Microsoft: AI Costs More Than Human Labor — The Cost Problem Is Now Real
Microsoft told Fortune that internal data shows the cost of using AI agents now exceeds the equivalent human labor costs. The report has drawn significant attention, gathering 74 points on Hacker News.
When it comes to AI profitability, major players like OpenAI and Anthropic are all posting losses, and Microsoft's statement has underscored the reality that AI is not a silver bullet for cost reduction. While token prices keep falling, agentic workflows involve multiple API calls, tool usage, and iterative processing, so actual operating costs often far exceed expectations.
This issue is becoming a major factor that makes ROI calculations difficult, particularly during the enterprise AI adoption phase.
From Model Labs to Agent Labs — An Industry-Wide Strategic Shift
OpenAI's Greg Brockman has stated plainly that "the model alone is no longer the product." This marks a major policy reversal from the company's previous stance of prioritizing model development above all else. Among major model developers, including the former head of OpenAI Labs, the consensus used to be that a good enough model would win.
In reality, however, winning as a product now requires a combination of model + harness + workflow + UI + memory + economics.
This trend is spreading to other companies as well.
AI21** has shut down its model team and fully pivoted to agent development
DeepSeek** has launched a "harness team" for the first time
Cursor** has released an SDK, making it possible to build custom agents
There is a caveat, though. If models come to be trained in tight coupling with their own proprietary agents, models could effectively become even more closed.
The MCP Protocol Goes Stateless — A Major Infrastructure Update
A new release candidate for the Model Context Protocol (MCP), "MCP 2026-07-28 RC," has been announced. The biggest change is that the protocol is going stateless.
Previous versions of MCP required handshake and session ID management, but the new specification eliminates both. With a design in which any request can hit any server instance, the challenges of scaling, load balancing, and sticky sessions are dramatically simplified.
The release also includes first-class extensions called MCP Apps and Tasks, strengthened authentication, and clarified deprecation policies. For infrastructure teams, this update represents a major operational change.
DeepSeek V4-Pro: 75% Price Cut Made Permanent
DeepSeek has officially made its 75% discount on V4-Pro a permanent measure. This puts inference costs at $0.435/M for input, $0.87/M for output, and $0.0036/M for cached input, reaching a blended rate of about $0.18/M.
External analyses estimate that running the Intelligence Index on V4 Pro costs roughly one-third of Gemini 3.1 Pro Preview, one-twelfth of GPT-5.5, and one-nineteenth of Claude Opus 4.7. DeepSeek's strategy of making intelligence "unmeasurably cheap" is having a decisive impact on the price war over LLM inference.
Agent Distillation: Inference Costs Cut to 1/100
A paper demonstrating the practical economic benefits of distilling entire agent workflows into model weights is attracting attention.
By compiling complete agent workflows — including multi-step API calls, tool usage, scratchpads, and decision structures — into the weights of a smaller model, inference costs can reportedly be cut to roughly 1/100 while nearly maintaining frontier quality.
Considered alongside Microsoft's AI cost problem described above, reducing agent runtime costs is an urgent priority, and this approach is expected to serve as a practical solution.
---