AI Token Pricing for Developers: The Ledger

Reading Time: 5 minutes

What if the cheapest token you will ever buy is the one you already paid for yesterday? Seedhi baat: most developers are overpaying for AI, and not because the models got expensive.

This week is about AI token pricing for developers: what the three meters are, and how to shrink the bill.

🪙 The Token Diaries, Episode 2 of 8 — Last week we mapped seven ways to read a single token. This week we go deep on the first lens, the developer’s, where a token is money and a limit. Read Episode 1 here.

🧾 AI Token Pricing: Three Meters, One Bill

Every API call runs three meters at once.

Input tokens are what you send. Output tokens are what the model writes back. Cached tokens are input the provider has already seen and remembers.

Look at the price sheet and one thing jumps out. Output is the expensive meter. On Anthropic’s current price list, output tokens cost five times what input tokens do, whether the model is the premium Fable 5.1 ($10 in, $50 out per million) or the budget Haiku 5.5 ($0.10 in, $0.50 out for shorter prompts).

So the first rule of the ledger is simple. A long, chatty answer costs far more than a long question.

📌 Real-world case: Anthropic’s own pricing page lists input and output rates side by side for every Claude model, and the five-to-one ratio holds across the current Claude lineup, from the priciest model to the cheapest. Longer replies, longer reasoning, longer bills. (Anthropic pricing docs)

🪙 Token Truth: You pay more for what the model says than for what you say.

🪟 The Context Window Is a Tab You Keep Open

A context window is how much the model can hold in view at one time. Bigger sounds better. But every token inside that window gets re-sent and re-billed on each call.

Here the vendors differ, and the fine print matters.

Anthropic now includes the full one-million-token window at standard rates on its 4.6-and-later models, with one exception. A 900,000-token request costs the same per token as a 9,000-token one. The exception is Haiku 5.5, the cheap model, which charges higher rates once a prompt crosses 100,000 tokens.

Same window. Different meter. Read the tier table before you paste a whole codebase into the prompt.

📌 Real-world case: On Haiku 5.5, input costs $0.10 per million tokens for prompts up to 100,000 tokens and $0.50 per million above that. A single request that crosses the line pays the higher rate even when part of it comes from the cache. (Anthropic pricing docs)

🪙 Token Truth: A bigger window is a bigger tab, and the tab stays open on your bill.

♻️ The Discount Hiding in Plain Sight

Now the good news.

Prompt caching lets the provider remember the part of your prompt that never changes: the system prompt, the tool definitions, the long reference document. Write it once, read it many times.

On Anthropic’s published rates, a five-minute cache write costs 1.25 times the normal input price. A cache read costs a tenth of it on most models, and even less on the newest ones. Do the arithmetic for two requests: without caching you pay 2.0 units, with caching you pay 1.35. The savings grow with every repeat.

Batch processing stacks another 50% off for work that can wait.

Researchers have now measured this on real agent runs. An academic study of more than 500 agent sessions, spanning OpenAI, Anthropic and Google, found that caching cut API costs by 41 to 80 percent and improved time to first token by 13 to 31 percent.

But there is a catch, and it is the best part of the paper.

📌 Real-world case: The same study found that blindly caching everything can backfire. The steadiest savings came from caching the stable system prompt, putting dynamic content at the end, and leaving changing tool results out. Naive full-context caching sometimes made responses slower. (arXiv: Don’t Break the Cache)

Sach batao, how many production prompts open with today’s date or a user name? That one dynamic line at the top quietly breaks the cache for everything after it.

🪙 Token Truth: A cached token is the cheapest token you will ever buy, if you keep the front of your prompt boring.

📉 Cheap Is Getting Cheaper

Not every job needs your best model. Summarising a ticket, tagging an email, extracting a field from an invoice: these are boring jobs, and boring jobs belong on small models.

The price gap is now dramatic. Anthropic’s Haiku 4.5 lists at $1 in and $5 out per million tokens. Haiku 5.5, for prompts up to 100,000 tokens, lists at $0.10 in and $0.50 out. That is a tenth of the per-token price for short prompts.

Good developers now route. A small model handles the high-volume, low-stakes calls, and the expensive model is reserved for the genuinely hard ones.

In practice that is a simple two-tier setup. Classification, extraction and short summaries go to the small model by default. Anything that fails a quick check, or needs real reasoning, gets escalated. You pay premium rates only for the calls that earn them.

📌 Real-world case: Anthropic’s price table puts Haiku 4.5 at $1 / $5 and Haiku 5.5 at $0.10 / $0.50 per million input and output tokens for prompts up to 100,000 tokens, so the same short, boring task costs roughly a tenth as much on the newer small model. Prompts above 100,000 tokens move to $0.50 / $2.50. (Anthropic pricing docs)

🪙 Token Truth: Pick the smallest model that does the job. Save the big one for the job that needs it.

💸 When the Flat Fee Dies

Here is the uncomfortable part. For years, developer tools hid the ledger behind a flat monthly fee. That is ending.

Remember GitHub Copilot from last week? Since June 1, 2026 its plans bill in AI Credits tied to token use. Each plan includes credits equal to its price, so Pro at $10 a month gets $10 of credits. GitHub said the old premium-request model was “no longer sustainable.”

And the bill is not only an individual’s problem.

📌 Real-world case: Uber used up its full-year budget for AI coding tools in roughly the first four months of 2026, driven largely by heavy use of agentic tools such as Claude Code. Its CTO told The Information the company was “back to the drawing board.” Uber then capped token spending on agentic coding tools at $1,500 a month per tool, with a dashboard so employees can see their own usage. (Outlook Business, citing Bloomberg and The Information, GitHub blog)

Notice what Uber did not do. It did not ban the tools. It put a meter in front of them.

🪙 Token Truth: When the flat fee disappears, the ledger lands on your desk.

🧮 A Developer’s Checklist

Five habits that pay for themselves:

Put stable text first. System prompt, tools and reference material at the front, dynamic bits at the end, so the cache survives.
Cache on purpose. Treat a 90% discount on repeat input as a design goal, not a lucky accident.
Route by difficulty. Small model by default, big model by exception.
Cap the output. Ask for the format you need. Long answers are the expensive meter.
Watch the tier line. Know where your model’s price steps up, such as Haiku 5.5’s 100,000-token threshold.

None of this needs a rewrite. Most of it is reordering a prompt, setting a cap and choosing a model on purpose. Do these five and your bill shrinks without the product getting any worse. That is the whole game of the developer’s ledger.

So, toh phir: if you opened your last AI invoice today, could you say which of the three meters, input, output or cached, caused most of it?

🔜 Next Tuesday: Episode 3 — Tokenmaxxing Inc.

📬 Episode 3 is on its way. Subscribe to get it in your inbox the moment it goes live.

What happens when a company turns token usage into a scoreboard? Leaderboards, mandates and a very old law about measures and targets.

Bring your own stories of AI dashboards at work. The best ones may show up in the series.

📚 References

What’s your Reaction?