Blog · 2026-07-13 · Vynaris Team
96% vs 30% vs 40%: Auditing LLM Router Savings Claims (2026)
Vynaris claims 96-98% savings, Not Diamond 20-40%, OpenRouter none. All three use different denominators. We recompute Vynaris and OpenRouter on one baseline and sanity-check Not Diamond's fee math: a 100k-task mixed workload saves ~84% naive, ~42% right-sized, ~0% on arbitrage-only routing. Prices
Three routers, three savings headlines that cannot be compared as printed. Vynaris claims 96 to 98% measured savings. Not Diamond claims at least 20 to 40%. OpenRouter claims no percentage at all, just "better prices". All three numbers are arithmetically honest and all three use a different denominator. We recomputed Vynaris and OpenRouter against one common baseline, and sanity-checked Not Diamond's fee and ROI math (its 20 to 40% is measured against your current spend, a denominator we cannot reproduce). On a realistic mixed workload, the spread collapses from 96-vs-30 to roughly 84% vs 20 to 40% vs approximately 0%, and which one you should care about depends entirely on how badly your traffic is over-provisioned today.
Every price below was fetched from the vendor's own pages on 2026-07-13.
The verdict table
Vynaris Not Diamond OpenRouter
------------------------------------- ----------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------
Headline claim 96 to 98% measured savings ([homepage](https://vynaris.com), verified 2026-07-13) "At least 20 to 40% cost savings", 10x ROI ([pricing](https://www.notdiamond.ai/pricing), verified 2026-07-13) No % claim; no markup on inference ([FAQ](https://openrouter.ai/docs/faq), verified 2026-07-13)
Denominator The **requested** model at list price, per request Your current model spend, per team Provider list price for the **same** model
Savings mechanism Downgrade to cheapest model that clears a quality bar Learned per-prompt model recommendation Provider arbitrage on the same model (inverse-square price weighting)
Fees on top Provider list + 3% of first $500/mo, 1% after ([pricing](https://vynaris.com/pricing), verified 2026-07-13) $0.05 per 1M tokens routed, plus your own model bill 5.5% Stripe credit fee, $0.80 min ([FAQ](https://openrouter.ai/docs/faq), verified 2026-07-13)
Our recomputed number, mixed workload ~84% vs a naive baseline; ~42% vs a right-sized one 20 to 40% (claim is plausible; fee is negligible) ~0% on closed models; you pay +5.5% in credit fees
Best when Most traffic is over-provisioned on frontier models You want accuracy-first routing and will write integration code You want one API for 400+ models at list priceA note on the Vynaris fee row: the pricing page gives the full schedule (3% of the first $500/mo, 1% after), while the homepage simplifies to "Original provider pricing + 1%. That's the whole model." The vendor's own two pages disagree at low volumes; we treat the pricing page as authoritative, but an audit article should flag the mismatch, so here it is flagged.
What each number actually measures
Vynaris: 96 to 98%, per-request, against the model you asked for. The mechanism is explicit in the docs: every response carries a direct_equivalent_usd field, defined as "what the requested_model would have cost at list price for the same tokens". Savings percent is computed against that. The docs example shows a request served for $0.000418 against a $0.0182 baseline, which our script confirms as 97.7% saved.
Not Diamond: 20 to 40%, at the team level. The pricing page says teams "achieve at least 20-40% cost savings" without quality degradation, and claims 10x ROI against its $0.05 per million tokens routed fee even in the conservative case. Note the homepage headline is more modest still: 30% cost savings and 5% accuracy gains, with the big number (a +39% accuracy Rootly case study) being about quality, not cost. Not Diamond is a recommendation API, not a proxy: per its docs, it returns a model choice and you make the call yourself.
OpenRouter: no savings percentage, by design. OpenRouter adds no markup on inference; its default router load-balances among providers of the same model, weighted by the inverse square of price, after filtering providers with outages in the last 30 seconds. Savings come from landing on the cheapest host of an open-weights model. For closed models with one canonical price, there is no arbitrage, and the 5.5% credit-purchase fee means you pay slightly more than direct. Cross-model routing exists but is opt-in, via an Auto Router that is itself powered by Not Diamond.
Recomputing on one baseline
We ran the numbers in a short script; the formula below is all of it, so you can reproduce every figure in a spreadsheet. One representative task: 1,200 input plus 300 output tokens.
cost per task = (1,200 x input $/1M + 300 x output $/1M) / 1,000,000Reference prices, verified 2026-07-13: GPT-5.6 Sol at $5.00/$30.00 and GPT-5.6 Luna at $1.00/$6.00 per million input/output tokens (OpenRouter models API); Command R7B at $0.0375/$0.15, the cheapest entry on the Vynaris models table. Per 1,000 tasks: Sol $15.00, Luna $3.00, Command R7B $0.09.
Replicating the 96 to 98%. Request GPT-5.6 Sol, get served a Command R7B-class model, and the receipt reads 99.4% saved ($15.00 down to $0.09 per 1,000 tasks). Request Claude Opus 4.x at Vynaris's listed $15.00/$75.00 and the same downgrade reads 99.8%. A more moderate Sol-to-Luna downgrade reads 80%. So the 96 to 98% headline is reproducible arithmetic, but only when the router downgrades frontier requests to very small models. The denominator is the model you asked for, and the number inflates in exact proportion to how oversized that ask was. To Vynaris's credit, the failure mode is disclosed: when a request needs the model you asked for, "the receipt reads −0%".
A realistic mixed workload. Take 100,000 tasks a month: 60% simple (classification, extraction), 30% medium, 10% genuinely frontier. That 60/30/10 mix — and the mapping of simple to R7B-class, medium to Luna-class, hard to Sol — is our illustrative assumption, not a measured distribution; the recomputed savings depend entirely on it, and they fall as your frontier share rises or as the quality bar lands above the cheapest listed model. With that mix, everything requested as GPT-5.6 Sol costs $1,500/mo. A quality-bar router serving R7B-class, Luna-class, and Sol respectively spends $245.40/mo. That is 83.6% saved against the naive baseline. Real, but not 96 to 98%, because a real workload escalates.
The right-sizing counterfactual. Now assume your team already did the obvious work: Luna for the 90% of easy-plus-medium traffic, Sol for the hard 10%. That bill is $420/mo. The router still finds about $175/mo (41.6%) by pushing the easy 60% below Luna to R7B-class models, but the headline savings just fell from 84% to 42%, and it keeps falling the better your defaults already are. The router's edge is finding the floor below where a human would bother going, per request.
Fees barely move any of this. Vynaris's convenience fee on the $245.40 of routed spend is $7.36/mo (3% of the first $500, 1% after). Not Diamond routes the full workload: 100,000 tasks x 1,500 tokens each = 150M tokens/mo, a $7.50 fee at $0.05/1M — assuming, as we do here, that the fee applies to all request-plus-response tokens; the pricing page does not spell out which tokens are billed. Against a claimed 20 to 40% of the $1,500 naive baseline ($300 to $600 saved) that is 40 to 80x ROI, so the "10x conservative" claim checks out on this shape. OpenRouter's 5.5% Stripe fee turns the $1,500 all-Sol bill into $1,582.50 — assuming all credits are bought by card; crypto and invoicing paths differ. Run your own token mix through the cost calculator before trusting anyone's headline, including this one.
The four setups side by side, monthly bill for the 100k-task workload:
Setup Monthly bill
--------------------------------------- -------------------
Naive: all-Sol, direct $1,500
All-Sol via OpenRouter (card credits) $1,582.50
Right-sized direct (Luna 90% / Sol 10%) $420
Quality-bar routed $245.40 + $7.36 feeWhen NOT to pick Vynaris
Honest cases where a competitor is the better call:
- You already right-size per endpoint. If every endpoint is pinned to the smallest model that passes your evals, a quality-bar router is squeezing the last 40%, not the first 96%. At that point OpenRouter's zero-markup access to 400+ models with automatic provider failover is the lower-friction win, and its fallback semantics are more mature: documented model-fallback arrays versus Vynaris's retry-on-502 with no documented automatic provider failover.
- You run Anthropic-native traffic. Vynaris's own docs state the /v1/messages wire format is not yet implemented; Anthropic-shaped traffic must go through the OpenAI-compatible endpoint, and Claude Code support is described as environment setup for future compatibility, not a guarantee.
- You need operational commitments. Vynaris publishes no rate limits, no uptime SLA, no status page, no regions (checked 2026-07-13; status.vynaris.com does not resolve). OpenRouter has a public status page and an enterprise tier with SOC-2, ZDR, and EU region locking (enterprise page); Not Diamond is SOC-2 and ISO 27001 compliant per its docs.
- Accuracy is the goal, not cost. Not Diamond's routing optimizes quality first with cost as a dial, and its headline case study is +39% accuracy. If you are routing to get better answers rather than a smaller bill, that is its home turf, provided you accept the integration cost: it returns a recommendation and you dispatch the call yourself (docs), plus 100 to 200ms per routed step (pricing page).
Where Vynaris fits
The labeled pitch: Vynaris is the option for the common bad case, a codebase that requests frontier models everywhere because nobody had time to eval alternatives per endpoint. It is an OpenAI-compatible gateway (one base_url change) that sends each request to the cheapest model clearing a quality bar, escalates when needed, and prints the cost and served model on every response, so the 96 to 98% claim is auditable per request rather than taken on faith. Compute is billed at provider list price plus a 3%-then-1% fee; enterprise is $5,000/yr flat, self-hosted on your infra. Get a key at vynaris.com.
FAQ
How much do LLM routers actually save?
On a naive all-frontier workload, a quality-bar router saves roughly 80 to 98% depending on how far requests can be downgraded. On an already right-sized workload, around 40% in our scenario (41.6% computed above; your mix will move that number). On a workload that genuinely needs frontier models, near 0%.
Is Vynaris's 96 to 98% savings claim real?
The arithmetic is real and reproducible: it measures served cost against the requested model at list price. It only applies when most traffic can be downgraded far below what was requested; escalated requests read 0% by Vynaris's own disclosure.
How does OpenRouter save money if it charges list price?
By load-balancing across providers of the same model weighted by inverse-square price, so open-weights models land on cheap hosts. For closed models with one price, you pay list plus a 5.5% credit-purchase fee (card purchases).
What does Not Diamond cost?
$0.05 per million tokens routed, on top of your own model bills, with 100 to 200ms added latency per step and a 50 requests/second rate limit per its docs. On 150M tokens a month (100k tasks x 1,500 tokens) the fee is $7.50.
Which LLM router is cheapest overall?
Fees are noise on all three (single-digit dollars to ~5.5% of spend). What dominates is the routing mechanism versus your workload: quality-bar downgrading (Vynaris) wins on over-provisioned traffic, learned recommendation (Not Diamond) on accuracy-sensitive traffic, provider arbitrage (OpenRouter) on open-model traffic.