Summary
Highlights
Microsoft is increasing prices for Copilot Pro and Copilot Pro Plus subscribers starting June 1st, 2026. Model multipliers, which determine the cost of using AI models within Copilot, will see significant increases, with some rising fivefold. This means users will pay substantially more for models like Cloud Opus 4.5 and 4.7, potentially reflecting the true API costs of these advanced language models.
The price increases have significant implications for developers using agentic workflows, especially those involving longer-running agents with extensive memory. These workflows, which involve multi-step interactions with the AI model, can quickly accumulate costs due to token usage. The speaker shares an anecdote of two prompts costing $27.12, highlighting the rapid escalation of expenses. Solutions like 'agent code' are emerging, where agents write functions that call APIs, consolidating multiple calls into one to save context and cost, though this introduces security and stability concerns.
Two main narratives explain the price hikes: first, that AI inference was heavily subsidized by companies like Anthropic and venture capitalists, and now users are seeing the true, expensive cost. Second, that companies like OpenAI and Anthropic are preparing for IPOs and need to show strong financial numbers. Additionally, high demand for AI amidst limited capacity in data centers contributes to the price surge. The speaker believes there's truth in both, noting that while preparing for IPOs, the training and maintenance of these complex models are also inherently very expensive.
The current widespread, often undisciplined use of AI tools, where users prompt models for basic tasks like Git commands instead of performing them manually, leads to inefficient use of software and hardware resources. This wastefulness is amplified by the fact that AI services lack the vendor lock-in seen in traditional cloud providers like AWS. Developers can switch AI models or even self-host open-source alternatives much more easily, which could lead to companies migrating to cheaper solutions if prices remain high.
As AI model costs rise, companies will seek developers who can minimize token usage and apply AI strategically. The speaker suggests a return to emphasizing strong technical skills rather than over-reliance on AI for every task. The idea that one engineer with AI can replace two is challenged, as AI-generated code often requires extensive human review to prevent 'quality drift' and technical debt. The speaker highlights that while AI can generate code, human judgment and oversight remain crucial for quality and maintainability.
The speaker criticizes the concept of 'vibe coding,' where developers generate applications with minimal effort using AI. While AI can produce visually appealing apps, the underlying code often contains 'code smells' and is difficult to review due to its sheer volume. This leads to a disconnect between the apparent productivity and the actual quality and maintainability of the software. The speaker observes that many AI-generated products, particularly from solopreneurs, struggle in the market because they lack innovation and are often inferior copies of existing solutions, making their high AI compute costs unsustainable.
The speaker argues that the belief that AI will replace developers is a 'wet dream' of Silicon Valley. True software development requires technical creativity and collaborative problem-solving, not just churning out code. The marketing strategies of some AI companies, like Anthropic, are criticized for alienating developers by suggesting their obsolescence. This could lead developers to turn away from such tools. Despite the current 'complicated market,' developers are advised to stay focused on improving their core skills and not succumb to anxiety driven by misleading AI narratives, as the industry will eventually stabilize.