Rewrite BitRouter vs LiteLLM around claims that hold up

[SPIKESPIGEL404]

Why

LiteLLM's GitHub tagline is now "The fastest, litest AI Gateway. Rust core with Python SDK" and they publish benchmarks at ~0.05ms added overhead / 31.7MB memory. Our comparison page claimed they were a Python gateway at ~20ms+ versus our ~5ms p50.

That's a fight we lose by two orders of magnitude, using their published numbers, on the one page whose entire job is credibility. This reframes the argument onto ground that survives their roadmap.

The new argument, in order

1. All-in-one proxy vs. a router that learns — promoted to lead. Feature coverage is a race they win by definition; a router that folds outcomes back into policy is the one difference they can't ship past.
2. Self-serve cloud vs. enterprise-only — verified against litellm.ai/pricing: quote-only, no public price, SSO/audit logs/RBAC behind it.
3. Rust-accelerated vs. Rust-native — demoted and reframed off language speed. Their recommended mode is a PyO3 bridge inside the Python wheel; Python still owns auth, routing, callbacks, spend tracking. The standalone Axum server is beta with fewer routes and no Docker support. Their own Stage 3 endgame keeps Python in a sidecar for plugins, guardrails, and SSO — which is why this claim doesn't expire when the Dec 1 2026 transition lands.

Concessions added

The page previously read as a strawman. It now concedes, in the prose and the table and the migration guide:

• LiteLLM's in-process library (from litellm import completion) — we have no answer, and the page says so
• 100+ providers plus embeddings / rerank / audio, which BitRouter doesn't ship out of the box
• Enterprise checkboxes they have shipped and we haven't
• Their sub-millisecond Rust overhead, stated plainly as their result

On the security paragraph

Describes the 2026 unauthenticated RCE by incident class — command injection in MCP endpoints chained with a host-header bypass upstream — without CVE numbers or the CVSS score, notes both were disclosed and patched responsibly, and explicitly disclaims immunity: our MCP gateway executes tools too, and Rust prevents none of that bug class. The claim is smaller blast radius, not safety.

Deliberate call. Citing a competitor's CVSS 10.0 reads as FUD and ages badly the first time we ship our own advisory. Easy to harden if you disagree.

Also fixed

Compare table	11 rows → 5; routing-overhead row dropped entirely rather than kept with a number we can't defend
Platform fee	2% stablecoin / 5% card → 0% markup on the OpenRouter and Portkey tables
Terminal animations	Asserted ~85ms p99 · Python GIL and LiteLLM: none of the above (false — they ship MCP). Rewritten with no numbers; marks changed red ✗ → amber ⚠
CompareTerminal	Registered in mdx-components.tsx but missing from COMPONENT_WHITELIST, so any page using it failed lint:docs. Added
Migration guide	Claimed LiteLLM was "self-hosted only" — flatly wrong, and it undercut our own self-serve argument

The terminal programs are corrected but not rendered on the page — no <CompareTerminal> blocks, same as before this PR. The whitelist entry stays because the registry/whitelist desync is real either way.

Numbers

No hard performance numbers are asserted for BitRouter. The Rust section reads "cold-starting in well under a second and holding tens of megabytes resident" — conservative qualitative bands, chosen because the section argues dependency surface rather than speed. Worth replacing with real measurements if this page ever needs to carry a benchmark claim.

Known follow-ups (pre-existing, not touched here)

• The migration guide uses 127.0.0.1:4356 while the rest of the site uses :8787 — and it's in a copy-paste path.
• The 0% markup fix may collide with a parallel branch that already renders none · 0% markup.

Verification

pnpm build clean · pnpm test 83/83 · pnpm lint:docs 49 docs pass · page rendered and checked locally

🤖 Generated with Claude Code