Two gateways, two different bets
Portkey and NeuroRoute solve the same first problem — one OpenAI-compatible endpoint in front of many model providers — and then diverge on what they build on top of that. Portkey's public materials describe a broad AI gateway: a claimed 1,600+ models behind a unified API, an MCP gateway that centralizes authentication and observability for the MCP servers you connect to it, prompt management, RBAC, and an option to self-host the open-source core for free. NeuroRoute is narrower and deeper on one specific bet: that the routing decision itself — which model actually answers a given request — is where most of the addressable cost and quality variance lives, so that's where the engineering weight goes, backed by hard spend controls and a genuinely provable data-retention posture.
This is a comparison, not a takedown. Where Portkey is ahead — and it is, in a few concrete places — this says so plainly. The goal is to give you the actual facts to make your own call, not a reason to pick one side.
Where each one makes sense
Portkey fits well if breadth of provider/model access, prompt management, and MCP governance across a large agent fleet are the primary concern, or if a free, self-hosted open-source deployment is a hard requirement — that option exists today, and NeuroRoute does not currently offer an equivalent self-hosted gateway package.
NeuroRoute fits well if the primary cost driver is model selection at real volume — routing each request to the cheapest model that clears a quality bar, with a hard, provable ceiling on spend — or if your organization needs a specific, documented retention posture (including a genuine zero-retention mode) rather than a general enterprise data-residency option.
Head-to-head, by category
Graded on public, verifiable material as of this writing: Portkey's own site, docs, and pricing page, against NeuroRoute's own shipped, in-production capabilities. "Partial" means the capability exists but is gated to a specific tier or works differently than described; "not shown" means we could not find public evidence either way, which is different from asserting it doesn't exist.
| Category | NeuroRoute | Portkey |
|---|---|---|
| Task-aware / quality-scored routing | Yes — 10+ task types classified per request, per-model×task quality scoring, weighting shifts with classifier confidence | Routing strategies + load balancing + fallbacks are documented; per-task quality scoring not shown in public materials |
| Cascade / escalate-on-failure | Yes, built in — cheapest eligible model first, escalate only on hard failure | Fallback chains are supported; cascade-by-quality-confidence not shown |
| Learned routing from feedback | Yes — embedding-KNN soft-pinning from real thumbs feedback, retain-mode orgs only | Not shown |
| Hard $ budget caps | Four tiers (run, key, org, platform), enforced pre-call, from the entry tier up | "Granular Budget & Rate Limits" — Enterprise tier only, per Portkey's own pricing page |
| Token compression | Six engines (normalize, command-output, rag-dedup, caveman, tabular, ML-based llmlingua2) | Not shown — Portkey's cost lever is caching, not input compression |
| Response caching | Exact-match (L1) + semantic (L2) + provider prompt-caching | "Simple Caching" from the Production tier ($49/mo) up |
| Zero data retention mode | Yes — three explicit modes (zero-strict / zero / retain), proven per-response via an X-Data-Retention header | Not shown as a named mode; VPC/private-cloud hosting is offered at Enterprise for data residency, a related but different guarantee |
| Encryption / customer-managed keys | Per-org AES-256-GCM, customer-managed key (CMEK) option, crypto-shred erasure | Not shown in public materials |
| MCP support | Exposes itself AS an MCP server (agents call NeuroRoute's tools directly, OAuth 2.1) | Runs as a gateway IN FRONT OF your own MCP servers (centralizes their auth/observability) — a different capability, not a lesser version of the same one |
| Guardrails | Input + output, PII (checksum-validated) + prompt-injection (heuristic + optional LLM tier) + custom webhook, fail-closed | PII redaction (homepage); "Deterministic" + "LLM & Partner" guardrails from Production up; custom hooks at Enterprise |
| Self-hosted gateway | Not offered — NeuroRoute is managed only. (Self-hosted vLLM *models* behind it are supported, which is a different thing.) | Yes — free, open-source, self-host option with community support |
| Raw model/provider count | 30+ models, individually quality-scored per task, across 11+ providers | 1,600+ claimed via unified API — materially broader raw catalog |
| Formal compliance certifications | Not certified. A security whitepaper, CAIQ-Lite response, and zero-retention checklist are available on request | Markets HIPAA compliance and SOC2 Type 2 support at Enterprise |
| Pricing model | Flat monthly platform fee + metered $/1M tokens, four tiers ($99–$149 + custom) | Free (self-host or limited SaaS) / $49/mo Production + log-volume overages / custom Enterprise |
Where Portkey is genuinely ahead
Three things are worth saying plainly rather than burying. First, the free, open-source, self-hostable gateway is real and NeuroRoute has no equivalent — if running the gateway itself on your own infrastructure is a requirement, not a preference, that alone may decide this. Second, the raw model count is materially larger; if your use case genuinely needs access to a very long tail of niche or regional models, Portkey's catalog breadth is a real advantage NeuroRoute doesn't match today. Third, Portkey markets HIPAA and SOC2 Type 2 support at its Enterprise tier — a formal certification a compliance team can point to in a checklist, which is a different (and for some buyers, more immediately useful) artifact than an architecture description, however detailed.
Where the two genuinely differ, not just compete
The MCP comparison in the table above is worth restating because it's easy to misread as one product being ahead of the other when they're actually doing different jobs. Portkey's MCP gateway sits in front of MCP servers you already run or connect to, centralizing how agents authenticate to and are observed calling them. NeuroRoute's MCP server is the reverse relationship: NeuroRoute itself is the thing an agent connects to, exposing routing, cost estimation, and savings-query tools directly. A team governing a large fleet of existing MCP servers wants what Portkey built; a team that wants its coding agent or IDE to call a routing decision directly wants what NeuroRoute built. Neither replaces the other.
Retention is a similar case of real difference rather than a simple ahead/behind. Portkey's Enterprise tier offers private cloud/VPC hosting, which addresses data residency — where your data physically sits. NeuroRoute's zero-retention modes address a different question — whether your data is stored at all, provable per response via a response header — and are available below Enterprise pricing, not gated to it. Both are legitimate answers to a security review; they're answers to slightly different questions.
Questions worth asking before you choose
- What fraction of our AI cost is driven by model choice versus by token volume on a small set of models we've already committed to? If it's the former, routing depth (NeuroRoute's core bet) matters more than gateway breadth.
- Do we need a formal SOC2/HIPAA certificate today, or an architecture we can audit ourselves? The honest answer changes which gap in the table above actually matters to you.
- Are we governing MCP servers we already operate, or do we want our own agents/IDEs to call the router directly as an MCP server? These are the two different MCP capabilities in the table, not degrees of the same one.
- Is a free, self-hosted gateway a hard requirement, or a nice-to-have? If it's hard, this decision may already be made.
- How many of the 1,600+ models in a broad catalog will we actually route to in the first year? A bigger number is only an advantage if your traffic needs the tail it represents.
The honest bottom line
If your evaluation is genuinely about routing quality and cost governance at volume — the specific bet this blog has been making the case for across every other post here — NeuroRoute is built for exactly that, with hard budget tiers and a retention story you can verify per response rather than take on faith. If your evaluation is about governing a large existing MCP fleet, needing the largest possible raw model catalog, or requiring a free self-hosted deployment, Portkey's public feature set covers ground NeuroRoute doesn't claim to. Both of those can be true at once, for different teams, and a vendor comparison that only tells you one of them isn't actually helping you decide.