Switching AI Models in Production Isn't Routing — Only 3 Parameters Are Spelled the Same
Stripe acquired OpenRouter for over $7 billion in August 2026, valuing the AI gateway's metering and settlement capabilities far above its $1.3B valuation just three months prior Only 36% of AI request-level capabilities are shared across the four major provider SDKs (OpenAI, Anthropic, Google, Mistral), with just 3 field names spelled identically across all four Model portability is largely a myth for production workloads that use advanced features like reasoning budgets, prompt caching, and st
Analysis
TL;DR
- Stripe acquired OpenRouter for over $7 billion in August 2026, valuing the AI gateway's metering and settlement capabilities far above its $1.3B valuation just three months prior
- Only 36% of AI request-level capabilities are shared across the four major provider SDKs (OpenAI, Anthropic, Google, Mistral), with just 3 field names spelled identically across all four
- Model portability is largely a myth for production workloads that use advanced features like reasoning budgets, prompt caching, and structured output dialects
- Routers deliver genuine value in failover routing, spend telemetry, and procurement simplification, but cannot abstract away vendor-specific behavior differences
- Cross-vendor swaps silently downgrade output quality without errors, as provider-specific parameters are dropped or defaulted rather than translated
Why It Matters
This analysis exposes a critical gap between the AI gateway industry's marketing promise of "model portability" and the technical reality of fragmented provider APIs. For AI practitioners building production systems, it means architecture decisions around model switching require far more careful consideration than a simple config change, and teams should budget for the significant engineering effort needed to maintain consistent behavior across providers.
Technical Details
- Measured four official SDKs as of August 19, 2026: openai@7.5.0 (37 raw fields, 30 canonical capabilities), @anthropic-ai/sdk@0.118.0 (19 fields, 19 capabilities), @google/genai@2.17.1 (37 fields, 27 capabilities), @mistralai/mistralai@2.6.3 (23 fields, 22 capabilities)
- Union across all four SDKs contains 88 distinct raw field names mapping to 36 canonical capabilities, with only 13 (36.1%) present universally
- Three top-level fields are identically spelled across all providers: model, temperature, tools; every other parameter requires manual translation
- Structured output counts as universal but differs significantly in JSON Schema keyword support, enforcement timing, and violation handling across providers
- OpenAI's newer responses endpoint diverges substantially from its legacy chat.completions endpoint, indicating even same-vendor portability is unreliable
Industry Insight
- The $7B acquisition price reflects Stripe's strategic bet on AI spend telemetry and settlement infrastructure, not model portability—buyers should evaluate gateways for procurement and observability value rather than abstraction promises
- Teams should deliberately build for the 36% intersection of provider capabilities when designing for portability, and explicitly price the engineering cost of handling the remaining 64% of vendor-specific features
- The convergence trend toward common shapes (tools, tool_choice) suggests standardization will emerge organically through aggregator pressure, but this process is slow and incomplete for frontier model features
Disclaimer: The above content is generated by AI and is for reference only.