"The AI Subsidy Is Ending: What Your LLMs Actually Cost"

# The AI Subsidy Is Ending: What Your LLMs Actually Cost

A $200 Claude Max 20x subscription buys you roughly $1,000 to $5,000 in API-equivalent compute. That is the estimate from ExplainX, backed by independent token-tracking across heavy Claude Code users. GitHub Copilot cost Microsoft up to $80 per user per month to serve at the $10 price point in early 2023. Across the entire industry, flat-rate AI subscriptions have been subsidised by venture capital, below-cost pricing strategies, and a land-grab that values market share over unit economics.

That era is ending. Not next year. Now.

The subsidy is real, and it is large

The Mind Momentum analysis published this month is the cleanest measurement I have seen. Over three months (March through May 2026), one heavy Claude Code user tracked every token they consumed and priced it at published API rates. The result: $600 in subscription payments against $5,883 of compute consumed. That is a 9.8x subsidy.

Zed Industries estimated the same gap at 15-30x for heavy agent SDK users running Opus. The Register reported in April that Anthropic was restructuring its enterprise pricing "in ways licensing analysts estimated would double or triple costs for heavy users." GitHub announced Copilot moves to usage-based pricing in June 2026. OpenAI already shifted enterprise customers to API-equivalent billing last month.

This is not a coincidence. Every provider is converging on the same answer: flat-rate subscriptions were never designed for the consumption patterns of autonomous agents and programmatic pipelines. A human sending a few dozen prompts per day looks nothing like an agent loop generating thousands of API calls an hour. The subsidy works until it does not.

Anthropic paused the change. The pattern did not.

On June 15, Anthropic was scheduled to split Claude Agent SDK usage off subscription rate limits and onto a separate credit pool billed at full API rates. A Pro subscriber would get $20 in monthly API credits. A Max 20x subscriber would get $200. When the credit runs dry, you pay API rates or your agents stop.

On the day itself, Anthropic walked it back. A line appeared at the top of the billing page: "We're pausing the changes to Claude Agent SDK usage described below. For now, nothing has changed."

Do not read that as the subsidy continuing. Read it as a delay while they figure out the rollout mechanics. GitHub already made the switch. OpenAI already made the switch. Anthropic stated its intention publicly. The direction of travel is clear.

The gap between what a subscription costs and what API service costs to deliver is a gap that only gets wider the more you automate. If your workflows depend on that gap, you have a pricing risk the size of your token volume.

The four pricing layers that hide the real cost

Most teams running AI workloads see only the top layer: the monthly subscription or API invoice. Below that, four structural dynamics determine what you actually pay:

Layer 1: Provider headroom. API pricing includes margin that covers training-cost amortisation, R&D, safety research, and profit targets. The per-token cost to serve is lower than the list price. OpenAI is estimated to lose money on every ChatGPT Pro subscriber because o3-pro single queries can cost up to $1,000 in compute. The gap between what you pay and what it costs to serve you is widening, not narrowing.

Layer 2: Caching discounts. Provider-native prompt caching at 90% off cached input tokens is the largest single lever. But the fine print matters: Anthropic charges a cache-write premium (1.25x for 5-minute TTL, 2x for 1-hour), and Google's explicit caching adds per-hour storage fees ($1/MTok/hr for Gemini 3.1 Pro). At low hit rates, caching costs more than it saves.

Layer 3: Routing inefficiency. Sending every request to the cheapest capable model cuts average cost by 40-70% on most workloads. Most teams do not do this. The gap between what you pay and what you could pay is entirely within your control.

Layer 4: The multiplier problem. Agent loops, retries, tool call fanout, and context-length growth compound in ways linear cost models miss. A single user action triggering 200 internal LLM calls costs 200x a single request. The 90th percentile request often costs 10x the median. If you are averaging instead of measuring percentiles, you are underestimating your exposure.

The industry subsidy sits on top of all four layers. When it disappears, the arithmetic of every layer is exposed at once.

What to do before the meter turns on

The playbook has three steps, and they are the same regardless of which provider changes its pricing first.

First, measure at API rates. Whatever you are paying in subscriptions, price your actual token consumption at standard API list prices for the models you use. The gap is your exposure. If it is small, you have nothing to worry about. If it is large, the next two steps are urgent.

A reliable way to do this: instrument every LLM call with a route tag, log the actual tokens consumed, and multiply by the relevant provider rate. Do not estimate. Do not average. Pull the per-request numbers. Cost achieves this by wrapping your provider clients directly and attributing every euro to a route, but you can also build it yourself with middleware if you have the engineering time.

Second, optimise the levers you control. Provider-native prompt caching is free engineering time and delivers the highest ROI. Exact-match response caching is a half-day project with any gateway. Model-tier routing is a two to three day project. Each one compounds. Together they cut a typical bill 60-80%.

The order matters: caching first (it is risk-free), then max_tokens discipline (thirty minutes), then routing (two to three days). Each layer shrinks the bill the next layer works on.

Third, design for API pricing. Build your agent pipelines, your caching strategy, and your model selection around the assumption that you will eventually pay API rates for everything. If your product's unit economics rely on a 10x subsidy, redesign the product. The teams that survive the transition will be the ones whose economics work at list price, not the ones who complain about how expensive things used to be.

A worked example: a $200 subscription vs API pricing

Here is the arithmetic for a real scenario. A solo developer runs Claude Code on a daily basis with Sonnet 4.6. A typical session moves roughly 2M input tokens and 200K output tokens. With prompt caching covering 70% of input, the API cost per session breaks down like this:

• 0.6M fresh input at $3/MTok = $1.80
• 1.4M cached input at $0.30/MTok = $0.42
• 200K output at $15/MTok = $3.00
• Total: $5.22 per session

Twenty working days of one session each: $104.40 at API rates. A Max 5x subscription at $100/month covers this and remains the cheaper option for this usage pattern. But push the usage to 40 sessions a month at $208.80, and the subscription no longer covers the full cost at metered rates. The Max 20x $200 credit now caps out.

This is the key insight: above a usage threshold that is lower than most people think, the flat rate no longer covers the bill. The provider either absorbs the loss (subsidy) or customers pay the difference. The current market bet is that the latter happens.

Now apply the same arithmetic to an agent loop. A single agent task that triggers 50 tool calls, each reading 4K of context and writing 500 tokens, costs roughly 250K input tokens and 25K output tokens per completed task. At Sonnet rates, that is about $1.13 per task in API cost. A $200 credit covers roughly 175 such tasks per month. A busy solo developer integrating agents into their daily workflow can hit that in two to three weeks. After that, the meter is on.

What this looks like for a team

For a small team of five developers each running similar agentic workloads, the monthly API-equivalent cost is roughly $522 per person - over $2,600 for the team. Under the current subsidy model, a Team Premium plan at $100/seat covers this. Under the metered model that is coming, the actual cost surfaces as a line item on the bill.

The engineering response is not to stockpile subscriptions. It is to measure your per-task cost, optimise the four layers I described above, and ensure your product margins work at API rates. Caching alone typically closes the gap by 30-60%. Routing closes it further. You may find that even at full API pricing, the discipline of optimisation brings your effective cost below what you are paying in subscriptions today.

The metric that matters

Track cost per completed task. Not cost per token, not cost per session, not aggregate monthly spend. Cost per task captures model tier, caching effectiveness, routing efficiency, retry overhead, and context length in a single number. It is the number you compare against the value the task delivers. It is the number that tells you whether the unit economics work.

If you know your cost per task at API rates today, you know your exposure when the subsidy ends. If you do not, you are flying blind.

I want to be clear: the subsidy ending is not a bad thing. Flat-rate subscriptions distort incentives. They encourage wasteful token consumption because the marginal cost of one more request feels like zero. Metered pricing aligns cost with usage and forces the discipline that separates a sustainable AI feature from one that burns cash.

The question is not whether the meter will turn on. It is whether you will know your numbers before it does.

---

Cost wraps your Anthropic, OpenAI, and Gemini clients in one line, attributes every euro to a route, and shows you per-request spend at API rates before the invoice lands. Free tier covers 100,000 events per month. No card needed.

If you are looking for the provider-specific side of this, read How to Reduce Your Anthropic Bill and the broader LLM Cost Optimisation guide.