Model Router

Vital signs

Why it's on the table

On the table, Model Router (Rt) is seat 4 of 58, in the Intelligence family. It is an emerging element — the job is real and here to stay, but the leaderboard still changes quarterly. Choose for this quarter, hold loosely, and watch the changelog. It is optional: plenty of companies run without it — until a specific trigger (scale, regulation, cost, or customers) makes it essential for them. It sits in the lowest paid band — lunch money against the hours it returns.

Model Router: the top 5 — v2026.Q3

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

1. 1OpenRouterOpenRouter, Inc.

   Pass-through token pricing (0% markup) · 5.5% fee on credit purchases ($0.80 min; 5% via crypto) · BYOK free to 1M req/mo, then 5%

   Best for Any team that wants every model behind one key and one invoice today, with zero infrastructure to run.

   The category's runaway aggregator: 25T tokens/week (up 5x in six months, on pace for a quadrillion/year), 8M+ developers, 400+ models across 60+ providers, and a $113M Series B at ~$1.3B (May 2026) with CapitalG, NVIDIA, ServiceNow, MongoDB, Snowflake and Databricks all on the round. Edge routing adds ~25ms; pricing is pure pass-through, so the only cost is the credit fee. Sacra pegs annualized platform revenue at ~$50M by Mar 2026 — proof the ~5% take rate model works at scale.

   Watch That 5.5% compounds painfully at volume — heavy spenders graduate to direct provider contracts or BYOK (which itself costs 5% past 1M req/mo). Fully managed only: your entire model traffic transits a single startup, and critics argue frontier-lab consolidation erodes the aggregator's reason to exist.

   $113M Series B led by CapitalG at ~$1.3B; 25T tokens/week, 8M devs, 400+ models (May 26–28, 2026) [src] · Valuation 2.4x in a year ($547M Jun 2025 → $1.3B May 2026); 100T tokens/month (TechCrunch, May 26, 2026) [src] · ~$50M annualized revenue by Mar 2026 on ~5% take; ~25ms added edge latency (Sacra) [src]
   Site →Docs →

2. 2LiteLLMBerriAI (YC W23; MIT OSS)

   Open source free forever (self-host) · Enterprise custom annual, sized to request capacity — never per-token · 30-day trial

   Best for Teams that need the router inside their own VPC — keys, logs, and spend controls that never leave your infrastructure.

   The open-source gateway standard: 53.8k GitHub stars, 1,000+ contributors, 240M+ Docker pulls, and testimonials from Netflix, NVIDIA, Okta and Stripe. Claims 140+ providers and ~1,900 models behind one OpenAI-compatible API with virtual keys, team budgets, fallbacks and Prometheus metrics all in the free tier. The 2026 Rust-core gateway answered the performance critics: 0.66ms added at p99 (vendor benchmark, ~2,800 RPS).

   Watch You run it — upgrades, scaling, and a famously sprawling codebase (2.5k open PRs) are your problem. Performance numbers are vendor-published; rivals (Bifrost) built entire marketing campaigns on LiteLLM's Python-era overhead. Enterprise pricing is quote-only and opaque.

   53.8k stars, 9.8k forks, 40k+ commits; 100+ provider endpoints (GitHub, Aug 2026) [src] · 240M+ Docker pulls, 1,005+ contributors; 140+ providers, 1,892 models claimed (vendor, Aug 2026) [src] · Rust gateway adds 0.66ms p99 at ~2,800 RPS — vendor benchmark, Aug 2026 [src]
   Site →Docs →

3. 3Vercel AI GatewayVercel

   0% markup, 0% platform fee — provider list price via credits · BYOK free · free tier (subset of models) · metered add-ons (team-wide ZDR/allowlist $0.10/1k req)

   Best for TypeScript/AI SDK teams and anyone who refuses to pay a take rate on tokens at all.

   The price disruptor: GA since Aug 21, 2025 with sub-20ms latency, hundreds of models, automatic failover — and genuinely zero markup, including on BYOK, where OpenRouter charges 5%+. Its monthly production index (a real telemetry set: open-weight models hit 29% of gateway tokens on <4% of spend, Jul 2026) shows serious traffic already flows through it. Vercel subsidizes the gateway to sell the surrounding cloud, which is exactly why it's cheap.

   Watch Youngest of the leaders; observability and governance arrive as metered add-ons (trace drains, tag writes, query fees) that quietly rebuild a bill. Loss-leader economics could change when strategy does. Deepest value only lands if you're inside the Vercel/AI SDK ecosystem.

   GA Aug 21, 2025: zero markup, 100+ models, sub-20ms latency [src] · "AI Gateway charges no markup and no platform fee on tokens"; BYOK with no fee (docs, updated Aug 1, 2026) [src] · Production index Jul 2026: token volume +29% MoM; open-weight models 29% of tokens, <4% of spend [src]
   Site →Docs →

4. 4Cloudflare AI GatewayCloudflare

   Core features free (analytics, caching, rate limits, dynamic routing) · unified billing +5% on credit purchases · Workers Paid ($5/mo) lifts log caps to 10M/gateway

   Best for Teams already on Cloudflare who want a free, edge-speed control plane — caching, rate limiting, DLP — in front of any model API.

   The strongest free tier in the category, run on the network a fifth of the web already transits. The Aug 27, 2025 refresh added dynamic routing (A/B tests, per-user limits, model chaining), Secrets Store key management, DLP scanning in the AI firewall, and 350+ models across major providers — with optional unified billing at a 5% credit fee only if you want one invoice.

   Watch Routing intelligence is thinner than dedicated routers — no learned model selection, and provider translation coverage trails OpenRouter's catalog. Observability depth is Cloudflare-grade metrics, not LLM-native evals. Least compelling if you're not otherwise a Cloudflare shop.

   Core AI Gateway features free; unified billing 5% surcharge on credits; DLP free on all plans (docs, 2026) [src] · Aug 27, 2025 refresh: dynamic routing, unified billing beta, Secrets Store, 350+ models across providers [src]
   Site →Docs →

5. 5PortkeyPortkey → Palo Alto Networks (closed May 29, 2026)

   Gateway fully open source (Mar 2026) · Developer free (10k logs/mo) · Production $49/mo (100k logs, +$9/100k) · Enterprise custom (VPC, SOC2/HIPAA)

   Best for Enterprises that treat the gateway as a governance layer — guardrails, RBAC, budgets, prompt management, audit — not just a switchboard.

   The governance-first gateway at real scale: 1T+ tokens and 120M+ requests daily across 24,000+ organizations, $180M+ in managed AI spend (Mar 2026), with the full enterprise gateway open-sourced that same month. Palo Alto Networks acquired it (closed May 29, 2026) to anchor Prisma AIRS — the clearest possible signal that routing is becoming the enforcement point for AI security.

   Watch The acquisition cuts both ways: the roadmap now serves Prisma AIRS, and the standalone product's independence is over. Buying it increasingly means buying Palo Alto. Pre-acquisition Portkey was also sub-scale commercially ($15M Series A) relative to its traffic.

   1T+ tokens/day, 120M+ requests/day, 24,000+ orgs; gateway fully open-sourced (Mar 24, 2026) [src] · Palo Alto Networks completed acquisition May 29, 2026; Portkey becomes core of Prisma AIRS (terms undisclosed) [src] · Production tier $49/mo for 100k logs; enterprise adds VPC deploy, SSO, custom guardrails (pricing page, Aug 2026) [src]
   Site →Docs →

Model Router: the top 8 compared

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

How to choose your model router

If you just hit two models in production and have no infra team

OpenRouter. One key, one invoice, ~25ms overhead; treat the 5.5% credit fee as rent until your spend justifies direct contracts.

If keys, prompts, or logs cannot leave your VPC — or procurement says self-host

LiteLLM open source. It's the de facto standard (Netflix, NVIDIA testimonials), and the Rust core removed the old performance excuse.

If gateway fees offend you and you're in the TypeScript/AI SDK world

Vercel AI Gateway — 0% markup and free BYOK is the best sticker price in the category; just watch the metered observability add-ons.

If you're already behind Cloudflare and mostly need caching, rate limits, and a kill switch

Cloudflare AI Gateway free tier first; add unified billing (+5%) only if the single invoice is worth it.

If the gateway is your AI governance and security enforcement point (regulated enterprise, agent fleets)

Portkey under Palo Alto — or self-host its now-open-source gateway if the acquisition roadmap worries you; Kong if you're already a Kong shop.

Model Router: the whole field

20 more tools tracked in this category, including 5 dead, renamed, or sunsetting — a reference that hides the graveyard isn't one. Verified 2026-08-06.

Model Router: the category in numbers

Edition v2026.Q3 · ranking, pricing and status verified 2026-08-06.

• OpenRouter raised a $113M Series B at ~$1.3B led by CapitalG (May 26–28, 2026) — the category's first unicorn; weekly tokens 5T → 25T in six months [src]
• Consolidation wave: Palo Alto Networks acquired Portkey (closed May 29, 2026) into Prisma AIRS; TensorZero shut down Jun 12, 2026; Martian and Unify both pivoted away from routing [src]
• Take rates converged: 0% token markup is now table stakes; money is made on ~5–5.5% credit/platform fees (OpenRouter 5.5%, Cloudflare 5%, Eden 5.5%) — and Vercel's 0%/0% undercuts them all (fee survey updated Aug 5, 2026) [src]
• Hyperscalers ship routing natively: Microsoft Foundry Model Router GA Nov 2025 now routes across 28 models including Claude, DeepSeek and Grok — squeezing standalone routers from above [src]
• Routers are the arbitrage layer: open-weight models hit 29% of Vercel AI Gateway tokens on under 4% of spend; Anthropic took 61% of spend on 32% of tokens (Jul 2026 production index) [src]
• Portkey open-sourced its full enterprise gateway (Mar 24, 2026) at 1T+ tokens/day across 24,000+ orgs — governance features are commoditizing fast [src]

Model Router: method & sources

Ranking criteria: effective take rate at real volume, catalog breadth, added latency, governance depth, and survivability — this category had a brutal 2026 (one shutdown, one acquisition, two pivots), so 'will it exist next year' is a first-order criterion. Conflicts resolved: (1) OpenRouter's ~$50M annualized revenue is Sacra's estimate, not company-disclosed — treated as directional; (2) a Medium post claiming Martian 'nearing $1.3B valuation' is unsourced and contradicted by its pivot to interpretability research — rejected; (3) Bifrost's '40x/50x faster than LiteLLM' and LiteLLM's '0.66ms p99' are both vendor benchmarks predating each other's latest rewrites — reported as claims, not facts; (4) model-count claims (400+, 600+, 1,892, 1,000+) are vendor-reported and drift weekly. Catalog counts and uptime figures are single-source (vendor pages). Adjacent elements: inference clouds serving open weights (Together, Fireworks, Groq, DeepInfra) are providers routers route TO → Ow; GPT-5-style in-product auto-routing lives inside the model app → Fm; observability-first tools (Langfuse, Helicone-as-monitor, Braintrust) → Ev; prompt/response filtering as a product → Gd. Cloud-desk aggregators reselling coding-agent subscriptions are out of scope (→ Ca). Ranking criteria: verified commercial traction, independent satisfaction surveys, agent benchmarks, and founder-fit (price floor, lock-in, surfaces). Editorial, never paid — the charter. Machine-readable twin: rt.json.

Our take

A router is optional until you have two models in production. Then it's plumbing you'll wish you'd laid earlier.

Combines with