The AI Price Hike Wave Is Here: Copilot Quadruples, the Free Lunch Is Over

The Free Lunch Is Disappearing

Over the past two years, AI tools have swept the globe with incredibly low prices—or even free access—letting developers enjoy service value far exceeding what they paid. However, in April 2025, back-to-back announcements from GitHub Copilot and Anthropic Claude are revealing a harsh reality: the era of AI companies burning cash to subsidize users has come to an end.

Price hikes, usage caps, feature cuts—the same playbook we've seen from Netflix, Uber, and DoorDash is now playing out at accelerated speed in the AI space.

GitHub Copilot: Signups Paused, Prices Quadrupled

Around April 20, 2025, GitHub Copilot made sweeping changes to its service plans—changes so dramatic they caught users completely off guard.

Here's what changed:

• New user signups were directly paused—they won't even let you in the door
• Existing users' usage was strictly limited, with mechanisms like weekly quotas
• Premium models (like Opus) were removed from the base plan—want them? Pay for a more expensive tier
• New pricing tiers were added, with prices quadrupling outright

What does this mean? The good old days of paying a few dozen dollars a month for unlimited access to top-tier AI models are officially over. GitHub clearly realized that moderate-to-heavy users were consuming computational resources far beyond what their subscription fees covered. Every active user was costing the company money—and the losses were only growing.

There's a deep technical cost logic behind this. The inference cost of large AI models is just as steep as training costs, and exhibits clear diseconomies of scale. Every time a user sends a request, GPU clusters in data centers must perform massive matrix computations. Take OpenAI as an example: estimates suggest that the marginal cost of each ChatGPT conversation is more than 10x that of a traditional search engine query. More critically, as model capabilities increase (parameters growing from 70 billion to hundreds of billions), inference costs rise exponentially rather than linearly. This explains why top-tier models like Claude Opus were the first to be removed from base plans—each invocation of these models may consume 20-50x the compute of lightweight models, making break-even impossible under a fixed subscription model.

Claude Code's "One-Day Trip": Cut and Then Restored

Anthropic's move was even more dramatic. Claude's code functionality (Claude Code) was originally included in the $20/month Pro plan, but around April 22, the feature suddenly disappeared from the $20 tier.

However, the decision immediately triggered fierce user backlash. Facing overwhelming public pressure, Anthropic reversed course almost instantly, quickly adding Claude Code back to the $20 plan.

This small episode exposed a critical fact: Anthropic is losing heavily on the $20 Pro plan. They wanted to cut features to stop the bleeding, but user dependency was already too high—abruptly reducing value would trigger massive backlash. The company found itself in a classic dilemma: continue subsidizing and keep losing money, or stop subsidizing and face user churn.

From Per-User to Per-Token Pricing: A Fundamental Business Model Shift

Beyond individual user adjustments, the billing model for enterprise AI is also quietly transforming.

Old model (per-user pricing): A fixed monthly fee per person ($20/30/100), including a certain number of API calls and token allowances.

New model (per-token usage pricing): Every API call, every token is precisely metered and billed.

To understand the impact of this shift, you need to understand what tokens actually are. A token is the fundamental unit that large language models use to process text—in English, one token corresponds to roughly 3/4 of a word; in Chinese, one character typically maps to 1-2 tokens. Mainstream model APIs like GPT-4 typically distinguish between "input tokens" and "output tokens," with output tokens priced 2-4x higher than input because generating text consumes more compute than understanding it. Take Claude 3 Opus as an example: its API pricing is roughly $15 per million input tokens and $75 per million output tokens. A moderately complex code completion request might consume thousands of tokens, meaning a heavy user's actual monthly compute consumption could easily exceed the fixed subscription fee by tens of times.

The logic behind this shift is clear. Under fixed-fee models, heavy users extract value far exceeding what they pay, and AI companies cannot effectively control costs. Per-token billing transfers the cost risk to users—the more you use, the more you pay—and companies no longer foot the bill for high-consumption users.

For enterprises with hundreds or thousands of employees, this means AI tool costs could increase by several times or even an order of magnitude.

The Classic VC Subsidy Playbook: AI Is Just the Latest Episode

If all this feels familiar, it's because this is the classic Silicon Valley venture capital subsidy playbook:

1. Burn cash for user acquisition: Attract massive user bases with prices far below cost
2. Build dependency: Get users deeply integrated into the product ecosystem, creating usage habits and switching costs
3. Harvest: Raise prices, cut features, push premium tiers

This strategy is a commercial variant of what economists call "predatory pricing," with the core logic being "grow market scale first, then build moats through network effects." During Uber's 2014-2016 peak, the company subsidized roughly 40% of the cost per completed ride. Netflix expanded rapidly in the early 2010s with rock-bottom prices, driving traditional rental giants like Blockbuster into bankruptcy—and now subscription fees have multiplied several times over, with ads added on top. DoorDash and other delivery platforms have also pushed delivery fees to jaw-dropping levels.

AI may be the most devastating version of this playbook. What makes AI unique is that its "lock-in effect" is far stronger than streaming or ride-sharing services. When a company's codebase, workflows, and employee habits have all been restructured around specific AI tools, the switching cost isn't just money—it's a combined toll of time, efficiency, and organizational change. When enough people depend on AI to write code and run businesses without possessing the underlying skills themselves, they're completely locked in. A 10x price increase? They'll grit their teeth and pay, because without AI they simply cannot function.

Who Will Be Hit Hardest?

Enterprise Users Are First in Line

Companies with 100, 1,000, or even 10,000 employees using AI tools face a massive expense if AI subscription prices increase 5x. More critically, these enterprises have often deeply embedded AI into their workflows—switching costs are extremely high, but bargaining power is limited.

Enterprise software procurement typically relies on Enterprise Agreements to lock in pricing and service terms, but the rapid iteration of AI tools has rendered traditional procurement frameworks ineffective. Most AI service providers currently offer enterprise contracts of only 1 year, retaining the right to unilaterally adjust pricing. Even thornier: AI tools often bypass IT procurement by being positioned as "productivity boosters," with business units subscribing directly—creating so-called "Shadow IT." When prices rise, enterprises face a dilemma: centralized negotiation requires first understanding actual usage across the entire company (data that's often scattered across departments), while decentralized purchasing means completely forfeiting bargaining power. This information asymmetry is precisely why AI service providers hold the upper hand in pricing games.

Developers Who Depend on AI for Coding

Developers who primarily rely on AI-assisted programming with weak foundational coding skills face the greatest risk. When tools like Copilot see price spikes, they can neither afford the steep costs nor independently complete development work without the tools.

The Entire AI Application Ecosystem

The ripple effects of price increases will spread across the entire ecosystem—from SaaS products to API consumers, from indie developers to large enterprises. Every scenario that heavily consumes AI compute will feel the cost impact. Products built on AI APIs will see their profit margins squeezed even further.

Enjoy the Present, Prepare for the Future

We are still in the "honeymoon phase" of AI tools.