Opinions expressed by Entrepreneur contributors are their own.
Key Takeaways
- Renting AI through an API means renting a product that can change without your knowledge or consent.
- Running open-weight models on hardware you control keeps your data, costs and accuracy in your hands.u003cbru003e
As a big believer in a domesticated superintelligence, I can’t pass up the Cheeky Pint podcast where OpenAI CEO suggested it’s a terrible idea to keep processing your personal data on your personal computers because his company’s data centers could process your data more efficiently. He walked the remark back after immediate backlash.
It’s hard to recognize, yet your business likely lives in that world of centralization already. When using inference via API, you often rent access to a black box that shifts its behavior, pricing and reasoning effort without your knowledge or consent.
The leaky bucket you never see
Ask yourself: how many times has your team asked for more budget due to inconsistent behavior from GPT or Claude? In Harvard Data Science Review, Stanford and UC Berkeley researchers Lingjiao Chen, Matei Zaharia and James Zou tested identical prompts on GPT-4 and GPT-3.5 across snapshots three months apart. GPT-4’s prime-number accuracy fell from 84.0% in March 2023 to 51.1% in June, while its chain-of-thought boost collapsed from 24.4% to -0.1%.
Everyday workflows suffered similar drops. On a happy-number task, GPT-4’s accuracy fell from 83.6% to 35.2%, while directly executable code plummeted from 52.0% to 10.0% because extra formatting broke pipelines. Instruction-following on answer extraction collapsed from 99.5% to 0.5% under the same model name and API.
A 2024 Carnegie Mellon and HKUST study by Wanqin Ma, Chenyang Yang and Christian Kästner found that 58.8% of prompt and model combinations lost accuracy after an API update, and 70.2% of those drops exceeded 5 percentage points. Across updates, 10.9% of individual predictions regressed and 63.8% of those failures occurred at maximum confidence, breaking working predictions 87.9% of the time even when overall accuracy rose.
Researchers Emanuele La Malfa and colleagues at Oxford and the Alan Turing Institute warned in a 2025 AAAI paper — first published in Journal of Artificial Intelligence Research — that language-model providers deprecate models on short schedules, collect prompts under commercial licenses and alter serving layers while keeping your pinned model ID unchanged. The end customer takes the hit and is forced to reinvest in adapting to another model that’s not smarter but cheaper to run.
The ultimate business model
Anthropic is currently defending against a consolidated consumer class action in California federal court (In re Claude Service Degradation Litigation, No. 3:26-cv-07699-RFL) alleging that the company wrongfully degraded its paid Claude Pro ($20/mo), Max 5x ($100/mo), and Max 20x ($200/mo) tiers between March 4 and May 6, 2026. In April 2026, the problem became so obvious that Anthropic had to publish an anti-crisis investigation into Claude Code complaints where it pointed to three changes they intentionally degraded without touching the model. They lowered default reasoning effort from high to medium, introduced a bug that discarded earlier reasoning after idle sessions, and added a brevity instruction that hurt coding output.
In September 2025, Anthropic acknowledged three infrastructure bugs that corrupted responses and misrouted traffic. One routing error grew from 0.8% of requests to 16% of Sonnet 4 traffic and hit roughly 30% of active Claude Code users. Customers paid full price for degraded service, and transparency arrived only after complaints.
The financial incentives behind this opacity are enormous. OpenAI closed a funding round at an $852 billion valuation and discussed a round above $1.2 trillion, while Anthropic prepared for a public listing above $2 trillion alongside roughly $895 billion in non-cancellable compute commitments. High switching costs and forced migrations serve those valuations because every broken workflow requires more API calls to test, more compute for quality assurance and more tokens to rewrite system prompts so your business is not sued by your customers one day for a change in your product you don’t even control.
In 1970, Nobel laureate George Akerlof showed in The Quarterly Journal of Economics how quality uncertainty creates a “market for lemons.” A 2026 CHI study by Alexander Erlei and colleagues at Göttingen, Cagliari, TU Delft and Columbia proved that this trap hits AI buyers. Without disclosure in a high-lemon market, 92.2% of delegations went to low-quality AI systems, whereas partial accuracy disclosure in a low-lemon market cut lemon delegations to 5.5% and raised earnings by 36.4 coins.
Regulators will not fix this soon. While Serena Booth argued at ICML 2025 that consumer protection laws apply to AI, a 2026 FAccT study by Abeba Birhane, William Agnew and colleagues found 249 instances across 100 news articles where AI firms used 27 capture mechanisms to fight oversight. Meanwhile, a 2026 report by Miles Brundage and 47 co-authors noted that frontier AI gets less independent scrutiny than consumer goods, with baseline AAL-1 audits costing $300,000 to $600,000 and AAL-2 reviews exceeding $1 million.
How to protect your business
Data centers are great for training large models because they concentrate highly interconnected hardware at scale. Once a model is trained, however, is there any specific reason you share trade secrets and intellectual property with a data center?
Tier one superintelligence models run on a domestic intelligence unit that costs less than three months of tokens. For most business workflows — drafting, classification, summarization, contract review and code assistance — an open-weight model on local memory keeps prompts inside your building.
In case you love to share your data for the greater good of pre-IPO corporations, here are a few tips from me, based on nearly two decades of experience in this field:
- You can ask your engineers to hash weights, lock configurations, cut token bills and guarantee that your tool will not lose accuracy overnight.
- Build a ten-prompt regression baseline for workflows tied to revenue or liability. Log outputs, timestamps and error rates across data slices monthly, so you catch silent serving changes before clients do.
- Audit your switching and token costs over a ninety-day window. Count the hours spent fixing broken prompts after updates and compare that bill against a local workstation.
- Demand contractual change notices and data sovereignty from vendors. If a provider will not guarantee advance notice before altering reasoning effort or routing infrastructure, move proprietary workflows to weights you control. Otherwise, it’s only a matter of time before you’re screwed by a hyperscaler. While some Silicon Valley CEOs ask you to give up your device’s ability to think independently, the real question is: are you willing to take action, or will you “follow the prompt”?
Good news: you can copy/paste this article into a vibecoding tool, so it will implement most of the life-saving solutions within 20 minutes.
You will spend some time on research and likely need more than one AI agent to make it work, yet not addressing this issue today might cost you your entire business tomorrow. The ball is in your court. For now.
Key Takeaways
- Renting AI through an API means renting a product that can change without your knowledge or consent.
- Running open-weight models on hardware you control keeps your data, costs and accuracy in your hands.u003cbru003e
As a big believer in a domesticated superintelligence, I can’t pass up the Cheeky Pint podcast where OpenAI CEO suggested it’s a terrible idea to keep processing your personal data on your personal computers because his company’s data centers could process your data more efficiently. He walked the remark back after immediate backlash.
It’s hard to recognize, yet your business likely lives in that world of centralization already. When using inference via API, you often rent access to a black box that shifts its behavior, pricing and reasoning effort without your knowledge or consent.
The leaky bucket you never see
Ask yourself: how many times has your team asked for more budget due to inconsistent behavior from GPT or Claude? In Harvard Data Science Review, Stanford and UC Berkeley researchers Lingjiao Chen, Matei Zaharia and James Zou tested identical prompts on GPT-4 and GPT-3.5 across snapshots three months apart. GPT-4’s prime-number accuracy fell from 84.0% in March 2023 to 51.1% in June, while its chain-of-thought boost collapsed from 24.4% to -0.1%.
Everyday workflows suffered similar drops. On a happy-number task, GPT-4’s accuracy fell from 83.6% to 35.2%, while directly executable code plummeted from 52.0% to 10.0% because extra formatting broke pipelines. Instruction-following on answer extraction collapsed from 99.5% to 0.5% under the same model name and API.
