The Unlimited AI Subscription Is Dying

The flat monthly AI subscription is staying. The unlimited part is doing a disappearing act. GitHub Copilot now meters much of its advanced model-driven usage through AI Credits. OpenAI Codex ties flexible usage to token consumption. Claude lets eligible paid users keep working beyond included limits with prepaid credits.

The pricing page is increasingly the entrance fee. The real bill is what an accepted workflow costs after tokens, retries, tool calls, and human supervision. This is evidence-based analysis using official provider announcements, help centers, and billing documentation current to July 29, 2026.

What changed with GitHub Copilot?

GitHub Copilot provides the clearest example of the shift. Under its previous system, premium model usage was largely measured through requests and model multipliers.

Under the current system, most model-driven Copilot activity is measured through GitHub AI Credits. The cost of an interaction depends on two things:

1. 1The model used
2. 2The number of tokens consumed

Those tokens include:

• Input tokens sent to the model
• Output tokens generated by the model
• Cached tokens reused as context

GitHub then converts the usage into AI Credits. One AI Credit equals one US cent.

A lightweight chat question might use only a fraction of a credit. A cloud agent working across a large repository with a frontier model may use substantially more.

Current Copilot individual-plan allowances

As of July 1, 2026, GitHub lists the following monthly allowances:

The split between base credits and flex matters because only one side is designed to stay fixed.

Base credits

These match the subscription price and, according to GitHub, do not change. A $10 Pro subscription includes 1,000 base credits. A $39 Pro+ subscription includes 3,900 base credits.

Flex allotment

This is additional included usage added on top of the base. GitHub describes flex as variable. It may change as model pricing, product features and the economics of AI evolve.

It is not simply a short-term promotional bonus. The current total allowance is therefore larger than the subscription price alone would suggest, but not every part of that total is guaranteed to remain fixed indefinitely.

What remains unlimited in Copilot?

The move to AI Credits does not mean every Copilot interaction is metered. On paid Copilot plans:

• Code completions remain unlimited
• Next Edit Suggestions remain unlimited
• Neither feature consumes AI Credits

This is an important carve-out. A developer who primarily uses Copilot for inline suggestions may notice little practical difference.

A developer who relies heavily on chat, agents, code review, CLI work or large-repository tasks is more exposed to variable consumption. GitHub is effectively keeping lightweight, always-on assistance inside the subscription while metering the more compute-intensive work.

What happens when Copilot credits run out?

Users who exhaust their included credits generally have three options:

1. 1Upgrade to a higher plan
2. 2Set a budget and pay for additional usage
3. 3Wait for the monthly allowance to reset

Additional spending is set in US dollars. Because one AI Credit equals one cent, a $10 additional-usage budget covers 1,000 extra credits.

That gives users control, provided they set the controls before the bill arrives. It also means the headline subscription price may not be the final monthly cost for heavier workloads.

Why GitHub made the change

The old request-based system treated an interaction as a request. The new system measures more closely what that interaction actually consumes.

Consider two Copilot tasks.

Task A

Explain what this function does.

The model reads a small amount of code and produces a short answer.

Task B

Review this repository, identify the authentication bug, update the relevant files, run the tests and explain the fix.

The agent may:

• Read hundreds of files
• Search the repository
• Build a plan
• Edit several files
• Run commands
• Review failures
• Retry the task
• Produce a final summary

Both tasks begin with one user instruction. They do not create the same cost.

Model choice also matters. A smaller model handling a routine question may cost far less than a frontier reasoning model processing a large context window and generating a long response.

Usage-based billing turns those differences into visible consumption.

Codex is moving in the same direction

OpenAI's Codex products reflect the same underlying shift. On April 2, 2026, OpenAI moved Codex pricing for Plus, Pro, Business and new Enterprise plans away from per-message estimates and aligned it with API token usage.

On April 23, OpenAI extended the change to remaining existing Enterprise, Edu, Health, Government and teacher plans. Codex usage is now calculated using credits per million:

• Input tokens
• Cached input tokens
• Output tokens

The practical effect is fairly straightforward. A short task with limited context may consume relatively few credits.

A long-running task using a premium model, a large codebase, several parallel instances or fast mode may consume considerably more. Users still receive access through their plan. Credits provide additional flexible usage when they need to continue beyond included limits.

The unit being sold is shifting from access alone toward the amount of work completed.

Claude adds usage credits beyond the plan limit

Anthropic still offers familiar individual subscriptions:

• Claude Pro
• Claude Max 5x
• Claude Max 20x

Each includes a different amount of plan usage. Eligible paid users can also enable usage credits.

After reaching the included limit, the user can choose to continue working through prepaid credits billed at Anthropic's standard API rates. Those charges are separate from the subscription.

Claude also provides controls including:

• Monthly spending caps
• Auto-reload settings
• Usage alerts
• Real-time consumption reporting
• Usage history

This creates a hybrid model:

A subscription for predictable baseline access, then consumption-based pricing beyond that baseline.

That hybrid model is becoming increasingly common.

Why the flat subscription worked at first

Early generative AI was mostly conversational. Users asked questions, drafted emails, summarized documents and generated ideas.

The workload was relatively short and easy to understand. AI companies also had strong reasons to subsidize usage.

They wanted to:

• Attract users
• Build daily habits
• Gather feedback
• Increase market share
• Establish a default interface
• Prove demand to investors and enterprise buyers

Simple pricing made adoption easier. Users did not need to understand tokens, caching, context windows, model pricing or agent runtime.

They only had to decide whether the product was worth the monthly fee. That simplicity helped the market grow.

AI usage no longer looks simple because the work is no longer simple.

Agents changed the economics

A chatbot waits for a question and produces an answer. An agent may:

• Break a goal into several tasks
• Search files or websites
• Call external tools
• Query a database
• Generate code
• Run tests
• Retry failed steps
• Operate in the background
• Coordinate with other agents

Each action may create another model call. One visible task can contain dozens of invisible interactions.

That is why agent pricing differs from ordinary software pricing. A project-management app may cost roughly the same to provide whether a user creates ten tasks or one hundred.

An AI agent that performs one hundred reasoning steps consumes meaningfully more infrastructure than one that performs ten. The additional work may produce far more value.

It can also produce a larger bill.

AI pricing is starting to resemble cloud computing

Traditional software subscriptions are usually priced around access. Cloud infrastructure is priced around consumption.

Businesses pay for resources such as:

• Storage
• Compute time
• Data transfer
• Database operations
• Function invocations

AI is moving toward the same logic. The emerging billing units include:

• Input tokens
• Output tokens
• Cached tokens
• AI Credits
• Tool calls
• Search operations
• Agent runtime
• Workflow executions
• Generated images, audio and video

The names differ between providers. The direction is consistent.

AI products are becoming metered infrastructure wrapped in consumer-friendly interfaces.

Usage-based does not automatically mean more expensive

Metering can be fairer for light users. Someone who uses an advanced feature occasionally may pay less than they would under a high flat subscription.

Teams can also control costs by routing simple work to smaller models and reserving frontier models for tasks that need deeper reasoning. Usage reporting can show:

• Which workflows consume the most credits
• Which models create the highest costs
• Which users require larger allowances
• Whether expensive models produce better outcomes
• Where budgets should be applied

The problem is not usage-based pricing itself. The problem is when the pricing becomes difficult to understand.

Why AI bills are getting harder to predict

A traditional SaaS comparison might look like this:

An AI cost comparison increasingly looks like this:

The monthly price is only the starting point. Two people paying for the same plan can create very different costs depending on how they work.

Seven AI pricing traps to watch

1. The subscription hides the usage ceiling

A plan may cost $20 each month while including only a defined amount of advanced usage. Always ask:

What happens when the included allowance runs out?

2. Different models consume credits at different rates

The strongest model may produce a better result. It may also consume an allowance much faster than a smaller model that could have completed the task adequately.

3. Long context is not free

Large documents, codebases, project files and conversation histories give the model more material to process. That context can materially increase consumption.

4. Agents create invisible work

The user sees one instruction. The system may perform many calls, tool actions and reasoning steps before returning the result.

5. Retries multiply consumption

An agent that fails, revises the work and tries again can use far more resources than the original instruction suggests. That may still be worthwhile, but it should not be invisible.

6. Variable allowances can change

A flex allotment, temporary boost or launch incentive may make a plan look generous today. Check which part of the allowance is fixed and which part the provider may adjust.

7. Automatic overage can create bill shock

Auto-reload and pay-as-you-go features allow work to continue without interruption. They can also increase spending without an obvious stopping point if budgets are not configured.

The better metric is cost per useful outcome

AI tools should not be judged only by subscription price or cost per token. A cheap model that repeatedly fails may cost more than an expensive model that completes the task correctly on the first attempt.

A $100 agent plan may also be excellent value if it reliably saves several hours of skilled work. The more useful question is:

What did it cost to produce an acceptable result?

That includes:

• Subscription cost
• Usage charges
• Failed attempts
• Human correction time
• Review time
• Integration overhead
• Value of the completed work

The lowest token price does not always create the lowest total cost.

Example: when the cheaper model loses

Imagine two models completing the same research task.

Model A

• Costs $1 to run
• Misses important details
• Requires 30 minutes of correction
• Must be run twice

Model B

• Costs $5 to run
• Produces a reliable result
• Requires five minutes of review
• Works on the first attempt

Model A looks cheaper on the pricing page. Model B may be cheaper for the business.

That is why AI pricing needs to be evaluated at the workflow level, not only at the model level.

What businesses should do now

1. Separate subscription cost from usage cost

Track both:

• Fixed monthly fees
• Variable credits and overage

A $20 tool with frequent extra usage may cost more than a $50 plan with a suitable included allowance.

2. Identify high-consumption workflows

Look for tasks involving:

• Large files
• Long conversation histories
• Full repositories
• Extended research
• Multiple tool calls
• Parallel agents
• Repeated generations
• High-resolution media

These workflows are the most likely to create variable costs.

3. Route work to the right model

Do not use the most expensive model by default. Smaller models may be sufficient for:

• Classification
• Simple extraction
• Formatting
• Routine summaries
• Basic transformations

Reserve premium models for work that benefits from deeper reasoning or greater reliability.

4. Set budgets before enabling overage

Use:

• Monthly caps
• Per-user limits
• Spend alerts
• Approval thresholds
• Team-level reporting

Define acceptable spending before the first unexpected bill.

5. Track fixed and variable allowances

Record whether included usage is:

• Fixed
• Variable
• Promotional
• Scheduled to expire

Do not assume every credit included today will remain unchanged.

6. Measure recurring workflows

Run a representative task several times. Measure:

• Credits consumed
• Average cost
• Failure rate
• Human-review time
• Monthly frequency

Then estimate monthly spend using real work rather than a pricing-page example.

7. Review AI tools more frequently than traditional SaaS

A tool that represented good value three months ago may now have:

• Different limits
• New credit rules
• Lower-cost models
• Higher overage rates
• A stronger competitor
• A better plan for the same workflow

Annual reviews are too slow for important AI tools.

What individuals should check before subscribing

Before paying for an AI product, answer five questions:

1. 1What is included in the monthly fee?
2. 2Which actions consume credits?
3. 3Do different models consume the allowance at different rates?
4. 4What happens when the allowance runs out?
5. 5Can additional spending be capped?

If the pricing page does not make these answers clear, treat that as part of the product evaluation. Good AI pricing should be understandable before the bill arrives.

Will unlimited AI disappear completely?

Probably not. Some products will continue offering generous flat access because:

• Their models are cheaper to operate
• Their tasks use limited context
• They apply fair-use restrictions
• Most subscribers use far less than the maximum
• Unlimited access remains a competitive advantage

But "unlimited" will increasingly come with conditions. Those conditions may include:

• Rate limits
• Fair-use rules
• Slower processing
• Restricted premium models
• Lower priority at busy times
• Separate agent allowances
• Extra credits for advanced features

The word may remain. The economics behind it will change.

AI is becoming a variable business expense

Most software is budgeted as a fixed cost. Ten seats multiplied by the subscription price produces a reasonably predictable bill.

Usage-based AI creates more variability. Costs can increase when:

• More employees adopt the tool
• Users select stronger models
• Workflows process more context
• Agents perform more steps
• Automated usage grows
• More business processes move into AI systems

That does not make AI poor value. It means AI needs to be managed more like cloud infrastructure than a simple app subscription.

Why this matters for your AI stack

The best AI stack is not the one with the most subscriptions. It is the one where every tool has a clear role and a defensible cost.

For each tool, you should know:

• Why it is in the stack
• Which workflow it supports
• What the subscription includes
• Which actions create additional charges
• What a successful outcome costs
• Whether another tool could do the job more efficiently

Without that visibility, small charges can accumulate across several products without anyone understanding the total. One tool may still look affordable.

The stack may not.

The Choosely verdict

Usage-based pricing is not automatically bad. It can be fairer for light users and more flexible for heavy ones. The problem starts when allowances, multipliers, retries, and overage rules are difficult to see before the work begins. Track fixed fees, variable spend, and cost per accepted outcome. A US$20 plan that needs four attempts and an afternoon of repairs is not a bargain. It is a variable bill with a tidy headline price.

The cheapest plan can still be the right choice. It just has to prove it at the workflow level.

Sources

• GitHub: Copilot is moving to usage-based billing
• GitHub: Updates to Copilot billing and plans
• GitHub Docs: Usage-based billing for individuals
• GitHub Docs: Models and pricing for Copilot
• OpenAI: Codex rate card
• OpenAI: Using credits for flexible usage in ChatGPT
• Anthropic: Manage usage credits for paid Claude plans
• Anthropic: Choose a Claude plan