Choosing an LLM vendor in 2026 is less like buying software and more like choosing a utility provider whose prices, products, and terms change quarterly. OpenAI now reports enterprise as more than 40% of its revenue; Anthropic reports over 1,000 business customers spending more than $1 million annualized. The vendors are ready for the enterprise. The harder question is whether your evaluation process is ready for the vendors.
Start With Workload Classes, Not Vendors
Before any bake-off, classify your workloads:
- High-volume, low-complexity (classification, extraction, routing) — optimize for price; small and open-weight models often win.
- Agentic and coding workloads — optimize for tool-use reliability and long-context behavior.
- Customer-facing generation — optimize for tone control, safety behavior, and latency.
- Regulated-data workloads — optimize for deployment options (VPC, on-prem, sovereign hosting).
Most enterprises discover they need two or three vendors, not one. Design for that from day one.
The Six Evaluation Dimensions
1. Capability Fit — Test on Your Tasks
Public benchmarks are marketing. Build a 100–300 item eval set from your real workload and run every candidate on it. This is the single highest-leverage step in the process.
2. Pricing Structure and Trajectory
Look past the sticker price to the structure: cached-input discounts, batch pricing, and rate-limit tiers change effective cost by multiples. Note that prices move fast — and not always down at launch: Anthropic's Sonnet 5 launched at introductory pricing ($2/$10 per million tokens) that stepped up after an initial window. Model your 12-month cost under announced pricing, not launch promos.
3. Data Terms and Training Rights
Verify: no training on your data by default, clear retention windows, and export controls. OpenAI, Anthropic, and Google all offer enterprise no-training defaults — but terms differ across API vs. consumer-grade products under the same brand.
4. Deprecation and Migration Risk
Model retirements are now routine and fast: DeepSeek fully retired legacy model IDs with hard cutoffs in 2026; xAI rebills deprecated slugs at new-model rates; Cohere has retired embedding models with breaking failures. Ask every vendor for their deprecation policy in writing, and budget engineering time for at least one forced migration per year per vendor.
5. Reliability and Concentration
Every major provider — including Anthropic and OpenAI — logged multi-model or majority-traffic incidents in the past year. Ask for status-page history, not SLA marketing. If the workload is revenue-critical, you need a tested failover path to a second provider (OpenAI-compatible APIs from DeepSeek, xAI, and others make this cheaper than it used to be).
6. Deployment Surface
Where can the model actually run? Hyperscaler marketplaces (Bedrock, Foundry, Vertex) simplify procurement and colocate models with your data. Private deployment matters for regulated work: Cohere's Command A+ runs on as few as one or two GPUs, and Mistral splits its control plane from customer-run workers so data stays in the customer perimeter.
A Simple Scoring Rubric
Weight the six dimensions for each workload class (capability 30%, price 20%, data terms 15%, deprecation 10%, reliability 15%, deployment 10% is a reasonable default), score candidates 1–5 from evidence, and rerun the exercise every two quarters. The rerun is not optional — this market reprices too fast for annual reviews.
The Multi-Vendor Baseline
The stable end-state for most enterprises: one frontier vendor for complex/agentic work, one low-cost vendor (often open-weight) for volume work, and an abstraction layer (gateway or router) so switching is a config change, not a rewrite.