Cross-Provider Arbitrage Playbook (2026)
Same-tier models from different providers price differently by 2–5×. Arbitrage is real money — if your architecture can capture it.
The current arbitrage matrix (July 2026)
Approximate ranges per 1M tokens for "frontier-adjacent" tiers:
| Tier | Cheapest | Priciest | Spread |
|---|---|---|---|
| Frontier reasoning | DeepSeek R2 ($0.55/$2.19) | Claude Opus 4.5 ($5/$25) | ~10× |
| Frontier general | DeepSeek V3.2 ($0.28/$0.42) | Claude Sonnet 5 ($3/$15) | ~35× on output |
| Cheap general | GPT-5 Nano ($0.05/$0.40) | Haiku 4.5 ($1/$5) | ~12× |
| Coding-specialised | Qwen 2.5 Coder 32B (~$0.20) | Claude Opus 4.5 | ~40× |
The output-token spread dominates the bill. Route generation-heavy tasks to the cheapest capable tier; route input-heavy tasks to whichever tier caches best.
The four arbitrage patterns
- Provider-tier arbitrage — route by task class, not preferred provider. 2026 stacks typically use 3–5 providers concurrently.
- Region arbitrage — Bedrock and Vertex price the same model differently by region. Bedrock cross-region inference (2025+) explicitly enables this.
- Batch vs real-time arbitrage — see the Batch API brief. Same model, 50% off, 24h SLA.
- Off-peak arbitrage — DeepSeek and some regionals price off-peak 50–75% cheaper.
The routing architecture
Request → cheap classifier SLM (~$0.05/M) → task_class, size_tier, urgency → versioned routing table → primary provider (health-checked) → fallback provider.
The classifier is the choke point. A misclassification that sends a hard task to a Nano model, or a trivial task to Opus, wipes the arbitrage gain.
The negotiation floor
For enterprise-committed spend (>$500K/year), published rates are not final:
| Commit tier | Typical discount |
|---|---|
| $500K–$2M/year | 10–20% |
| $2M–$10M/year | 20–35% |
| $10M+/year | 35–50% + custom SLAs |
| Multi-year | +5–10% |
Always negotiate: floor price, price-parity clause, rate-limit lift, cache-hit accounting.
The lock-in trap
Arbitrage requires optionality. Avoid:
- Provider-proprietary features on hot paths (Anthropic file API, OpenAI Assistants threads).
- Fine-tunes on non-portable base models.
- Prompt formats that only work on one provider's tokenizer.
Golden rule: every hot-path prompt must run against at least two providers in your CI eval.
The 2026 gateway that makes arbitrage automatic
LiteLLM, Portkey, Cloudflare AI Gateway, and Kong now support cost-aware fallback routing — define a preference list per task class; the gateway picks the cheapest healthy provider. Programs that adopt this report an additional 15–25% cost reduction on top of existing routing gains.