Same model, same quality. 71% cheaper token bill.

All your favorite AIs for about 71% cheaper. Tok Net AI is your drop-in API. It does much more than just reduce costs. See all features down below.

Contact us See how it works
One-line SDK swap OpenAI · Anthropic · Google · open weights
toknet.ai8.4B tokens saved with us
Without Tok Net AI 41.8M tok
With Tok Net AI 9.2M tok
Spend avoided $4,113/day
quality floor: 97.2% held p50 overhead: 38 ms

There are already teams shipping with us today, and these are the numbers we achieved:

71%

Median token reduction across production accounts

4.2×

More usable context inside the same window

38ms

Added latency at p50, including routing

99.96%

Gateway uptime over the last 12 months

How it works

Three lines of setup, then it just gets cheaper.

No rewrites, no prompt migration, no proxy you have to babysit. Point your existing client at the gateway and keep shipping.

Swap the base URL

Keep your SDK, your model names, your streaming. Only the endpoint changes.

client = OpenAI(
  base_url="https://gw.tok-net.de/v1",
  api_key=TOKNET_KEY,
)

You watch the bill drop

Per-feature cost, every compression diff, every routing decision.

week 1  $18,402 → $5,338
evals   97.4% → 97.9%
alerts  0 floor breaches
Token & cost optimization

Our layer between your prompt and the invoice.

Every request passes through our systems before it reaches a model. Quality floors and budgets sit on top, so cheaper never means worse.

$100 of today's inference bill−71% end to end
$29 you still pay $71 never spent
quality floor held
01

Context optimization

02

Rephrasing model

03

Semantic caching

04

Cost-aware sub-agents

Quality floors

A minimum eval score per route. Nothing is ever traded below it.

Never more than the original API

What the provider would have charged you directly is a hard ceiling. We only bill a share of what we save you, so every request comes out below the original price. Never above it!

And ten smaller ways the bill goes down
  • Provider prompt-cache pass-throughReuse the provider's own cached prefix
  • Batch API discountsNon-urgent jobs go at the batch rate
  • Output token capsStop a runaway answer before it bills
  • Agent loop capsA stuck agent cannot spend all night
  • Tool schema slimmingUnused tool definitions come off the request
  • Embedding cacheNever embed the same text twice
  • Retry cost avoidanceA failed call does not bill you twice
  • Duplicate request collapsingIdentical concurrent calls share one answer
  • Cheaper-tier pinningPin routes that never needed the big model
  • Per-feature token accountingSee which feature is actually expensive

The platform.

Savings are the reason teams switch. This is the reason they stay: one gateway that already does the boring, load-bearing work. Open a category to see inside it.

Universal APIOne endpoint and one key in front of every provider, with the failure handling already written 10
  • 200+ models, one schemaEvery provider behind a single request format
  • Drop-in OpenAI compatibilityKeep your existing SDK and model names
  • Chat, responses, embeddingsThe endpoints your app already calls
  • Batch & fine-tuning endpointsLong-running jobs at the cheaper batch rate
  • Fallback logic across providersA second provider takes over on error
  • Load balancingSpread across keys, regions and deployments
  • Retries with backoffTransient failures retried before your app sees them
  • Latency & least-busy routingEach call goes to the fastest healthy endpoint
  • TPM / RPM rate limitsStay inside provider quotas without hand-tuning
  • Tag-based routingRoute by team, feature or environment
ObservabilityLogs, traces, feedback, metadata, filters and alerts, for every call rather than a sampled slice 10
  • Request-level logsEvery call, not a sample of them
  • OpenTelemetry tracesSpans that slot into your existing tracing
  • Prometheus metricsScrape latency, errors and spend directly
  • Spend dashboardsCost broken down by day, model and endpoint
  • Attribution per key, team, user, tagKnow exactly which feature spent what
  • Custom metadata & filtersAttach your own fields and query on them
  • Feedback and eval scoresSend quality signals back against a request
  • Compression diff per requestSee what was removed and what stayed
  • Alerts to Slack, email, PagerDutyKnow when spend or errors move
  • Export to your warehouse or bucketStream raw logs to where you keep data
GuardrailsModel-side and partner guardrails, on the request going out and the response coming back 10
  • PII detection & maskingNames, emails and card numbers redacted in flight
  • Prompt-injection detectionCatch instructions hidden in retrieved content
  • Content moderationBlock unsafe input and output per policy
  • Blocked words and patternsYour own deny list and regex rules
  • LLM-as-a-judge checksA second model scores the answer before it ships
  • 40+ partner guardrailsPlug in the vendor you already use
  • Tool permission policiesDecide which tools an agent may call
  • Bring your own guardrail APIPoint at your own service over HTTP
  • Pre- and post-call hooksRun checks going out and coming back
  • Enforced per key or per teamDifferent rules for different consumers
Prompt managementPrompts stop living in application code and start behaving like anything else you ship 4
  • Versioned templatesEvery change is a version you can point at
  • Typed variablesFill slots without string-concatenating prompts
  • PlaygroundTry a change against real models before shipping
  • Shared prompt libraryOne place instead of six codebases
Team memoryShared context your whole team's apps and agents read from, instead of each one relearning the same facts 10
  • Named memory entriesFacts stored under a key you choose
  • Per-user and per-team scopePrivate notes or shared team knowledge
  • Injected as request contextPulled into the prompt automatically
  • Full CRUD over the APICreate, read, update and delete programmatically
  • Markdown valuesWrite it the way the model reads best
  • Metadata tagsAttach your own structure to each entry
  • Visibility rulesPeople only see what their scope allows
  • Author and timestamp per editKnow who changed a fact and when
  • Survives sessions and restartsNot tied to a conversation or a process
  • Shared across agentsEvery agent starts from the same context
Pricing

You pay a slice of the savings. Never more.

Two ways to use Tok Net AI. Pick one to see how it is priced.

Drop-in proxy

One line changes. Nothing else does.

Your app
Tok Net AI gateway
compress · cache · route
Any model
GPT Claude Gemini Llama +200
Savings share

We bill 15% of the spend we take off your bill. Nothing else. You get the whole platform for free.

Promise: you always save money with us = You never pay more than you would have paid for the original API.
  • No seat fees, no minimums
  • Saving measured per request, auditable
  • Every provider through one key
  • Routing, caching, floors, budgets included
Contact us

Self-hosted, private VPC and custom compressors: talk to us about Enterprise.

FAQ

The questions we actually get asked.

What if we don't save anything?

Then you pay us nothing. We bill 15% of the saving we measure, so no saving means no invoice, and the fee can never exceed the saving itself. That is the promise: your bill with us is never higher than the same traffic sent straight to your provider.

How do you work out what we saved?

Per request. We price the call as it would have run unchanged at your provider's list rate, subtract what it actually cost after compression, caching and routing, and bill 15% of the difference. Every line sits in the dashboard with the prompt diff and the model that answered it, and the whole ledger exports to your warehouse, so you can check our arithmetic against your provider invoice.

Will the answers get worse?

No. We have a quality floor of 97.2%. There will never be a request where we deliver a response that is below this threshold.

How much of our code has to change?

One line: the base URL. Your SDK, model names, streaming and tool calls all stay as they are, and you keep using your own provider keys. Compression and caching need no configuration, so the first reduction usually shows up within the hour; routing gains arrive after the first eval run, typically two or three days in.

Do you train on our prompts?

No. We do not train models on your traffic.

Which providers and models are supported?

Anthropic, OpenAI, Google, Mistral, and any OpenAI-compatible endpoint, including vLLM and Ollama behind your own network. Routing pools are yours to define — you choose which models are eligible, and we never route outside that set.

Get started

Start saving money today.

And not just that. You will get all the features from our platform for free.

Contact us Read the benchmarks