sky4me presents

The Cost of AI

Two ways to pay, subscriptions versus tokens, and how to cut the bill without cutting the quality.

01 Part one

Two ways to pay

How you get billed

Subscription

A flat fee for a seat.

Fixed monthly cost
Priced per user
Usage capped by rate limits
Same bill whether you use it or not

Tokens (API)

A meter that runs per call.

Pay for every token in and out
Cost scales with usage
No cap, the meter never stops
You pay for exactly what you use

What a subscription actually buys

A fixed, predictable monthly budget
Claude Pro at $20 per month
ChatGPT Business at $25 per user per month
Zero marginal cost per query, up to the cap
Rate limits that throttle heavy days
A bill you pay even when the seat sits idle

When each model wins

Subscription fits

Steady, human-paced use.

Daily hands-on-keyboard work
A fixed, known team size
Predictable workloads
Budget certainty over efficiency

Tokens fit

Spiky or automated use.

Bursty, uneven demand
Automation running at scale
Many light or occasional users
Paying only for real usage

Flat is a bet.

The subscription gamble

02 Part two

What tokens actually cost

How token billing works

Input tokens: everything you send
Output tokens: everything the model writes
Output is priced well above input
The whole context is resent every turn
Images and files count as tokens too
You pay for the model's reasoning, not just its answer

Output price per million tokens (July 2026)

$5 Haiku 4.5

$10 Sonnet 5

$25 Opus 4.8

$30 OpenAI terra

$50 Fable 5

/ Input versus output

Output is the expensive side

5x

Claude Opus output vs its input

6x

OpenAI terra output vs its input

10x

cheapest to priciest output token

You pay per turn.

Every message resends the whole context

The crossover

The flat plan

One price, whatever you do.

Same bill at low or high volume
Cheap once you use it hard
Wasted when the seat is idle

The meter

Illustrative: 20M output tokens/month on Sonnet 5 is roughly $200.

Cheap when volume is low
Climbs with every extra call
Assumes a set monthly volume

03 Part three

Cutting the bill

Two places to cut cost

The API bill: prompt caching
The API bill: batch processing
The API bill: model routing
Context per turn: offload the context
Context per turn: compress the output
Every lever below trades something for the saving

Prompt caching

How it works

Reuse a computed prefix instead of paying for it again.

Cache a stable prefix: system prompt, docs
Cached reads are about 90% cheaper
Best for large, repeated context

The catch

The cache is fussy about what it will reuse.

Exact-prefix match only
Short time-to-live, about 5 minutes
Cache writes cost a 1.25x premium
A dynamic prefix never hits

/ Batch processing

Half price, if you can wait

50 %

cheaper per token

1 batch

many jobs, one submission

~24 h

results come back later, not live

Model routing

How: send easy work to a cheap, fast model
How: escalate only the hard work to a frontier model
How: classify each task by complexity first
How: the cheapest capable model wins
Catch: a wrong route ships a worse answer
Catch: the routing logic is code you own and maintain

Offload the context

How it works

Keep raw bytes out of the model's context.

Hold data in a sandbox, index and search it
Feed the model only the derived answer
About 96 to 98% of context saved

The catch

The authors publish the misses in BENCHMARK.md.

Summaries are lossy
Search can miss the chunk you needed
Round-trips add their own overhead

Compress the output

How it works

Output is the pricey side, so write less of it.

A terse output style, filler words dropped
About 65% fewer output tokens
Targets the most expensive tokens

The catch

The authors publish this in HONEST-NUMBERS.md.

Adds 1 to 1.5k input tokens per turn
Net real-session saving is 14 to 21%
Net-negative on work that was already terse

No free tokens.

Each optimization trades something

/ What the levers give you

The savings, sourced

90 %

off cached input reads

50 %

off with the batch API

96 %

context saved by offloading, up to 98%

Ready to cut your AI bill

Let's talk

sky4me helps you pick the right cost model and engineer the token bill down, without trading away quality.

Book a call