Microsoft switched Copilot Cowork to usage-based pricing in June 2026, replacing unlimited tokens with per-seat metering. The company adopted DeepSeek V4 as a cost-effective open-source alternative after internal costs for unlimited inference proved unsustainable per Pandaily.
Key facts
- Copilot Cowork shifted from $30/seat flat rate to per-token pricing in June 2026
- DeepSeek V4 achieves 500K context with 90% less KV cache using FlashMemory
- Microsoft committed over $50B in AI infrastructure by 2026
- DeepSeek raised $7.4B at $50B valuation in first external round
- DeepSeek V4 costs $0.002 per 1K input tokens, $0.008 per 1K output
Microsoft's Copilot Cowork, the enterprise AI assistant embedded across M365, now charges per token consumed rather than a flat seat fee. The change, effective June 2026, follows months of escalating inference costs that made unlimited pricing untenable for Microsoft's largest customers.
Key Takeaways
- Microsoft switched Copilot Cowork to usage-based pricing, adopting DeepSeek V4 to cut inference costs by ~40%.
- The move breaks Microsoft's exclusive reliance on OpenAI for first-party AI.
Why DeepSeek V4 Won the Bid
DeepSeek V4 entered production at Microsoft after beating proprietary models on cost-per-token by an order of magnitude. The model achieves 500K context with 90% less KV cache using FlashMemory, per DeepSeek's June 9 technical disclosure. Microsoft's infrastructure commitments — over $50B by 2026, disclosed June 15 — made efficiency non-negotiable.
DeepSeek's open-weight license lets Microsoft self-host and fine-tune without per-token royalties, a structure impossible with OpenAI's GPT-4 or Anthropic's Claude. The move also reduces Microsoft's dependence on OpenAI, its largest investment at $13B+, according to our knowledge graph.
The Pricing Shift
Copilot Cowork now meters by token volume, with enterprise tiers starting at $0.002 per 1K input tokens and $0.008 per 1K output tokens for DeepSeek V4. Prior unlimited plans cost $30 per seat per month. For a 10,000-seat deployment generating 100M tokens daily, the new model cuts costs by roughly 40%, per Microsoft's internal estimates shared with Pandaily.
Microsoft's move mirrors a broader industry pivot from flat-rate to consumption-based AI pricing. OpenAI introduced usage tiers in March 2026; Anthropic followed in May.
Competitive Implications
DeepSeek's selection marks the first time Microsoft has deployed a non-OpenAI model as the default inference engine for a first-party product. The lab raised $7.4B at a $50B valuation on June 18, its first external round led by founder Liang Wenfeng's $2.9B contribution.
Microsoft continues to invest in its own reasoning model, MAI-Thinking-1, unveiled June 3 with 35B active parameters and 97% on AIME 2025. But for cost-sensitive enterprise workloads, DeepSeek V4 now carries the load.
What to watch
Watch for Microsoft's Q4 FY2026 earnings call (expected late July) where Copilot Cowork revenue mix and customer churn from the pricing change will be disclosed. Also monitor whether DeepSeek V4 expands to Azure OpenAI Service as a first-party offering.
Source: pandaily.com
