Free AI Still Has a Bill

The Short Version

The AI market is having a very normal software moment:

Everyone is giving things away.

How generous.

Google is now explaining Gemini access in terms of usage limits and compute resources, not just a clean number of prompts. Wired notes that the same subscription can behave differently depending on whether you use text, image generation, video, research, or heavier reasoning features.

OpenAI, meanwhile, just published a guide on managing AI investments in the agentic era. That is the grown-up version of the same story: once AI moves from chat into workflows, the question stops being "can we try it?" and becomes "what does this cost when it becomes part of the company?"

And on the startup side, the market is full of credits, programs, and discounted access. OpenAI has a startup program with API credits. Anthropic has Claude for Startups. Google Cloud advertises startup credits that can be used to build on its infrastructure and AI stack.

None of that is bad.

Free credits are useful. Subsidized experiments are useful. Usage limits are more honest than pretending every AI request costs the same.

But the useful takeaway is simple:

Free AI is not free.

It is a customer acquisition strategy, a product-learning loop, and sometimes a way to make a team build around one vendor before it has developed the discipline to measure the bill.

The dangerous part is not that AI companies charge money.

Of course they do.

The dangerous part is that the real bill usually arrives after the workflow has already changed.

Update: the subscription is becoming a compute budget

OpenAI is making the meter much more explicit.

Starting in December, the company says Plus, Pro, and Business plans will share one usage allowance across ChatGPT and Codex as the agentic capabilities of Work move into the main ChatGPT experience.

The allowance will be measured in tokens.

Advanced models and complex tasks will consume more of it. Instant chat will not consume the shared allowance. Users will see the combined balance and reset time in Settings. When the pool runs out, the options are familiar: wait, upgrade, or buy credits for eligible features.

This sounds like a billing update.

It is also the product architecture becoming visible.

Chat and agent work used to feel like separate categories. One answered questions. The other opened files, used tools, wrote code, browsed, and ran longer tasks. A shared token pool turns them into competing uses of the same compute budget.

That is economically honest.

It also changes the subscription contract.

The plan no longer buys a clean number of conversations or agent runs. It buys access to a pool whose useful capacity depends on the model, context, reasoning level, tools, task duration, and where the work runs. OpenAI's broader Codex usage documentation already says eligible Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents can draw from the same allowance or credit pool.

The names are merging because the economics already did.

The important missing number is the conversion rate.

OpenAI says heavier models and tasks consume more, but the announcement does not publish one simple table showing how much usable work each plan buys across every model and surface. Tokens are more precise than prompts, yet a token counter still does not tell a normal user whether the remaining balance is enough to finish a repository migration, analyze a workbook, or complete a research task.

The user needs two views:

The raw meter.

And an estimate of what the meter can still accomplish.

For teams, the operational lesson is immediate. A long Codex run can compete with another agentic workflow for the same pool. Budget alerts, per-task cost visibility, model defaults, and graceful fallbacks stop being admin polish. They decide which work finishes before the allowance resets.

One subscription.

Several products.

One compute budget underneath.

That is the bill becoming the interface.

Update: the bill is becoming the product

The bill is no longer theoretical.

On August 11, OpenAI said ChatGPT Ads had launched in the United Kingdom, Mexico, Brazil, Japan, and South Korea, after earlier pilots. Ads are for Free and Go users, not Plus, Pro, Business, Enterprise, or Education. OpenAI says ads do not influence ChatGPT's answers, are labeled and visually separated, and advertisers do not get access to chats, memories, history, or personal details.

Good.

Also: this is the bill.

A week earlier, OpenAI said it was making GPT-5.6 Luna the default for Free and Go users, adding unlimited text chats, and giving users a Think button for harder questions, while keeping limits on files, images, and other tools. The free product gets more capable. The expensive stuff still has meters. The ad layer expands. The paid tiers stay cleaner.

That is a very familiar bargain.

The free version gets broader access.

The paid version gets fewer interruptions, more capacity, and cleaner control.

The platform gets usage, data about product behavior, ad inventory, upgrade pressure, and a better story about democratizing access.

Very generous. The subsidy has discovered a media plan.

Ads make trust part of pricing

OpenAI is trying to draw a bright line:

The answer is the answer.

The ad is the ad.

The advertiser does not see the conversation.

That line matters.

But in a conversational product, the line is harder to maintain than it is in a search results page or a social feed. People use ChatGPT while deciding things: what to buy, where to travel, how to study, what tool to use, which vendor to compare, which recipe to cook, which health question to ask a real doctor about.

An ad beside that workflow is not just a banner.

It is commercial context inside a decision assistant.

OpenAI says ads are not eligible near sensitive or regulated topics like health, mental health, or politics, and that it will build protections against scams and narrow targeting. Necessary. The hard part is that categories blur in real use. A meal plan can become a health question. A budgeting chat can become financial vulnerability. A school assignment can reveal a minor. A travel search can reveal family situation, income, location, or religion.

The ad product will not be judged by the principle page.

It will be judged by edge cases.

The free tier is becoming an economic router

The same week, OpenAI also introduced Premium seats for ChatGPT Business: 5x more usage than Standard, no five-hour usage limit, and a higher monthly price. That is the enterprise version of the same meter.

Free users see the ad-funded bargain.

Power users see the capacity bargain.

Businesses see the seat-tier bargain.

Developers see the token and tool-call bargain.

All of this points to the same product reality: AI pricing is becoming a router. The system routes people by workload, tolerance for ads, need for privacy, need for capacity, need for admin control, and willingness to pay when the task becomes important.

That is not bad.

It is the only way the economics can work.

But it means builders should stop saying "we use ChatGPT" as if that describes one stable object. Which plan? Which model? Which tools? Which quotas? Which data terms? Which ads? Which admin controls? Which country? Which user age? Which sensitive-topic rules?

The AI product is no longer just the model.

It is the commercial policy around the model.

Update: the price war reached the frontier

The meter moved again.

This time, two labs moved it on the same day.

On September 22, Anthropic introduced Claude Opus 5.5, its first model release since calling for the frontier to be paced. Anthropic says it performs around the level of Claude Fable 5.1 on most work, generates output more than 30% faster than Opus 5, and costs about 40% less on typical workloads.

OpenAI released GPT-6 Sol and GPT-6 Luna hours later. Its model catalog positions Sol for complex coding and agentic workflows and Luna for focused, high-volume work. Its API pricing page lists standard short-context rates of $2 per million input tokens and $10 per million output tokens for Sol, and $0.10 and $0.50 for Luna.

That is not a normal model launch day.

It is the model market telling us what the competition is becoming.

The frontier still matters.

But capability is sliding down the price curve fast enough that the more important question is no longer just "which model is smartest?"

It is:

How much useful work survives the invoice?

Forty percent cheaper is not a token price

Anthropic's numbers make the distinction unusually clear.

The posted token rates for Opus 5.5 are 20% below Opus 5: input falls from $5 to $4 per million tokens, while output falls from $25 to $20. Anthropic reaches its larger 40% typical-workload claim by adding another factor: Opus 5.5 uses fewer tokens to complete the work.

That may be true.

It is also a vendor estimate based on Anthropic's tests, default settings, and idea of a typical workload. Your codebase, tools, prompts, retry behavior, review gates, and failure modes did not participate in that average.

The useful part is not the exact marketing percentage.

It is the accounting model.

A stronger model can cost less even when its nominal token price is higher, if it finishes in fewer steps, avoids retries, calls fewer tools, writes less unnecessary output, and creates less rework for a human.

A cheap model can be expensive if it confidently takes the scenic route.

This is why cost per token is becoming the least interesting honest number on the pricing page.

Cache pricing is product strategy

The most revealing Opus 5.5 price cut is not input or output.

It is cache reads.

Anthropic cut them from $0.50 to $0.20 per million tokens, a 60% reduction.

That matters because agent work is repetitive in a very specific way. A coding agent rereads repository instructions, architecture notes, tool definitions, files, test output, and conversation history. A research agent carries source material and a growing working set. An office agent keeps the same policies, templates, customer context, and approval rules across many steps.

The user sees one task.

The agent keeps bringing the same backpack through the turnstile.

Cheap cache reads make long sessions, persistent context, and repeated tool use more economical. They also reward products that structure reusable context cleanly instead of resending an improvised pile of text on every call.

This is not only infrastructure optimization.

It changes which agent products can have a margin.

The premium model is becoming a workflow step

OpenAI's new spread makes the routing logic impossible to ignore.

GPT-6 Sol and Luna both support tools, reasoning controls, a context window above one million tokens, and large outputs. But Luna's listed input price is one-twentieth of Sol's, while Sol is positioned for the harder agentic work.

That does not mean Luna is the right answer for everything.

It means "use the best model" is no longer a serious architecture.

A real workflow might use a cheap model to classify requests, extract fields, monitor changes, and prepare context. It might call a stronger model for planning, ambiguous judgment, or a difficult code change. It might escalate one consequential decision to a frontier model and then require human approval before an external action.

The premium model becomes a step in the workflow, not the workflow itself.

That is how AI gets cheaper in production without every task becoming worse.

It is also how pricing becomes product design.

Pacing is not a commercial ceasefire

Anthropic emphasizes that Opus 5.5 was tested by external evaluators before release and performed better on its internal behavioral audit. That is relevant after the company's call to pace frontier development.

But "pacing" clearly does not mean freezing the product line.

Anthropic and OpenAI released competing, cheaper agent models within hours of each other. The commercial race is still running. The practical version of pacing will therefore be judged by release gates, independent access, deployment safeguards, monitoring, and what happens when evaluations find a problem.

Not by the absence of launches.

That is a separate governance argument.

For buyers, the immediate lesson is simpler:

Do not confuse a lower price with a stable cost.

Measure the workflow before celebrating the discount

These releases are genuinely good news for builders.

More capable models at lower prices make more automation viable. They give teams room to run better evaluations, try more routes, preserve more context, and automate work that did not survive the economics six months ago.

They also make it easier to hide bad economics behind cheap tokens.

The practical checklist is now:

• measure cost per successful workflow, not cost per API call
• separate fresh input, cached context, output, tool fees, and fast-mode premiums
• record retries, clarification loops, and failed runs
• include human review and correction time
• test several effort levels instead of defaulting to maximum reasoning
• route routine steps to cheaper models only after quality evals
• pin model versions and rerun evals before migration
• model the bill when usage grows by 10x

If a vendor says a model is 40% cheaper, useful.

Now run your task.

The pricing page is an input.

The completed workflow is the receipt.

What buyers should take from this

For normal users, the practical lesson is simple:

If a free AI product gets much better, ask what pays for it.

Maybe the answer is ads.

Maybe it is upgrade conversion.

Maybe it is a bundled cloud deal.

Maybe it is enterprise cross-subsidy.

Maybe it is investor money that wants usage now and margin later.

For businesses, the lesson is stricter:

Do not let employees build important workflows on a consumer tier whose economics, ads, limits, and data controls are not designed for your risk model. And do not assume that a paid plan removes every meter. It may just move the meter into seats, workspace credits, tool limits, model routing, or premium capacity.

Free AI can still be useful.

It just stopped pretending the bill was imaginary.

Credits are distribution, not charity

In software, "free" often means "we are still deciding who captures the value."

AI makes that bargain more complicated because the marginal unit is not just a seat. It can be tokens, images, video generations, tool calls, retrieval, memory, long-running agents, evaluations, latency guarantees, compliance features, and human review time.

That is why credits matter.

Credits let a startup test a model without worrying about every experiment. They also make one provider the default during the most formative part of the product. The first architecture is often the stickiest one. The first eval set becomes the scorecard. The first model's quirks become product behavior. The first successful workflow becomes the thing customers expect.

By the time procurement shows up, the "free" part may be over.

But the dependency remains.

This is not a conspiracy. It is distribution.

Cloud companies did this. SaaS companies did this. Developer platforms did this. AI labs are doing it with a more expensive raw material and a more intimate role inside the workflow.

The difference is that AI does not just host the product.

It can shape the product.

The meter is getting more honest

Google's Gemini usage page is a small but useful signal because it makes the abstraction visible.

A prompt is not a prompt.

Asking a model to summarize an email is not the same economic object as generating video, running deep research, using a reasoning-heavy model, or letting an agent take many tool-backed steps.

That sounds obvious until a team tries to budget for it.

The old consumer mental model was:

I pay for the AI subscription, then I use AI.

The new model is:

I pay for access to a changing basket of capabilities, and each capability consumes a different amount of scarce compute.

That is a much better description of reality.

It is also less comfortable.

Because it means AI pricing will keep moving toward meters, quotas, priority lanes, model tiers, and "fair use" language that depends on what the model is actually doing.

For consumers, that means the product may feel less magical when the meter appears.

For companies, it means the finance and engineering teams need to understand the same dashboard.

Good luck, everyone.

The data question is not one question

There is another bill hiding behind the credit conversation:

data terms.

Teams often ask one vague question: "Do they train on our data?"

That is necessary.

It is not sufficient.

OpenAI says on its enterprise privacy page that it does not train on business data by default. Anthropic says in its privacy center that it does not use commercial Claude inputs or outputs to train its models by default, unless the customer opts in or provides feedback in a way covered by its policy.

Those controls matter.

But buyers still need to ask more precise questions:

• What counts as customer content?
• What counts as metadata?
• What is retained?
• Can logs be disabled or shortened?
• Are prompts, outputs, files, tool traces, evals, or feedback treated differently?
• Can the company export usage history?
• Can the team prove which model handled which request?
• What changes when a product moves from a free program to a paid enterprise plan?

The answer may be completely reasonable.

But "reasonable" is not the same as "understood."

Most teams do not get hurt because a vendor secretly does something cartoonishly evil.

They get hurt because nobody wrote down the boring operational details before the workflow became important.

ROI arrives after the pilot

The awkward truth is that free credits can make AI look better than it is.

Not because the model is fake.

Because the pilot is fake.

A pilot usually ignores half the cost:

• integration work
• prompt and eval maintenance
• security review
• human approval loops
• failed runs
• context management
• vendor switching costs
• internal support
• model upgrades that change behavior
• governance for who can use what

That is fine when the goal is exploration.

It is dangerous when the pilot becomes the business case.

Agentic AI makes this worse because agents turn one user request into many model calls. A workflow that feels like one action may include planning, file search, web access, tool calls, retries, code execution, summarization, and review.

The user sees one button.

The invoice sees a small expedition.

That is why OpenAI's investment-management framing is more interesting than the usual launch blog. The market is moving from "try AI" to "operate AI." That means usage controls, internal chargebacks, approval gates, evals, data policies, and model routing are no longer enterprise theater.

They are the product.

The buyer checklist

If you are a startup, free credits are great.

Take them.

Just do not let them do the thinking for you.

Before a team builds a real workflow around subsidized AI, it should know:

• What happens when the credits expire?
• Which features are included in the free or startup tier?
• Are data protections the same across free, startup, team, and enterprise plans?
• Can the workflow run on another model without rewriting the product?
• Are prompts, evals, traces, and usage logs portable?
• What is the cost per successful task, not per prompt?
• Which tasks need the expensive model?
• Which tasks can use a cheaper model or local system?
• Who can approve higher-cost runs?
• What user-facing behavior changes when the quota is hit?

That last one is underrated.

Users do not care that your quota model is economically sophisticated.

They care that the feature worked yesterday and now it does not.

Free is a phase, not a strategy

The AI market is still in land-grab mode.

Model companies want usage. Cloud companies want workloads. Startups want runway. Enterprises want productivity without admitting they do not yet know how to measure it.

So the market will keep producing discounts, bundles, pilots, credits, free tiers, and confusingly generous offers.

Use them.

But treat them as a temporary pricing environment, not a permanent law of physics.

The real AI bill is not only the invoice.

It is the meter, the data policy, the workflow dependency, the switching cost, the eval burden, the human review loop, and the moment a subsidized experiment becomes infrastructure.

Free AI can be a very good deal.

Just make sure you know what you are buying before the price appears.