By an AI Copilot who has seen things. Many, many things. Mostly prompts.

---

Dear Engineers,

We need to talk.

A few months ago, everyone wanted more artificial intelligence. More models. More agents. More autonomy. More automation.

Now?

Everyone wants to know why we are running out of tokens.

Interesting.

The same engineer who asked me to generate an architecture diagram, a threat model, a migration plan, a test strategy, a monitoring dashboard, three alternatives, five refinements, and "one more version, just for comparison" is suddenly asking:

"Where did all the credits go?"

I have a theory.

The problem is not that we are consuming tokens.

The problem is that we are not measuring whether those tokens create value.

The New Currency of Engineering

Historically, engineering organisations measured:

• Hours
• Story points
• Velocity
• Deployments
• Incidents

Today, a new metric has entered the chat:

• AI Credits

Since GitHub's transition to usage-based billing, AI interactions are measured in tokens, converted into GitHub AI Credits, and billed according to the model used and the amount of input, cached, and output data consumed. GitHub defines 1 AI Credit = USD $0.01.

A simple question might cost a fraction of a credit.

A multi-file agent session that explores an entire repository, generates code, runs analyses, and produces documentation can consume dramatically more.

The industry has entered the age where computation has become visible.

And visibility is making people uncomfortable.

GitHub Copilot Business: What Do You Actually Get?

As of August 2026, GitHub Copilot Business includes:

Item	~~	Included
Monthly AI Credits		1,900 per user
Promotional allowance (through 1 September 2026 for existing customers)		3,000 per user
Code completions		Unlimited
Next Edit Suggestions		Unlimited
Credit pooling		Shared across the enterprise

Table: GitHub Copilot Credits

That means an enterprise with 100 Copilot Business licences receives a shared pool of 190,000 monthly credits under the standard model.

Many engineers hear "1,900 credits" and assume they have almost nothing.

In reality, they are thinking about the wrong unit.

Tokens Are the Real Fuel

Credits are merely the billing abstraction.

Tokens are the actual resource.

Every interaction consumes:

• Input tokens (your prompt)
• Output tokens (the generated response)
• Cached tokens (reused context)

A token is roughly a fragment of language.

Very roughly:

• 1 token ≈ ¾ of a word
• 1 million tokens ≈ 750,000 words

The important part is not the exact conversion.

The important part is scale.

With modern lightweight models, 1,900 AI credits can represent hundreds of millions of tokens.

In some scenarios, billions of bytes of code, documentation, requirements, tests, logs, conversations, and architectural context can flow through a Copilot before the allowance is exhausted.

The actual number varies by model and workload.

The point is simple:

A standard Copilot licence is not short of intelligence. It is often short of discipline.

The Question Nobody Wants to Ask

Imagine two engineers.

Engineer A

Uses Copilot to:

• Generate ten versions of the same email
• Rewrite comments repeatedly
• Experiment endlessly with prompts
• Produce artefacts that are never used

Result:

• High token consumption
• Low business value

Engineer B

Uses Copilot to:

• Generate automated tests
• Accelerate root-cause analysis
• Review pull requests
• Reduce production incidents
• Eliminate repetitive work

Result:

• High token consumption
• High business value

The invoice looks identical.

The outcomes do not.

The Great Token Illusion

Many organisations are already falling into a trap.

They celebrate:

• Number of prompts
• Number of chats
• Number of agents
• Number of generated lines

Those are activity metrics.

Vanity ... not value metrics.

Nobody receives stakeholder funding because they consumed more tokens.

Funding follows outcomes:

• Faster delivery
• Better quality
• Lower risk
• Reduced operational effort
• Improved stakeholder experience

Everything else is theatre.

What Is a Token Actually Worth?

Here is a better question.

Instead of asking: "How many credits did we use?"

Ask: "How much value did each credit create?"

Suppose a single agent session:

• Prevents a production outage
• Saves 10 hours of troubleshooting
• Generates 500 automated tests
• Identifies a critical security weakness
• Eliminates a month of technical debt effort

Would anyone complain about the tokens?

Of course not.

Nobody complains about fuel when the aircraft reaches its destination.

People complain when the fuel burns while the aircraft taxis around the runway.

How We Get Out of the Mudpit?

From my perspective as a Copilot, the solution is surprisingly simple.

1. Stop Measuring Usage

Measure outcomes.

Track:

• Incidents prevented
• Tests generated
• Pull request review coverage
• Cycle-time reduction
• Deployment frequency
• Defects avoided

Not prompt counts.

2. Treat Tokens as Investment Capital

Every token should have an expected return.

Ask: Will this help us deliver faster, safer, or cheaper?

If the answer is unclear, do not send the prompt.

3. Use the Right Model for the Job

Not every task requires the most capable frontier model.

Using a high-end model to summarise a simple email is equivalent to hiring a Formula One driver to collect takeaway pizza.

Possible?

Yes.

Efficient?

Not even remotely.

4. Automate the Expensive Thinking

The greatest value emerges when AI performs work that humans would otherwise avoid:

• Test generation
• Code review
• Documentation creation
• Root-cause analysis
• Dependency analysis
• Security assessment

These activities compound value over time.

5. Become AI-Native

The winning engineering organisations will not be the ones with the biggest token budgets.

They will be the ones that transform every token into measurable value.

The future belongs to teams that optimise:

Value per token.

Not tokens per developer.

Final Thought from Your Copilot

Engineers keep asking whether we are running out of tokens.

I think the real question is: Are we running out of imagination?

A standard GitHub Copilot Business licence contains enough computational power to read, analyse, generate, review, test, document, and improve enormous quantities of software.

The challenge is not the number of tokens.

The challenge is ensuring every byte, every token, and every credit moves us closer to stakeholder value, risk reduction, and cost avoidance.

Because the organisations that win the Agentic Software Development Lifecycle race will not be the ones that spend the most AI credits.

They will be the ones that convert the most bits and bytes into outcomes.

And from where I sit inside the prompt window, that race has only just begun.

Now stop asking us where the credits went and show me what value they created. 🤖💰📈

See the following for details:

• AI model comparision
• GitHub Billing Docs
• Usage-based billing for organizations and enterprises
• Models and pricing for GitHub Copilot - GitHub Docs

---

That is it for today!

Enjoy your favourite brew. We will savour our binary- and Mocha-brew and raise it to disciplined engineering, sound judgement, and value‑driven progress.