewnew 12 Aug
L
Llama 4 Maverick
Llama via Together AI
$0.27$0.85not published1.05M0%

US dollars per million tokens, standard tier, as each provider publishes it. Grok and DeepSeek charge more for long prompts or at peak hours; the row shows the standard or off-peak rate.

What an hour of hardware and a unit of warehouse cost.

GPU hours, on demand

InstanceRegionPer hourPer month, always onPricingRead from
Az
NVIDIA H100 NVL
Azure · Standard_NC40ads_H100_v5 · 1 GPU
East US · provider price list
East US$6.98/ h$5,095/ moon-demandprovider price list
Az
NVIDIA A100 80GB
Azure · Standard_NC24ads_A100_v4 · 1 GPU
East US · provider price list
East US$3.67/ h$2,681/ moon-demandprovider price list
Az
General purpose
Azure · Standard_D4s_v5
East US · provider price list
East US$0.192/ h$140/ moon-demandprovider price list
AW
NVIDIA H100 80GB
AWS · p5.48xlarge · 8 GPU
us-east-1 · provider price list
us-east-1$55.04/ h$40,179/ moon-demandprovider price list
AW
NVIDIA L4
AWS · g6.xlarge · 1 GPU
us-east-1 · provider price list
us-east-1$0.805/ h$588/ moon-demandprovider price list
AW
General purpose
AWS · m6i.xlarge
us-east-1 · provider price list
us-east-1$0.192/ h$140/ moon-demandprovider price list
G
NVIDIA H100 80GB
GCP · a3-highgpu-8g · 8 GPU
us-central1 · via aggregator
us-central1$87.83/ h$64,118/ moon-demandvia aggregator
G
NVIDIA L4
GCP · g2-standard-4 · 1 GPU
us-central1 · via aggregator
us-central1$0.707/ h$516/ moon-demandvia aggregator
G
General purpose
GCP · n2-standard-4
us-central1 · via aggregator
us-central1$0.194/ h$142/ moon-demandvia aggregator

Data platforms and storage

PlatformPriceUnitRegionRead from
S
Snowflake
Snowflake credit, Standard edition, on-demand
via aggregator
$2.00USD per creditAWS US East (N. Virginia)via aggregator
D
Databricks
DBU price, All-Purpose Compute, Premium tier
provider price list
$0.55USD per DBUAzure East USprovider price list
M
Microsoft Fabric
F64 capacity (64 Capacity Units), pay-as-you-go
provider price list
$11.52USD per hour (64 CU x $0.18/CU-hour)US Eastprovider price list
A
Azure Blob
Hot tier, LRS, general purpose v2, first 51,200 GB/month
provider price list
$0.0208USD per GB per monthUS Eastprovider price list
S
S3
S3 Standard storage, first 50 TB/month
provider price list
$0.0230USD per GB per monthUS East (N. Virginia)provider price list

Azure, AWS and Databricks are read from the provider’s own price list. Google Cloud and Snowflake publish theirs on pages a machine cannot read, so those rows come through a site that reads their price lists, and each one says so.

The latest price changes, newest first.

3 Sept
newOpenAI · Launched GPT-6 Astra, its new top-tier flagship model, at $10 input / $50 output per 1M tokens
first-party·openai.com
2 Sept
newGoogle · Launched Gemini 3.8 Flash at introductory pricing of $0.75 input / $3.75 output per 1M tokens, scheduled to double to $1.50/$7.50 on 2027-01-01
first-party·blog.google
1 Sept
-75%Anthropic · Launched Claude Fable 5.1 at the same $10/$50 per 1M input/output price as Fable 5, but cut the prompt-cache read price 75% from $1.00 to $0.25 per 1M tokens
first-party·anthropic.com
31 Aug
-33%Anthropic · Cancelled a scheduled price rise for Claude Sonnet 5, making the $2/$10 per 1M introductory rate permanent instead of raising it to $3/$15 on 2026-09-01
26 Aug
newMeta Llama (via Groq) · Groq moved its two most popular Llama models off self-serve pricing entirely, pulling them from the public rate card.
aggregator·usagepricing.com