In Short: Route by Task, Verify by Evaluation
Microsoft Foundry's model catalogue is deliberately plural: Microsoft's MAI family, OpenAI, Anthropic's Claude models and open-source options, all behind the same deployment, safety and observability tooling. That plurality is the point. The organisations getting model choice right are not picking a champion; they are routing tasks to the cheapest model that clears the quality bar, and verifying that bar with evaluations built from their own work, not public leaderboards.
This guide is the selection framework we use, plus the honest state of the catalogue as we see it from client estates.
The Catalogue in Brief
MAI. Microsoft's own family stepped out of OpenAI's shadow with MAI-Thinking-1, the reasoning model served at list pricing around 2 dollars per million input tokens and 8 dollars per million output tokens, with a 256K context window. Its significance is economic as much as technical: it reset the price floor for reasoning-grade work, and our comparison against Claude and GPT-5 covers where it stands on quality. For regulated estates, Frontier Tuning adds reinforcement-learning customisation of MAI models inside the customer's compliance boundary.
OpenAI and Anthropic. The frontier families, available for the work where their particular strengths earn the premium. Notably, Microsoft itself runs Copilot Cowork on Anthropic's Claude models, which settles any question about whether multi-vendor is an officially sanctioned pattern.
Open source. Strong and improving for well-bounded tasks - classification, extraction, summarisation at volume - and for estates with data residency or cost constraints that favour self-managed deployment.
The Framework: Three Questions per Task
1. What does failure cost? Tasks where a wrong output is cheap to catch (draft generation, tagging with human review) tolerate cheaper models. Tasks where errors propagate (agent decisions, customer-facing answers, anything feeding automated action) justify reasoning models and evaluation rigour.
2. What does the token bill look like at volume? Cost per outcome, not per token, is the metric: a cheaper model that needs two retries and a review loses to a dearer one that lands first time. Model the realistic monthly volume before committing; our agent cost guide has the worked arithmetic.
3. How switchable is the choice? Prices and models in this market move quarterly. Keep prompts portable, keep an evaluation set that makes a switch measurable in an afternoon, and treat the model as configuration. The estates hurting today are the ones that welded an architecture to a 2024 model choice.
The Routing Pattern That Keeps Winning
In practice, mature Foundry estates converge on a three-tier routing:
- Volume tier - small, fast, cheap models for classification, extraction, routing and formatting; open source or small MAI/OpenAI variants
- Reasoning tier - MAI-Thinking-1 or peers for multi-step analysis, agent planning and anything where correctness across steps compounds
- Premium tier - frontier models reserved for the tasks where output quality is the product itself
The router is often embarrassingly simple - a task-type lookup - and the savings against a single-premium-model estate are routinely large. The evaluation set is what makes the tiers trustworthy: fifty to a hundred real cases per task type, scored on every candidate, re-run when models or prices change.
What We Tell Clients
Spend your sophistication on evaluation, not selection. The model market will keep moving under you; a good evaluation harness converts that churn from risk into a quarterly cost-reduction opportunity. Start with the routing pattern, benchmark on your own tasks, and let the platform's uniform tooling do what it is for: making the swap boring.
Sources and Further Reading
Frequently asked
The catalogue spans Microsoft's own MAI family (including the reasoning model MAI-Thinking-1), OpenAI models, Anthropic's Claude models and a wide set of open-source options, all behind consistent deployment, safety and observability tooling.
Microsoft's flagship reasoning model, served through Foundry with list pricing around 2 dollars per million input tokens and 8 dollars per million output tokens and a 256K context window. It matters because it reset the price floor for serious reasoning work, making high-volume agent workloads economically viable that were marginal a year earlier.
Standardise the platform, not the model. Route by task: a cheap fast model for classification and extraction, a reasoning model where multi-step correctness pays, a premium model where quality is the product. Single-model estates overpay simple tasks and underserve hard ones.
Build a small evaluation set from your real tasks - fifty to a hundred representative cases with known good outputs - and score candidates on quality, latency and cost per outcome. An afternoon of task-specific evaluation beats any public leaderboard, because leaderboards do not measure your task.
Keep prompts and tools model-agnostic where practical, hold evaluation sets that make switching measurable, and treat model choice as a routing configuration rather than an architecture. Foundry's catalogue makes multi-vendor practical; discipline makes it real.
