Microsoft cut the input price of MAI Code 2 to $0.50 per million Tokens, while DeepSeek withdrew its per-call fee for web access on the same day after only 18 days in effect. We believe cost competition among developers has entered a “task-by-task accounting” phase: cheaper models are only the starting point, while search, context length, and routing strategies now determine the real bill.

What this is

Microsoft introduced MAI Code 2 at Ignite 2026 and integrated it with GitHub Copilot and Microsoft Foundry. It supports a total context of 1.05 million Tokens, with input pricing halved to $0.50 per million Tokens and output priced at $2.50. The half-price promotion lasts until May 17, 2027.

DeepSeek, meanwhile, removed the surcharge for automatically triggering web search and page retrieval in model conversations. Standalone calls to its Search or Fetch interfaces still cost RMB 5 per 1,000 successful requests. It is important to note that retrieved page text still counts toward input Tokens, so “free search” does not mean “free retrieval.”

Industry view

These adjustments show that vendors are beginning to price against actual production bills rather than comparing benchmark rankings alone. The article’s author proposes splitting workloads as follows: assign roughly 80% of code reading, modification, and testing to MAI Code 2; use DeepSeek’s built-in search for real-time retrieval; and reserve a small share of complex tasks for high-reasoning models. In this developer’s sample, the monthly bill fell from $200–$350 to under $40. However, this is a single developer’s usage sample and cannot be extrapolated directly to every team.

We also see a risk: the half-price period has a clearly defined end date, and web scraping will continue to increase context Token consumption. If companies design their architectures solely around today’s low prices, costs could rise again after May 2027. Moving every task into a low-cost model may also increase the quality, latency, and retry costs of complex code. Lower prices do not automatically offset growth in total usage.

Impact on regular people

For enterprise IT: Procurement teams should ask less about “Which model is strongest?” and focus more on the completion cost of each task category. Contracts should also specify separate pricing for input, output, search, and retrieval, preventing high-frequency context consumption from erasing the benefits of low-cost entry points.

For individual professionals: Developers will increasingly coordinate multiple models, making everyday code changes and information lookups cheaper. Whether this actually reduces spending still depends on limiting web results, compressing context, and controlling failed retries.

For the consumer market: Falling model invocation costs could accelerate the adoption of coding assistants, real-time search, and automation tools among small and midsize software businesses. Consumers may see less dramatic price cuts than more generous free tiers and additional features, while intensive services will still impose usage limits.