diff metis competitors/
The commodity gateways intercept HTTP and proxy it. Metis runs the same canonical IR through the gateway and through the agent loop, with a pattern store that learns from outcomes and an evaluator that scores them. The features that show up green below are the moat — not the table-stakes rows above them.
Sourced from docs/market-research/03-routing-layers.md and a 2026-05-16 GitHub spot-check. Tracking compiled against LiteLLM 47.2k stars, Portkey 11.7k, and Helicone 5.7k; re-verify before quoting externally.
| capability | LiteLLM | Portkey | Helicone | Metis |
|---|---|---|---|---|
| Wire fidelity | ||||
| OpenAI Chat shape inbound | yes | yes | yes | yes |
| Anthropic Messages shape inbound | partial | partial | passthrough | yes |
| Lossless thinking / reasoning blocks | no (#27512 open) | lossy on convert | passthrough only | yes |
| Lossless tool_use round-trip | regressions (#27469) | normalized | passthrough only | yes |
| cache_control preserved end-to-end | bug-of-the-week | lossy | passthrough only | yes |
| Replays survive provider wire changes | no (logs are wire-format) | no | no | yes (canonical IR) |
| Routing | ||||
| Manual / configured rules | yes | yes (conditional) | light | yes (yaml policy) |
| Per-request override (model in body) | yes | yes | yes | yes |
| Pattern-learned routing from outcomes | no | no | no | yes (slot 4, K-NN) |
| Planner-worker delegation as a tool | no | no | no | yes (delegate()) |
| Capability validation pre-dispatch | partial | partial | no | yes (per-model) |
| Cost attribution | ||||
| Per-key cost rollup | yes | yes | yes | yes |
| Per-user / per-team rollup | per req/user | per req/team | per user/session | yes |
| Cost-per-quality (evaluator joined) | no | no | no | yes |
| Baseline counterfactual (savings_pct) | no | no | no | yes |
| Prompt cache hit-rate per model | partial | partial | yes | yes |
| Quality signal | ||||
| Built-in evaluator (heuristic + LLM judge) | no | no | no | yes (hybrid) |
| Per-turn / per-session / per-workload verdicts | no | no | no | yes |
| Budget-capped LLM judging | no | no | no | yes ($0.10/session, $1/day) |
| Compliance | ||||
| Audit log (key lifecycle + policy events) | no | yes (SaaS) | yes (SaaS) | yes (12-event) |
| GDPR export / forget endpoints | no | vendor-managed | vendor-managed | yes |
| Trace retention sweep with audit exemption | no | vendor-managed | vendor-managed | yes (90-day) |
| Export redaction (4 modes) | no | no | no | yes |
| SOC2 gap audit published | no | Type 2 | Type 2 | Type 1 Q3 2026 |
| Deployment | ||||
| Self-host (BYO infra) | yes | limited | via base URL | yes |
| Helm chart published | community | no | no | yes |
| Docker compose recipe | yes | no | no | yes |
| Local dev mode (no cloud control plane) | yes | no | limited | yes |
| Prometheus /metrics endpoint | plugin | vendor-managed | vendor-managed | yes |
Legend: green = shipped / live-validated; amber = partial or vendor-managed; dim = not available.
The customer trial recipe drops the gateway in front of your existing tools, runs your prompts, and reports cost, cost-per-quality, and the routing chain trace for every turn. Setup is 60-90 minutes; smoke real-API spend is under $1.