Most teams can't say what their AI costs — or whether they got what they paid for.
Reliova specs what your workload actually needs, verifies that the capacity behind it is what it claims to be, and supplies it — on one OpenAI-compatible key, one endpoint, one invoice.
Three steps. Where you start depends on where you are.
Building something new? Start at Spec. Already running on a provider? Start at Verify. Both are free, and neither commits you to buying anything.
Spec
Tell us what the feature does and how many people will use it. We model what it will actually cost per month at each model tier, and name the cheapest model that is genuinely good enough for the job.
Verify
We fingerprint your provider's output against a reference implementation and tell you whether the model you pay for is the model you get — at the precision you were promised.
Supply
If our capacity is the right answer, we supply it — screened, continuously monitored suppliers behind one key. If it isn't, we say so and tell you where to buy instead.
What the cost model answers
Nobody can forecast an inference bill from a pricing page. Token counts per call, input-to-output ratios and model choice move the number by an order of magnitude. We work it out from the shape of your feature.
- Tokens per call — estimated from your actual prompt and output shape
- Monthly bill by model tier — the same workload priced across the roster and the closed models you'd otherwise use
- Good-enough analysis — whether this job actually needs a frontier model, which most don't
- What breaks at scale — concurrency, rate limits, and what the bill does when usage grows 10×
Inference is nearly deterministic
With a fixed sampling seed, over 98% of an LLM's output tokens reproduce exactly. Compare a provider's output against a reference implementation and the divergence tells you whether the weights, the quantization or the sampling configuration changed underneath you. Legitimate differences produce far smaller divergence than a real change — that gap is the signal.
- ~100Changed sampling configuration
- ~1,0004-bit weight quantization
- ~10,000KV cache quantization
- variesServing bugs & silent model swaps
We publish the method and hand you the harness, so your own engineers can reproduce every number in the report.
Frontier open models, screened suppliers
Every model here runs on capacity we have tested and keep testing. A supplier that fails an audit comes off the roster.
The roster tracks the frontier as it moves. Need a model, a region or an SLA that isn't here? Ask — we will source it or tell you straight that we can't.
Three ways to buy, one endpoint
Most workloads are a mix. Run user-facing paths on demand, reserve throughput for traffic you can forecast, and send everything that can wait to the batch tier.
On-demand
Pay per token as you go. Fund a balance, draw it down, top it up when you like. The right place to start when your volume is still growing.
Reserved throughput
Lock a dedicated throughput block at a fixed monthly rate. Your cost stops moving with your traffic, and the headroom is held for you.
Batch tier
For work with no deadline: embeddings, backfills, evaluation runs, document pipelines. Submit the job, collect the output, pay less.
Sit behind your brand, not ours
If you ship a product with AI inside it, inference is a supply problem you did not sign up to solve. Reliova runs underneath: your customers see your product, your pricing, your brand — you see one contract and one invoice.
What you get underneath
Bring capacity. We bring the demand book.
Idle accelerators earn nothing. Reliova aggregates enterprise demand into steady, forecastable offtake on one clean contract — so your clusters run loaded, including through the hours they would otherwise sit quiet.
What we ask, and what we give
Straight answers
How do I forecast what an AI feature will cost before we launch it?
Estimate tokens per call from your real prompt and output shape, multiply by expected call volume, and price that across model tiers. A pricing page alone tells you almost nothing, because token count per call and the input-to-output ratio move the bill by an order of magnitude. Reliova builds this model free of charge before you commit to anything.
How do I know if my inference provider is silently degrading quality?
You test it. LLM inference is nearly deterministic: with a fixed sampling seed, over 98% of output tokens reproduce exactly. Comparing a provider's output to a reference implementation reveals changed weights, quantization or sampling config. Roughly 1,000 tokens detects 4-bit quantization; about 100 detects a changed sampling configuration. Reliova runs this audit free of charge.
Does my AI feature really need a frontier model?
Usually not. Classification, extraction, summarisation, routing and most agent loops run on open-weight models at a fraction of the cost with no measurable quality loss. The tasks at the edge of model capability are the ones worth paying frontier prices for — and the point of the analysis is to tell the two apart on your own workload rather than in general.
Which models does Reliova supply?
GLM-5.3 and GLM-5.2 from Zhipu, DeepSeek V4 Pro and V4 Flash, and Kimi K3 from Moonshot AI. All are open-weight models with a 1M-token context window. The roster changes as the frontier moves.
Do we have to change our code?
No. The endpoint is OpenAI-compatible: point your existing SDK at the Reliova base URL and use a Reliova key. Switching between models on the roster needs no code change either.
How does procurement review this?
The models on the roster are published as open weights, so running them is a matter of deploying weights rather than transacting with the labs that released them. Where each workload is served, and how data moves through it, is set out in writing before anything is signed.
What throughput does an account get?
2,000 to 3,000 requests per minute and 10 million to 50 million tokens per minute, sized to the workload. Reserved throughput plans hold that headroom for you exclusively.
Can we resell capacity to our own customers?
Yes — resale is explicitly permitted. Platforms and agencies get per-tenant keys with separate quotas and usage records, and Reliova never appears to your end users.
What if Reliova isn't the cheapest option for us?
Then we say so and tell you where to buy instead. The analysis is worth having either way, and a recommendation you can't trust is worth nothing to either of us.
Start with the free analysis
Building something new, or already running on a provider — either way, tell us the shape of it. We come back within two business days with a scope, and the analysis five business days after that. No charge, no commitment.