Copilot goes cheap as Microsoft phases out OpenAI and Anthropic models to cut costs
Microsoft is actively replacing third-party models from OpenAI and Anthropic with its in-house MAI models across Copilot products like Excel and Outlook to significantly reduce operational costs. The transition involves a potential shift in pricing strategy, where cheaper, less capable in-house models become the default, while premium third-party models may incur surcharges. Despite marketing claims of using "clean, commercially licensed data," technical disclosures reveal Microsoft utilizes the
Analysis
TL;DR
- Microsoft is actively replacing third-party models from OpenAI and Anthropic with its in-house MAI models across Copilot products like Excel and Outlook to significantly reduce operational costs.
- The transition involves a potential shift in pricing strategy, where cheaper, less capable in-house models become the default, while premium third-party models may incur surcharges.
- Despite marketing claims of using "clean, commercially licensed data," technical disclosures reveal Microsoft utilizes the legally ambiguous Common Crawl dataset, mirroring industry-standard practices.
- Performance benchmarks indicate Microsoft's new reasoning model, MAI-Thinking 1, trails behind competitors like Sonnet 4.6 and Opus 4.6, performing closer to older models like Deepseek V3.2.
Why It Matters
This development signals a critical inflection point in the AI industry where major tech giants prioritize cost efficiency and vertical integration over reliance on specialized AI vendors. It highlights the growing tension between reducing infrastructure expenses and maintaining high-quality user experiences, potentially forcing a re-evaluation of subscription value propositions for enterprise and consumer users alike.
Technical Details
- Model Integration: Microsoft’s proprietary MAI models are currently processing tens of thousands of requests weekly in Excel and Outlook, with additional deployment planned for GitHub Copilot and a proprietary transcription model for Teams.
- Performance Gap: While Microsoft claims parity with Sonnet 4.6 and Opus 4.6 in coding tasks based on human evaluation, independent benchmarks show MAI-Thinking 1 lags significantly, aligning more closely with Deepseek V3.2.
- Data Provenance: Contrary to public statements about exclusively using clean commercial data, the technical paper confirms the use of Common Crawl, a publicly available web dataset with unresolved legal status regarding AI training rights.
- Pricing Architecture: The underlying technical strategy supports a tiered service model, likely separating basic functionality (powered by MAI) from advanced capabilities (powered by OpenAI/Anthropic) to facilitate usage-based billing adjustments.
Industry Insight
- Cost-Driven Consolidation: Expect other large enterprises to accelerate the development of internal LLMs to mitigate rising API costs, leading to a bifurcation in the market between high-end specialized models and cost-effective generalist in-house solutions.
- Value Proposition Risks: Consumers and businesses may face "sticker shock" if subscription prices remain static while model quality degrades; transparency regarding model swaps and performance tiers will become a key differentiator for trust.
- Legal Ambiguity in Training Data: The discrepancy between marketing claims and actual data sourcing (Common Crawl) underscores the need for stricter regulatory frameworks and clearer disclosure standards regarding AI training data provenance and licensing.
Disclaimer: The above content is generated by AI and is for reference only.