Running autonomous agents around the clock can quickly drain enterprise budgets when token spend escalates. To address these expenses, MAI has launched a lineup of seven new models, aiming to give organizations lower-cost options for continuous operational workloads.

According to MAI, enterprise customers need accessible pricing to run agents doing real engineering work without runaway bills. By moving beyond standard cloud API access to licensed weights, MAI is altering the unit economics of autonomous workflows. For companies scaling continuous automation, the question remains whether these architectures can reliably curb escalating token bills.

Benchmark Performance and Cost Disparities

The pricing gap appears most clearly in dedicated operational tasks. In automated security testing, MAI Cyber 2 Flash beats GPT-6 Astra’s quality at less than 2% of the cost per task on the CyberGymE2E benchmark.

"MAI Cyber 2 Flash beats GPT-6 Astra’s quality at less than 2% of the cost per task on the CyberGymE2E benchmark."

This gap reshapes expectations for high-frequency defense tasks. Within the MDASH platform, MAI Cyber already accounts for almost 40% of scans, and the latest MAI Cyber 2 Flash model is priced 86% lower on output tokens than MAI Cyber 1.1 Flash. The engineering workload models follow a similar pattern: MAI Code 2 reaches 98% of the coding quality of GPT-6.1 Sol at about 40% of the price, and it matches 96% performance for just 5% of the price. In routine work, MAI Code 2 Flash matches the performance of GPT-6 Luna at 65% of the price and Haiku 5.5 at 20% of the price. In a blind side-by-side evaluation of frontend tasks, developers preferred MAI Code 2 over GPT-6 Sol 64% of the time, although MAI explicitly notes that these models are not yet tuned for fundamental science research or very long horizon agentic work.

Hardware Efficiency and Enterprise Workloads

These pricing reductions connect directly to underlying infrastructure and task routing. Microsoft confirmed that Maia 200, its second-generation AI accelerator, is now in production running our largest models, achieving 1.7x more tokens per watt to reduce data center power demands.

"Maia 200, Microsoft’s second-generation AI accelerator, is now in production running our largest models, achieving 1.7x more tokens per watt."

Model routing also shifts the financial equation. The Decision 1.1 model, the latest decision model from Microsoft, is 50x faster and 79x cheaper than GPT-Sol 6.1. In creative and audio domains, MAI Image 2.7 holds the top spot on Artificial Analysis image editing while outperforming GPT Image 2.6 Sunburst (max) at one fifth of the price, supported by MAI Image 2.7 Flash at ~$10 per 1,000 images. For speech, MAI Realtime 1 outperforms GPT-Realtime 2.1 at rates starting at $0.01 per audio minute. Clinical applications show similar scale, as Dragon Copilot Transcribe, built on MAI Transcribe 2, covers 10 million patient encounters a month and is in the hands of a third of US physicians.

A central concern for organizations remains whether licensed model weights provide sufficient containment and control when managing autonomous agents across sensitive enterprise environments. MAI maintains that dedicated weights offer the direct oversight that organizations need to prevent operational drift while reducing inference overhead.

The economics of enterprise AI are shifting from raw capability benchmarks to operational sustainability. By paring back prices on routine tasks and backing inference with Maia 200 silicon, MAI is making continuous autonomous workloads viable for production budgets.

The catch

Corporate budgets get relief from runaway token bills as MAI shifts to licensed model weights for continuous operational workloads. Custom Maia 200 silicon cuts backend energy expenditures, but enterprises adopting local weights now bear the risk of managing operational drift on their own. The inference standard got radically cheaper for routine tasks, leaving enterprises to handle the fact that these budget models cannot plan far ahead.

MicrosoftModel InferenceInference CostEnterprise LLMsAI Infrastructure