diff metis competitors/

Metis vs LiteLLM, Portkey, and Helicone.

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.

Pick Metis when…

  • Your workloads mix easy and hard turns and you want the router to learn which is which.
  • You use Anthropic thinking blocks, cache_control, or tool_use heavily and can't afford the LiteLLM bug-of-the-week.
  • Your buyer story needs per-team cost attribution, audit log, and a SOC2 roadmap — not just "we cache calls."
  • You want fan-out turns to spawn planner + worker sessions, with cost attributed to the role rather than just the model.

Pick a commodity gateway when…

  • Single-model workloads — every turn already needs the same model. Slot 4 can't find a cheaper alternative. Caching saves ~22% on long sessions; routing saves ~0%.
  • Very short sessions (under ~6 turns) rarely cross the cache-write break-even — the cache discipline pays off later, not first call.
  • No quality signal at all — no rubric, no test suite, no human feedback. The evaluator has nothing to learn from; pattern store still tracks cost and latency, but success-aware routing needs a verdict.
  • You want a fully-hosted control plane and don't mind vendor lock-in on observability. Metis is a runtime you operate; LiteLLM Cloud / Portkey / Helicone are SaaS dashboards over your traffic.

Run the comparison on your own traffic.

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.