Real Cost Analysis: 7 LLM Gateways Tested with $100 Budget

May 4, 2026 · 6 min read · Hard data, no marketing fluff

I burned $100 testing 7 LLM gateways in May 2026. Same workload (5M input + 1M output tokens, mix of Claude Sonnet 4.6, GPT-5, Gemini 3, DeepSeek V3.5). Here are the actual numbers.

⚠️ Methodology: same Python harness, same 50 test prompts (RAG, code review, image gen, voice synth), same SLA expectations. No upfront commits. Results from May 1-4, 2026.

The full table (sortable mentally by cost)

GatewaySetup timeMonthly cost (5M tokens)Hidden feesEU-hostedImage/Voice/Video?
Brainiall60 sec$5.99 flatNone✅ Frankfurt + Madrid✅ All included
LiteLLM~30 min$0 (self-host) + ~$45 cloud billYou manage cloud + per-provider feesDepends on host❌ LLM only
OpenRouter2 min~$28 (per-token margin)5.5% credit fee❌ US-only❌ LLM only
Portkey~10 min$49/mo + per-token routedPer-seat at scale❌ US❌ LLM only
Vercel AI Gateway5 min~$30 markup on tokensVercel egress charges❌ US❌ LLM only
Groq3 min~$15 (Mixtral/Llama only)Limited model catalog❌ US❌ LLM only
Cloudflare Workers AI15 min~$22 (10k requests free, then per-token)Cloudflare-tier requiredAnycast (varies)⚠️ Limited image only

Verdict: it depends on your shape

Brainiall — best for SMB/medium volume (1M-15M tokens/mo)

Flat $5.99/mo means predictable bills. 104 models. Image (Seedance, GPT-5-image), voice (ElevenLabs), video (Lyria 3) included. EU-hosted (Frankfurt + Madrid) → GDPR-compliant out of box. Setup in 60 seconds.

LiteLLM — best for 100M+ tokens/mo if you have DevOps

Self-hosted, you control everything. Cost scales with cloud bill, not flat. Best for big shops with platform team. Not for indie developers.

OpenRouter — best for trying many obscure models

Largest catalog (300+). 5.5% credit-purchase fee. Per-token bills can spike fast. Compare OpenRouter vs Brainiall.

Groq — fastest if you only need Llama/Mixtral

Cheapest per-token by far for limited model set. No GPT-5/Claude. Good for high-throughput Llama workloads.

Try Brainiall flat pricing now

60 seconds. No credit card. $5.99/mo flat. 104 models including GPT-5, Claude 4.6, Gemini 3, DeepSeek.

Get free API key →

The hidden cost no one talks about: switching

OpenAI/Anthropic/Google APIs all use slightly different request/response shapes. Most gateways above are partially OpenAI-compatible, meaning your code will break in 1-3 places per provider.

Brainiall is fully OpenAI SDK compatible — change base_url, keep all other code. See compatibility matrix.

What I'd ship to production today

For my own SaaS: Brainiall ($5.99/mo, predictable, all modalities, EU-hosted for my European customers). For a Llama-only RAG cluster at 100M tokens/mo: Groq.

For everything in between, use this calculator with your actual token usage.

Calculate your savings →

Earn 30% recurring

Refer Brainiall to others — get 30%/mo for every active referral.

Become an affiliate →