Claude's 15-Hour Lockout Revealed the Cost of Relying on a Single AI Provider

We quickly grow accustomed to cheap machine intelligence. But we pay for it with more than tokens: we also surrender our ability to keep working when something breaks.

What happened to Belo

In April 2026, fintech company Belo lost access to Claude for more than 60 employees. The notice cited a violation of the terms of use without providing details. Access was restored about 15 hours later, and the lockout was described as a likely false positive. This account is based on the company's public statements and press coverage, not an independent investigation.

The company had backup access to Gemini, but it could not switch quickly: accumulated history, familiar tools, and integrations were tied to a single provider. Fifteen hours is little in the history of technology, but for a company whose processes have stopped, it is tangible downtime.

Now imagine that first-line support has already been replaced by agents, part of accounting has been automated, and coding agents write the software. An outage in that scenario would be far more painful.

Tokens get cheaper while tasks can get more expensive

Prices really are falling. According to the Stanford AI Index 2025, the price of one million tokens at GPT-3.5-level MMLU performance fell from $20 in November 2022 to $0.07 in October 2024. Claude Opus pricing declined from $15/$75 at the launch of Claude 3 Opus to $4/$20 for Opus 5.5. But these are different models and generations, so they cannot be compared directly.

An agent works differently from a chat interface. It reads files, retains context, calls tools, checks its work, and sometimes revisits the same error several times. If a token becomes four times cheaper but a task consumes eight times as many tokens, the final cost doubles. This is an illustrative calculation, not a forecast. In June 2026, Gartner warned that by 2028, spending on AI coding could exceed the average salary of a developer, though that is specifically a forecast.

Subscriptions carry the same uncertainty. Anthropic introduced weekly quotas in 2025 and doubled five-hour limits for some plans in May 2026. Terms change in both directions, making it risky to build multi-year economics around today's limits.

Chinese and open-weight models: an alternative, not a panacea

DeepSeek V4 Pro costs $0.66 per million input tokens and $1.98 per million output tokens off peak; prices double during peak hours. That does not mean it will perform the same work more cheaply: token usage, rework, and quality vary. DeepSeek also discontinued off-peak discounts in 2025, so prices are not fixed forever.

On October 29, 2026, Mistral released the weights for Mistral Large 4: around three trillion parameters, with approximately 52 billion active per request. The model can be self-hosted, but doing so requires accelerators, memory, power, cooling, and specialists. Its license adds conditions for organizations that offer the model as a service. Independence comes not from the word "open," but from a tested ability to keep operating at an acceptable cost.

People as part of the infrastructure

According to Challenger, Gray & Christmas, from January through September 2026, US employers cited AI as the reason for around 120,000 layoffs, approximately 21% of the total. These are employers' stated reasons, not proof that every position was replaced. Klarna reported that its AI assistant handled 80% of chats and generated around $59 million in annual savings. All estimates were provided by the company itself.

The risk is that, after several years, dismissed specialists will move to other employers or change professions. If they return, they will ask an uncomfortable question: what will stop the company from dismissing them again when the next cheaper model arrives? This is a risk scenario, not a universal outcome.

What to do in practice

• Track three separate metrics: token cost, cost per successful agent action, and the total cost of the process.
• Maintain a tested fallback path: a second API, an open model, or a combination of both. Run real tasks through it periodically.
• Do not tie integrations and prompts to a single provider; store history and context in a portable format.
• Do not eliminate all process expertise: retain people who understand how the work gets done without an agent.

The central idea is simple: automation is cost-effective as long as you can turn it off. If you cannot, it is no longer savings but dependency.