Microsoft Drops Anthropic: AI Token Costs Explode Enterprise Budgets

When Microsoft says stop: the wake-up call for all businesses

Microsoft just terminated its internal Anthropic licenses. The reason: in just a few months, the shift to token-based billing caused their annual budgets to explode. If a tech giant with virtually unlimited resources deems these costs unsustainable, what should French and Belgian SMEs and mid-market companies expect?

This decision isn't trivial. It reveals a reality that many executives discover too late: the true cost of generative AI in business bears no resemblance to the advertised rates. Between initial estimates and the final invoice, the gap can reach 300 to 500% depending on usage patterns. And this phenomenon will intensify in 2025-2026.

This article breaks down the mechanics of token-based billing, identifies the hidden costs no one mentions, and offers concrete strategies to control your AI budget without sacrificing performance.

Understanding token-based billing: the usage pricing trap

What is a token and how are they counted?

A token represents approximately 0.75 words in French. Each interaction with an AI model like GPT-4, Anthropic's Claude, or Google's Gemini consumes input tokens (your request) and output tokens (the generated response). Billing distinguishes between these two flows, with different rates.

Here are the average rates observed in 2025 for premium models:

• GPT-4 Turbo: $10 per million input tokens, $30 output
• Claude 3 Opus: $15 per million input, $75 output
• Gemini Ultra: $12.50 per million input, $37.50 output

These figures seem modest. They become staggering at organizational scale.

The multiplier effect nobody anticipates

A simple conversation with an AI assistant consumes between 1,000 and 4,000 tokens. But professional use cases involve much heavier contexts: reference documents, conversation history, detailed system instructions. A single business query can reach 50,000 to 100,000 tokens.

Let's take a concrete example. A team of 20 salespeople uses an AI assistant to draft commercial proposals. Each proposal requires:

• Client context (history, expressed needs): 15,000 tokens
• Internal instructions and templates: 8,000 tokens
• Proposal generation: 12,000 output tokens

That's 35,000 tokens per proposal. With 10 proposals per salesperson per week, the team consumes 7 million tokens weekly. Over a year: 364 million tokens, representing between EUR 15,000 and 40,000 depending on the model used. For a single use case.

The five hidden costs of enterprise AI

1. Prompt drift: when users optimize for quality, not cost

Employees naturally learn to get better responses. How? By providing more context, requesting more detailed answers, following up to refine results. Each quality improvement translates to increased token consumption.

At AISOS, we observe that average consumption per user increases by 15 to 25% each month during the first six months of deployment. Without caps, initial budgets become obsolete within a quarter.

2. System tokens: the invisible tax

Every API call includes system instructions that define AI behavior. These instructions are billed with each request, even if they never change. A 2,000-token system prompt repeated 10,000 times daily represents 20 million monthly tokens: between EUR 200 and 600 per month for text that nobody reads.

3. Failures and retries: paying for what doesn't work

AI models don't always succeed on the first try. Format errors, incomplete responses, timeouts: each failure consumes tokens. Robust architectures include automatic retry mechanisms. Result: 10 to 20% additional consumption to handle edge cases.

4. Model versioning: planned obsolescence

Providers update their models regularly. Each new version can modify behaviors, requiring prompt adjustments and testing phases. These iterations consume tokens without producing direct value. The most active companies can dedicate 5 to 10% of their annual budget to this.

5. The dependency effect: when switching becomes impossible

Once your workflows are built around a specific model, migrating to a less expensive alternative means rewriting prompts, retesting use cases, training teams. This migration cost strengthens the original provider's negotiating power. Price increases become difficult to contest.

Why Microsoft sounded the alarm

The Microsoft case illustrates a systemic phenomenon. According to available information, the company found that its internal teams had consumed the equivalent of their entire annual Anthropic budget in just a few months.

Several factors explain this drift:

• Viral adoption: once access is open, usage multiplies exponentially
• Lack of governance: without quotas or monitoring, no regulation mechanism exists
• Quality of Anthropic models: Claude produces long, detailed responses, therefore expensive ones
• Unforeseen usage: teams invent applications not anticipated in the initial budget

Microsoft isn't abandoning AI. The company is rationalizing its investments by favoring its own models via Azure OpenAI, where it has better control over costs and margins. This decision is strategic, not defeatist.

AI budget optimization strategies for SMEs and mid-market companies

Implement usage governance from the start

Before any deployment, clearly define:

• Authorized use cases and their business priority
• Quotas per team, per project, per user
• Overage alerts (50%, 80%, 100% of budget)
• The arbitration process when limits are reached

This governance isn't bureaucratic constraint. It's the condition for transforming AI into a controlled investment rather than a financial sinkhole.

Choose the right model for each task

Not all use cases require GPT-4 or Claude Opus. A simple classification can reduce costs by 40 to 70%:

• Simple tasks (classification, extraction, short reformulation): lightweight models like GPT-3.5 or Claude Haiku
• Intermediate tasks (standard writing, document analysis): mid-range models like GPT-4 Turbo or Claude Sonnet
• Complex tasks (advanced reasoning, original creation, critical decisions): premium models

AISOS audits reveal that 60 to 75% of enterprise queries can be handled by mid-range models without perceptible quality loss.

Optimize prompts to reduce consumption

Every token counts. Simple techniques can reduce consumption by 20 to 40%:

• Compress context: summarize documents rather than including them entirely
• Limit response length: specify maximum word count or bullet points
• Use structured formats: JSON or lists rather than prose
• Cache system instructions: some providers offer this option

Negotiate contracts adapted to your reality

Public rates are starting points, not final prices. Beyond a certain volume, negotiate:

• Tiered pricing with volume discounts
• Volume commitments with associated rebates
• Monthly or quarterly billing caps
• Test credits for development phases

Companies that negotiate regularly achieve 15 to 30% reductions compared to standard rates.

Consider open source and self-hosted alternatives

For non-critical use cases or very high volumes, open source models like Llama 3, Mistral, or Falcon offer comparable performance at virtually zero marginal cost once infrastructure is deployed.

The economic calculation becomes favorable when:

• Your consumption exceeds 50 million tokens per month
• Your use cases are stable and well-defined
• You have technical skills to operate the infrastructure
• Confidentiality requirements justify internal hosting

Anticipating cost evolution in 2025-2026

Current trends paint a contrasted landscape for the coming years:

Factors driving unit cost reduction:

• Improved model efficiency (more performant with fewer parameters)
• Increased competition between providers
• Computing infrastructure optimization

Factors driving total cost increase:

• Multiplication of use cases within organizations
• Increasing complexity of tasks assigned to AI
• AI integration into critical processes (therefore difficult to reduce)

The most realistic projection: unit costs will decrease by 20 to 30%, but volumes will increase by 100 to 200%. Companies' overall AI budgets will continue to grow, but more predictably if best practices are in place.

Transforming budget constraints into competitive advantage

The Microsoft case isn't a defeat for enterprise AI. It's a signal of maturity. Organizations that survive the euphoria phase will be those that have learned to measure, optimize, and prioritize their AI investments.

For SMEs and mid-market companies, this discipline is even more crucial as budget flexibility is limited. But it's also an opportunity: a mid-sized company that masters its AI costs can deploy use cases that competitors will deem too expensive.

Three priority actions to launch this week:

• Audit your current token consumption and identify the three most expensive areas
• Implement a monitoring dashboard with overage alerts
• Evaluate whether your premium use cases truly justify premium models

Mastering AI token costs is no longer a technical subject reserved for IT teams. It's a general management issue, on par with payroll or procurement. Leaders who integrate it into their financial management now will gain a decisive head start.