Copperway vs LLM gateways
what the binary actually ships
Copperway SRA intro is €0.10/M in + €0.10/M out. OpenRouter publishes 5.5% on credit purchases and 5% BYOK after allowance — no markup on model list. Plus modeled semantic compression on input. Catalog width is not the fight.
Direct answer
OpenRouter is a US-hosted credit marketplace: Free, pay-as-you-go, and Enterprise; 5.5% when you buy credits; 5% on bring-your-own-key after the allowance (openrouter.ai/pricing). Those two percents are different — do not mix them. Copperway SRA intro is €0.10/M input + €0.10/M output, not a percent of upstream. Five percent of a $15/M frontier output rate is $0.75/M; SRA output is €0.10/M (~$0.109/M at planning FX 0.92). OpenRouter's live homepage states 200T+ Monthly Tokens — monthly (openrouter.ai, 2026-08-19).
OpenRouter volume, as they state it
The figure OpenRouter puts on the homepage today is 200T+ Monthly Tokens. Window: monthly. URL: https://openrouter.ai/. Checked 2026-08-19. They do not publish an all-time “tokens served” total on that page. A Series B post dated 28 May 2026 separately says weekly volume grew from 5 trillion to 25 trillion tokens over the six months before that post (openrouter.ai/blog/announcements/series-b). A DeepSeek V4 note published 30 June 2026 cites a 450 trillion token sample from 1 January–14 June 2026 request logs — that is not an all-time served headline (deepseek-v4-adoption).
Two OpenRouter fees, not one
OpenRouter does not mark up token list prices. Pay-as-you-go credit purchases carry a 5.5% platform fee. BYOK after the monthly list-price allowance carries a 5% fee. The user-facing example “5% of $15/M output = $0.75/M” is the BYOK overage math, not the 5.5% credit-purchase fee. On the same $15/M output, 5.5% of list is $0.825/M.
Copperway's SRA intro is a flat €0.10/M in + €0.10/M out, not a percent take. Standard Gateway is compressed upstream pass-through + €0.15/M. We do not publish a 3.97% product take.
The second lever is compression. Copperway's Semantic Compression Engine (SCE) is modeled at 46% fewer input tokens billed upstream. OpenRouter prompt caching helps repeated prefixes. It is not SCE on novel RAG-shaped prompts.
Most expensive OpenRouter SKUs vs SRA intro
List prices below are from OpenRouter's models API on 19 August 2026, ordered by listed output $/M. Not usage rank. Batch SKUs omitted. Claude Sonnet 4.6 is included as the $15/M output worked example, not because it is the most expensive.
| Model | List in | List out | 5% BYOK take in | 5% BYOK take out | 5.5% credit fee in | 5.5% credit fee out | Copperway SRA in | Copperway SRA out | SCE input saving (46%) | SRA vs 5% BYOK out |
|---|---|---|---|---|---|---|---|---|---|---|
| OpenAI o1-pro | $150.00/M | $600.00/M | $7.50/M | $30.00/M | $8.25/M | $33.00/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $69.00/M | $29.891/M cheaper |
| OpenAI GPT-5.5 Pro | $30.00/M | $180.00/M | $1.50/M | $9.00/M | $1.65/M | $9.90/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $13.80/M | $8.891/M cheaper |
| OpenAI GPT-5.4 Pro | $30.00/M | $180.00/M | $1.50/M | $9.00/M | $1.65/M | $9.90/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $13.80/M | $8.891/M cheaper |
| Anthropic Claude Opus 4.7 Fast | $30.00/M | $150.00/M | $1.50/M | $7.50/M | $1.65/M | $8.25/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $13.80/M | $7.391/M cheaper |
| OpenAI GPT-5.2 Pro | $21.00/M | $168.00/M | $1.05/M | $8.40/M | $1.155/M | $9.24/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $9.66/M | $8.291/M cheaper |
| OpenAI GPT-5 Pro | $15.00/M | $120.00/M | $0.75/M | $6.00/M | $0.825/M | $6.60/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $6.90/M | $5.891/M cheaper |
| OpenAI o3-pro | $20.00/M | $80.00/M | $1.00/M | $4.00/M | $1.10/M | $4.40/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $9.20/M | $3.891/M cheaper |
| Anthropic Claude Opus 4.1 | $15.00/M | $75.00/M | $0.75/M | $3.75/M | $0.825/M | $4.125/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $6.90/M | $3.641/M cheaper |
| OpenAI o1 | $15.00/M | $60.00/M | $0.75/M | $3.00/M | $0.825/M | $3.30/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $6.90/M | $2.891/M cheaper |
| OpenAI GPT-4 | $30.00/M | $60.00/M | $1.50/M | $3.00/M | $1.65/M | $3.30/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $13.80/M | $2.891/M cheaper |
| Anthropic Claude Opus 5 Fast | $10.00/M | $50.00/M | $0.50/M | $2.50/M | $0.55/M | $2.75/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $4.60/M | $2.391/M cheaper |
| Anthropic Claude Fable 5 | $10.00/M | $50.00/M | $0.50/M | $2.50/M | $0.55/M | $2.75/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $4.60/M | $2.391/M cheaper |
| Anthropic Claude Opus 4.8 Fast | $10.00/M | $50.00/M | $0.50/M | $2.50/M | $0.55/M | $2.75/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $4.60/M | $2.391/M cheaper |
| OpenAI GPT-5.5 | $5.00/M | $30.00/M | $0.25/M | $1.50/M | $0.275/M | $1.65/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $2.30/M | $1.391/M cheaper |
| Anthropic Claude Sonnet 4.6 | $3.00/M | $15.00/M | $0.15/M | $0.75/M | $0.165/M | $0.825/M | €0.10/M ($0.109/M) | €0.10/M ($0.109/M) | $1.38/M | $0.641/M cheaper |
Modeled comparison. Not a quote. SCE 46% is modeled and fail-open. SRA intro is subsidized. Campus is pre-construction; executed offtake is zero. Terms · Gateway pricing disclaimer.
Worked example — Anthropic Claude Sonnet 4.6 list $3/$15 per million. Five percent of $15/M output is $0.75/M. Copperway SRA output is €0.10/M (~$0.109/M). That SRA line is $0.641/M cheaper than the 5% BYOK take on that output rate, before counting modeled SCE input savings of $1.38/M (46% of $3/M list input).
What Copperway's code does
Five stages, in order: ingress, PII mask, optional SCE, a single upstream, rehydrate. The PII engine is regex: email, 13–16 digit card-shaped numbers, sk- and Bearer keys, US Social Security numbers. Values sit in memory for that request and return in the response. They are not written to disk. This is not named-entity recognition and not a GDPR Article 9 classifier.
Production chat routing maps a frontier catalog model id onto already-configured upstreams (OpenAI workspace key, Cerebras, Gemini, Moonshot). If there is no match, the historical default remains: Cerebras if that key is set, otherwise the workspace OpenAI key. Images go to Gemini on the masked prompt. Streaming is blocked when any PII was masked. Embeddings are not a first-class Copperway route. Cascade, parallel, and agent orchestration remain stubs. The 25-model JSON catalog is display. It is not a 500-model marketplace.
Copperway Playground is live: 50 free queries per day per signed-in account. Paid credits (€10 PAYG or €9.97/mo) will be available soon. Eligible countries at launch: Cyprus and Greece. Try the playground.
Job matrix — failure mode, not a score
| If you need | Who sells that | Copperway |
|---|---|---|
| 500+ models, auto-route, one invoice | OpenRouter | No. Curated list. One upstream per request. |
| 0% markup, every modality, coding-agent CLI | Vercel AI Gateway | No. Keep the baseURL snippet only. |
| Self-host / Helm / air-gap | LiteLLM | No. Hosted path, not a chart you run. |
| Retry mesh, semantic cache, guardrails | Portkey / Prisma AIRS | No. SCE is compression, not a cache. |
| Spend-by-user dashboards | Helicone | No. Telemetry to Postgres only. |
| Frankfurt router + OpenAI SDK + PII scrub | Requesty (UK; AWS eu-central-1) | Closest clone. We convert on operator jurisdiction and on SRA plus compression. |
| French HQ + multimodal + EU endpoint | Eden AI | Do not copy 500 expert models. Bedrock / Vertex / Azure can still sit behind the flag. |
| €0.10/M SRA + SCE + PII vault + OpenAI SDK | Copperway | Try free AI · Gateway · Reserve SRA |
Requesty's own docs are the cleanest primary source on residency: the EU endpoint guarantees Requesty's processing in Frankfurt. A global model ID still leaves the EU for inference. UK company, AWS. That is the same CLOUD Act test we apply to OpenRouter. Detail: European alternative to OpenRouter. The vs-listicle stays at OpenRouter alternatives.
If you are a publisher
This is a product-capability note, not a securities story and not a live-megawatt commissioning. Reasons a desk can print without laundering a catalog claim:
- Named fee undercut, sourced. OpenRouter publishes 5.5% on credit purchases and 5% BYOK overage. Copperway SRA intro is €0.10 + €0.10 per million. Link both pricing pages. Do not write “5% cheaper models.” Do not invent a Copperway percent take.
- A second lever OpenRouter does not ship on novel tokens. Semantic compression can shrink billed input. Prompt caching is not the same mechanism. Print the fail-open caveat.
- Counter-programming. An EU-incorporated gateway is refusing to clone a 500-model marketplace. The news is the job split, not a fake rank.
- Independent docs on residency. Requesty states that a global model still leaves the EU. Use that, not AGICY paraphrase, when you write “residency is not sovereignty.”
- Honesty constraint is itself copy. Campus pre-construction. Executed offtake zero. PII mask is four regexes. Playground limits are published. That is rarer in this SERP than another “Europe’s AI Gateway” H1.
- Tryable today. Copperway Playground is a live software path, not a hall in Vasilikos. Do not caption it as a Cyprus GPU cloud.
Do not report live MW, 500 Copperway models, open-source Rust runtime, zero CLOUD Act exposure, or a completed EU AI Act conformity assessment. HoldCo is not incorporated. See Terms · Gateway pricing disclaimer.
What we will not claim
- Portkey-class automatic fallback or vector semantic cache.
- OpenRouter-class catalog width or spend-index auto router.
- Helm install, embeddings as a first-class route, live Galaxy inference.
- That an EU logo on a US or UK router repeals US law. PII vaulting reduces what leaves in the clear. It does not kill the CLOUD Act.
Self-hosted OpenAI-compatible upstream
Self-hosted OpenAI-compatible engines (for example EULLM Engine on port 11434) are bring-your-own upstreams. AGICY does not ship that Rust binary; Copperway remains the gateway product. Local audit JSONL ≠ EU AI Act conformity assessment.
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:11434/v1",
api_key="unused",
)Diligence note: EULLM Engine examination · EU open-source AI Act stack.
What to open next
- Developers — Copperway Playground · Gateway product hub
- DPOs — CLOUD Act buyer guide
- Catalog shoppers — LiteLLM vs Portkey vs Copperway
- Offtake — SRA · planned GPU leasing