Varsten helps your team spend less on AI by reusing safe repeat answers, sending each request to the right model, and checking that cost savings do not hurt quality.
Drop-in proxy
Keep your provider SDK. Point its base URL at Varsten and swap the key. Streaming, tool calls, and your existing code stay exactly as they are.
from openai import OpenAI
client = OpenAI(
base_url="https://proxy.varsten.ai/v1", # was https://api.openai.com/v1
api_key=os.environ["VARSTEN_API_KEY"], # your Varsten key
)
stream = client.chat.completions.create(
model="gpt-4o",
messages=messages,
stream=True,
)Response reuse
When the same request shows up again, Varsten can serve the stored response instead of paying for another completion. Near-duplicate matching can be enabled only on routes where it is safe; otherwise the request streams straight through untouched.
Routing & evals
Before a route moves to a smaller model, Varsten replays your real traffic through both and grades them with a position-swapped judge. Savings are measured as an A/B difference against a concurrent holdback and reported with confidence intervals — not an estimate. If quality slips past tolerance, the route rolls back.
Reliability
The data plane fails open. If anything upstream is unreachable, requests pass through to your original provider unchanged — you stop saving, you never stop serving. Strict read and total timeouts mean a hung upstream can't pin a connection.
Pricing
Start with Free to monitor AI spend month by month and review savings recommendations. Move to Performance when you want Varsten to optimize traffic directly: 25% of verified savings, with you keeping the other 75%.
For teams that need AI spend monitoring and savings recommendations.
For teams ready to automate savings, billed monthly in arrears only from the dollars it proves.
Savings proof
The calculator shows the shared-savings economics. The proof rules explain what Varsten can bill and what stays off the invoice.
Interactive calculator
Uses a conservative 20% savings assumption. Real billing uses verified savings only, never projections.
Verified savings
Varsten only charges when attribution can defend the delta.