AI Vendor Lock-in Leaderboard — May 2026
How hard is it to leave each AI vendor? Ten providers scored 0-5 across the seven dimensions of our Model Exit-Cost Audit. Lower total = easier to switch out. Updated monthly — current snapshot May 2026.
The 7-Dimension Exit-Cost Audit
Lock-in is not a single thing. It compounds across prompt portability (will my prompts re-translate?), tool/function-calling schema (does my agent code re-target?), output format drift (will my downstream parsers break?), latency profile match (will my SLOs hold?), eval re-run cost (do I have to redo the entire QA pass?), compliance re-certification (does the new vendor force a SOC 2 / DPA refresh?), and behavioural drift (does the answer quality vary in production?). The lower the total score, the cheaper it is to switch. Read the full framework in our Model Exit-Cost Audit writeup.
| Rank | Vendor | Total /35 | Prompt | Tool/Function | Output | Latency | Eval | Compliance | Behavioural |
|---|---|---|---|---|---|---|---|---|---|
| #1 | OpenAI Direct Proprietary tool-calling schema (function_calling) and persistent quirks in JSON output handling. Highest exit cost in the closed-frontier tier. | 26 | 4 | 4 | 4 | 3 | 4 | 3 | 4 |
| #2 | Azure OpenAI Single-provider gateway (OpenAI only). Same exit cost as OpenAI Direct plus Azure IAM/networking coupling on top. | 26 | 4 | 4 | 3 | 3 | 4 | 4 | 4 |
| #3 | Anthropic Direct Distinct prompt structure (XML-tagged), proprietary tool-calling, and the new Opus 4.7 tokenizer adds ~35% drift on translated prompts. | 24 | 3 | 4 | 3 | 3 | 4 | 3 | 4 |
| #4 | Google Direct (Vertex) Vertex-only IAM, GCP egress fees, and Google-specific safety filter behaviour. Strong compliance posture but the IAM coupling is sticky. | 24 | 3 | 4 | 3 | 3 | 4 | 4 | 3 |
| #5 | AWS Bedrock Multi-provider but with AWS IAM/VPC/PrivateLink coupling and Converse-API normalization layer. Strong compliance, weaker provider portability. | 21 | 3 | 3 | 3 | 2 | 3 | 4 | 3 |
| #6 | Together AI Open-weights focus; OpenAI-compatible. Easy to leave per provider, but model availability shifts as Together rotates its hosted catalog. | 14 | 2 | 2 | 2 | 2 | 2 | 2 | 2 |
| #7 | OpenRouter OpenAI-compatible API; switching providers is a one-line config change. Limited enterprise procurement features (PO/MSA/DPA) on smaller plans. | 13 | 1 | 2 | 2 | 2 | 2 | 2 | 2 |
| #8 | Swfte Connect Provider-agnostic abstraction designed to drop 4 of 7 dimensions to near-zero. OpenAI-compatible inbound API; pluggable outbound providers. | 12 | 1 | 1 | 2 | 2 | 2 | 2 | 2 |
| #9 | DeepSeek (self-host) Apache 2.0 weights. Switching is a model file swap. Operational cost dominates; vendor lock-in is functionally zero. | 7 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
| #10 | Gemma 4 (self-host) Apache 2.0. Identical exit profile to other self-hosted open weights — operational cost is the primary concern, not lock-in. | 7 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
How to read the lock-in leaderboard
0-7: Effectively zero lock-in. Self-hosted open weights or thin OpenAI-compatible aggregators. Switching cost is configuration, not engineering.
8-17: Manageable lock-in. Provider-agnostic gateways and well-designed multi-cloud platforms. Expect a one-quarter migration window if you have to switch.
18-24: Material lock-in. Single-provider gateways and managed platforms with proprietary surfaces. Plan for two quarters and a meaningful re-architecture if you switch.
25-35: Severe lock-in. Direct closed-frontier providers with proprietary tool schemas and tokenizers. Expect three quarters or more, plus a full eval and compliance re-run. Frequently the biggest unbudgeted cost on AI roadmaps.