Anthropic has moved enterprise plans to per-token billing. The seat fee covers platform access; all usage across Claude, Claude Code, and Cowork is billed separately at standard API rates.
The change was documented in Anthropic’s enterprise billing help center article (updated ~April 7, 2026) and reported by Implicator AI and NPI Financial. Legacy enterprise plans with bundled usage will transition at their next renewal.
“The seat fee only covers access to the platform and doesn’t include any usage. All usage across Claude, Claude Code, and Cowork is billed separately at standard API rates, based on what your team actually consumes.”
This is not a surprise. It is the predictable consequence of a flywheel that has become overloaded.
The flywheel problem
We have been writing about Claude Code’s cache and cost mechanics for two weeks. The investigation started with a cache regression (Part 1), expanded into the three-layer gating model (Part 2), examined what “5x” legally means (Part 3), and culminated in a practical optimization guide that achieved 99.5% cache hit rate across a 61-hour session (How to Get 99.5%).
Throughout that investigation, a structural observation kept surfacing: Anthropic’s consumer subscription model appeared to be selling surplus compute capacity from an infrastructure built for enterprise peak demand. The GPU clusters exist to serve government contracts and large API customers with strict SLAs. The consumer tiers — Pro, Max 5x, Max 20x — filled the idle cycles, converting waste into revenue.
Think of the infrastructure as a flywheel — a massive rotating system built to store and deliver energy on demand. Anthropic’s GPU clusters must be provisioned for peak enterprise load, which means the flywheel is always spinning, always consuming power, whether or not anyone is drawing from it. A flywheel only works if it has load — useful work to absorb its rotational kinetic energy. Without load, it dissipates that energy as no-load losses: friction, heat, wasted torque. Consumer subscriptions were the load that converted those losses into useful work, making the infrastructure economically viable during off-peak hours.
But load is only useful until it exceeds the torque the flywheel can deliver. When demand from consumer and enterprise users together surpasses the rotational energy available, the flat-fee model breaks — because a flat fee has no mechanism to regulate demand. The vendor’s only tools are throttling (which they deployed via TTL gating and quota caps) and pricing (which they are now deploying via per-token billing).
Today’s announcement is the pricing lever being pulled.
What changed
Before: Enterprise plans bundled usage into the seat fee. Customers paid a predictable monthly amount regardless of token consumption.
After: The seat fee covers access. Every token consumed across Claude, Claude Code, and Cowork is billed at standard API rates — the same rates that apply to direct API customers.
The consumer tiers (Pro, Max 5x, Max 20x) have not been explicitly changed in this announcement. But the direction is clear: Anthropic is moving from “all you can eat” toward “pay for what you use.” Whether consumer tiers follow the same path is a matter of when, not if.
Why cache optimization just became a line item
Under flat-fee billing, cache efficiency saved you quota — an abstract percentage that reset every 5 hours. Under per-token billing, cache efficiency saves you dollars.
The math is simple. Our interceptor telemetry from 14,000+ API calls shows:
| Metric | Value |
|---|---|
| Total tokens processed | 3.04 billion |
| API-equivalent cost (with caching) | $1,834 |
| API-equivalent cost (without caching) | $14,965 |
| Cache savings | $13,131 (87.7%) |
Under per-token billing, the difference between optimized and unoptimized usage is $13,131 over 10 days. That is not a quota percentage. That is a line on an invoice.
The tools we built for cache optimization — claude-code-cache-fix (prompt cache interceptor) and claude-code-meter (cost analytics with live dashboard) — were designed around quota efficiency. Under per-token billing, they become cost controls. The same fixes that saved quota now save money directly.
What to do about it
If you are on an enterprise plan affected by this change, or if you expect consumer tiers to follow, three things matter immediately:
1. Know your cost profile
Install claude-code-meter and run claude-meter analyze. It shows your API-equivalent cost by token type, cache savings, and per-model breakdown. Under per-token billing, these numbers are your actual bill.
2. Optimize your cache
Install claude-code-cache-fix. The interceptor fixes client-side cache bugs that cause unnecessary cache misses — each miss is a full-price rebuild of your context at API input rates. Our data shows 87.7% cost reduction from cache optimization alone.
The 15-minute setup from our optimization guide applies directly: upgrade to the latest CC version, install the interceptor, disable git-status injection, and set up the status line for real-time visibility.
3. Understand your token mix
Not all tokens cost the same. Our OLS regression across 57 sessions shows:
- Output tokens dominate cost (Pearson r = 0.57 with quota drain)
- Cache read tokens are nearly free (r = 0.28, coefficient ~7e-9)
- Cache creation tokens are expensive (the rebuild penalty)
Under per-token billing, the optimization strategy is: maximize cache reads (cheap), minimize cache creation (expensive), and be deliberate about output token generation (most expensive).
Built by the community, for the community
These tools are not the work of one person or one company. The cache-fix interceptor has 8 contributors across 4 countries, is translated into 4 languages, and has been independently audited by a security researcher who assessed it as a legitimate tool with no network exfiltration.
The investigation that produced these findings was community-driven from the start. @VictorSun92 wrote the original monkey-patch fix. @jmarianski reverse-engineered the cache structure via MITM proxy and Ghidra. @ArkNill’s hidden problem analysis provided the foundational dataset. @bilby91 at Crunchloop deployed the interceptor to production and validated it against real Agent SDK workloads. @fgrosswig built a forensic dashboard that visualizes the token flows. @TomTheMenace, @arjansingh, @beekamai, @JEONG-JIWOO, @X-15, and @thepiper18 contributed Windows support, VS Code integration, safety hardening, and translations.
Under flat-fee billing, this community effort saved people quota frustration. Under per-token billing, it saves them money. The community dashboard and GitHub Discussions are open to anyone who wants to contribute data or build on the work.
The broader pattern
Every major AI provider running agent workloads is expected to move to usage-based billing. The flat-fee era was a market-entry strategy, not a sustainable business model. Compute is a finite resource, and when demand exceeds supply, the market clears via price.
This is not a criticism of Anthropic. It is physics. A flywheel requires continuous torque to maintain its angular momentum against friction. When the compute capacity that supplies that torque is finite and demand for useful work is growing, the cost of torque must eventually be passed to the consumer. The question was never whether this would happen, but when.
For Claude Code users, the practical implication is clear: cache optimization is no longer a nice-to-have. It is cost management. The tools exist. The data exists. The 15-minute setup pays for itself on the first session.
Tools referenced in this post:
npm install -g claude-code-cache-fix@latest
npm install -g claude-code-meter@latest- Interceptor: https://github.com/cnighswonger/claude-code-cache-fix
- Meter + dashboard: https://github.com/cnighswonger/claude-code-meter
- Live dashboard: https://meter.veritassuperaitsolutions.com
- Optimization guide: How to Get 99.5% Cache Hit Rate
H/T @mhbosch for surfacing the pricing coverage.
Have questions about optimizing your Claude Code costs? We help teams build AI-augmented development workflows that scale. Get in touch.