Originally published on The Searchless Journal
The numbers stopped adding up months ago. AI startups burning through cash on API calls that exceeded payroll costs. Enterprises quietly questioning why their AI budgets tripled quarter over quarter. Engineering teams spending more time optimizing token usage than shipping features.
Then Flo Crivello, CEO of AI startup Lindy, did what an increasing number of founders are doing in private but few discuss publicly. He fired Claude.
Not because Claude underperformed. Because the cost curve, in his words, "crashed to the ground" when he switched to DeepSeek hosted on US infrastructure. The 25-person startup had been spending more on AI models than on its entire engineering team. Crivello called it "unsustainable" and told CNBC the switch saved millions.
He is not alone. Across the AI ecosystem, a cost rebellion is reshaping how companies think about model selection, vendor lock-in, and the fundamental economics of building with large language models.
The Price Gap That Broke the Model
Consider the current pricing landscape. Claude Opus 4.7 costs $5 per million input tokens and $25 per million output tokens. GPT-5.5 from OpenAI is even steeper at $5 input and $30 output. Meanwhile, GLM-5.2 from Chinese AI lab Zhipu charges $1.40 per million input tokens and $4.40 per million output. Some third-party providers offer it even cheaper.
That is a 3.5x difference on input and nearly 6x on output. For a company running agentic workloads where models make dozens of tool calls per task, token consumption multiplies rapidly. A single complex coding task might burn through a million tokens. At Opus pricing, that task costs roughly $15 in output alone. At GLM pricing, it costs under $5.
Snowflake CEO Sridhar Ramaswamy ran the definitive benchmark. His team tested GLM-5.2 against Claude Opus 4.7 across 103 coding tasks, each run three times. The models had to write code that worked on both DuckDB and Snowflake. The results were closer than anyone at Anthropic or OpenAI would feel comfortable admitting.
When each model got three attempts per task, the solve rates were 66% versus 67%. Neck and neck. GLM-5.2, a Chinese model costing a fraction of Claude, matched the frontier model on practical coding tasks.
Where Cheaper Models Fall Short
The Snowflake benchmark reveals important nuances. First-attempt accuracy diverged significantly. Opus hit 53.7% on the first try. GLM managed only 47.6%. The Chinese model was less consistent, more likely to produce a wrong answer before eventually finding the right one through iteration.
GLM also burned more tokens to get there. It averaged 99 runs per task versus Opus's 80 and consumed 860 million tokens total, nearly double Opus's 439 million. In one extreme case, GLM fired off 411 tool calls in 24 minutes, checking row counts, distributions, null values, and column types, only to fail all three attempts. Opus solved the same task with 49 calls in 9 minutes.
More checks did not lead to more correct results, Ramaswamy noted. GLM's weakness was giving up too early on some tasks while obsessively checking the wrong things on others. The claim that GLM produces cleaner code did not hold up under scrutiny.
But here is the catch. Even with higher token usage, GLM's total cost per task remained dramatically lower. The price gap is wide enough to absorb significant inefficiency and still deliver savings.
The Agentic Token Problem
For simple chatbot applications, model pricing matters less. A few thousand tokens per conversation adds up slowly. But agentic AI changes the calculus entirely.
Agents do not just answer questions. They plan, execute, verify, and iterate. A coding agent might read files, write code, run tests, analyze failures, and rewrite solutions. Each step consumes tokens. A single complex task can easily consume hundreds of thousands of tokens across multiple model calls.
This is where the pricing gap becomes existential for startups. Lindy's AI costs exceeded personnel costs. Read that again. A 25-person company was spending more on API calls than on salaries, benefits, and office space combined. That is not a sustainable business model. It is a ticking bomb.
The problem compounds as agents become more capable. Better models mean more complex tasks, which means more tool calls, which means more tokens. Capability improvements do not reduce costs. They increase them. Every breakthrough in agentic reasoning makes the pricing problem worse, not better.
Sam Altman acknowledged this publicly, saying AI cost became a "huge issue" for OpenAI's customers. When the CEO of the most valuable AI company admits pricing is a problem, you know the pressure is real.
The China Factor
The geopolitical dimensions of this cost rebellion are impossible to ignore. Chinese AI labs benefit from substantially lower labor costs, government subsidies, and access to cheaper computing infrastructure. They can price aggressively in ways that Western labs, burdened by massive fundraising rounds and sky-high valuations, simply cannot match.
OpenAI is reportedly seeking a valuation approaching $1 trillion in its delayed IPO. Anthropic's explosive growth could push its valuation past OpenAI's for the first time. These valuations assume ever-climbing revenue trajectories. They assume that customers will keep paying premium prices for marginal quality improvements.
But the market is speaking. When a startup CEO switches to a cheaper model and says "it's a matter of survival for the business," that is not a preference. It is a forced hand.
The quality gap is narrowing faster than the price gap. GLM-5.2 is not as good as Opus 4.7 across the board. But it is good enough for many use cases, and it costs a fraction as much. For startups operating on thin margins, "good enough" at one-third the price beats "slightly better" at full freight.
What This Means for AI Strategy
Companies building AI products need to rethink their model strategies. The default approach of picking the best available model and optimizing later is a luxury few can afford.
Multi-model architectures are becoming standard. Use frontier models for complex reasoning tasks where accuracy is critical. Route simpler tasks to cheaper models. Build abstraction layers that let you swap models based on cost-performance tradeoffs.
Token budgeting should be a first-class engineering discipline. Every prompt, every tool call, every retry consumes tokens. Teams need dashboards showing token consumption per feature, per user, per task. Optimization is not optional.
Vendor diversification protects against both price hikes and availability issues. Companies locked into a single provider face existential risk if pricing changes or rate limits tighten. The Lindy switch from Claude to DeepSeek took engineering effort but ultimately reduced costs by orders of magnitude.
Evaluation pipelines become critical when switching models. Snowflake's benchmark worked because they had rigorous evaluation criteria. Without strong evals, you cannot know whether a cheaper model actually performs adequately for your use case.
The Valuation Pressure Cooker
The cost rebellion creates a dangerous feedback loop for AI labs. If customers migrate to cheaper alternatives, revenue growth slows. Slower growth threatens valuations. Lower valuations reduce the capital available for training next-generation models. Fewer breakthroughs widen the window for competitors to close the quality gap.
OpenAI's IPO postponement may reflect this dynamic. Going public requires predictable revenue growth. But if enterprise customers are actively reducing their AI spend by switching to cheaper models, the growth narrative weakens. Anthropic faces similar pressure. Its valuation assumes Claude maintains a meaningful quality premium over alternatives. The Snowflake benchmark suggests that premium is shrinking in coding, one of the most commercially important use cases.
The infrastructure investments make this even more precarious. Data centers, chip orders, energy contracts. These are multi-billion-dollar commitments based on revenue projections that assume continued premium pricing. If the floor falls out of model pricing, the infrastructure bets become massive liabilities.
Looking Ahead
The AI cost rebellion is not a temporary market correction. It is a structural shift. Chinese labs will continue pricing aggressively. Open-source models will continue improving. Enterprise buyers will continue optimizing for cost.
Western AI labs have two paths. Cut prices aggressively to retain customers, accepting lower margins and slower growth. Or demonstrate clear, measurable quality advantages that justify premium pricing. The second path becomes harder with each benchmark showing marginal quality differences at massive price premiums.
For companies building with AI, the message is clear. Stop overpaying for marginal quality gains. Build flexible architectures. Evaluate models on your actual workloads. And treat token costs with the same financial discipline you apply to any other infrastructure expense.
The era of "just use GPT-4" is over. The era of strategic model selection has begun.
Top comments (0)