OpenRouter
The widest model catalogue you can reach through one endpoint — 425 text models from 80+ providers, 10 trillion tokens a day — and the spend controls now actually work. Two things to price in: the 5.5% fee on every credit purchase, and the fact that since 19/08/2026 the neutral middle layer belongs to Stripe.
PRICING
| Free | Free | 25+ free models, 4 providers, 50 requests/day |
| Pay-as-you-go | 5.5% platform fee | on credit purchases, $0.80 minimum; 500+ models, 80+ providers; token rates at provider cost |
| BYOK | 5% above allowance | free to $25,000/month of list-price inference on PAYG, $200,000 on Enterprise |
| Enterprise | Custom | fee discounts, SSO/SAML, managed policy enforcement, contractual SLAs, invoicing |
The fee lands on the top-up, not the token. A $5 top-up pays the $0.80 minimum — 16% — so consolidate credit purchases. Failed and fallback-triggered requests are not charged, and credits expire after 365 days.
checked 2026-09-03 · vendor pricing page
Element scores
Strengths
Model access is the product and nothing else comes close: 425 text models live on 03/09/2026, 573 across all modalities, 80+ providers, frontier launches usually on day one, all behind an OpenAI-compatible base_url. Everything around it grew up in 2026 — guardrails with real DLP and budget enforcement, per-model-group zero data retention, workspaces with SSO and SCIM, an Analytics API, response caching that bills zero on a hit, embeddings and rerank routers, image, video and audio endpoints, an Agent SDK with MCP, and the Ori harness that runs Claude Code or Codex on any model with one bill.
Honest dings
The 5.5% fee sits on every credit purchase with a $0.80 floor, so small top-ups pay 10-20% and a $200k monthly spend pays about $11k before a single token; BYOK stops being free above $25,000/month of list-price inference. Credits expire after 365 days. There is no self-hosting and no fine-tuning, and a third-party 2026 comparison measured 40-55 ms of added p50 latency against gateways at 10-20 ms. And the neutrality argument now needs a second look: Stripe announced on 19/08/2026 it is acquiring the company (about $7.5B, per the New York Times via CNBC). OpenRouter says product, mission and commitments are unchanged.
Sources (16) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- openrouter.ai/blog/announcements/openrouter-is-join… — Official 19/08/2026: joining Stripe; 10+ trillion tokens/day, 400+ models, 10M+ developers and companies (accessed 2026-09-03)
- cnbc.com/2026/08/19/stripe-openrouter-fintech-ai-mo… — Independent confirmation of the acquisition; ~$7.5B price per the New York Times, terms not disclosed by the companies (accessed 2026-09-03)
- openrouter.ai/blog/announcements/series-b — Official 28/05/2026: $113M Series B led by CapitalG; weekly volume 5T to 25T tokens in six months, 8M+ developers (accessed 2026-09-03)
- openrouter.ai/pricing — Official pricing: free / pay-as-you-go 5.5% platform fee / enterprise; 500+ models, 80+ providers, SSO/SAML, budgets, 50 req/day free limit (accessed 2026-09-03)
- openrouter.ai/api/v1/models/count — Live model count queried 03/09/2026: 425 text models, 573 across all output modalities (accessed 2026-09-03)
- openrouter.ai/blog/announcements/guardrails — Guardrails: DLP with seven built-in sensitive types plus custom regex (redact or block, 403), budget enforcement per member and per key (402), ZDR and provider restrictions (accessed 2026-09-03)
- openrouter.ai/docs/guides/features/zdr — Zero data retention enforced globally, per model group, per guardrail or per request; conservative default when a provider policy is unclear (accessed 2026-09-03)
- openrouter.ai/docs/guides/features/response-caching — Response caching is model-agnostic, runs before the provider, and bills zero on a hit; cache status visible in the Activity log (accessed 2026-09-03)
- openrouter.ai/docs/guides/ori/harness — Ori harness runs claude, codex, grok, hermes, opencode and others on any OpenRouter model with org guardrails and one bill (accessed 2026-09-03)
- openrouter.ai/docs/guides/ori/eval — Ori Eval pins one harness and one model per run for reproducible comparisons, with a GitHub Actions recipe (accessed 2026-09-03)
- openrouter.ai/docs/llms.txt — Documentation index evidencing embeddings and rerank APIs, MCP server and MCP tools in the Agent SDK, audio/PDF/image/video inputs, BYOK and private models (accessed 2026-09-03)
- openrouter.ai/docs/features/prompt-caching — Prompt caching with provider sticky routing, activated only when cache-read pricing beats prompt pricing (accessed 2026-09-03)
- truefoundry.com/blog/openrouter-pricing — Independent pricing analysis: fee compounding at volume (~$11k/month on $200k of credits), BYOK allowance and 5% overage (accessed 2026-09-03)
- requesty.ai/blog/best-llm-routing-platforms-compare… — Third-party gateway comparison measuring 40-55 ms p50 overhead for OpenRouter vs 10-20 ms for self-hosted gateways (competitor-published, treated as a claim) (accessed 2026-09-03)
- braintrust.dev/articles/openrouter-alternatives-2026 — Independent review of fee structure, BYOK allowances and cost-attribution granularity (accessed 2026-09-03)
- menlovc.com/perspective/openrouter-now-processes-mo… — Investor write-up 26/05/2026: ~1.5 quadrillion tokens/year run rate, ~50 employees, $1.3B Series B valuation (accessed 2026-09-03)