The Cerver Blog
The economics of AI sessions.
Field notes on what AI coding actually costs — token spend, model reliability, multi-harness routing — and why the session is the unit that finally makes the bill make sense.
Agencies: build the AI backend once. Bill it forever.
Every client brief has "and an AI assistant" in it now. With one client per project it's the same two calls every time — plus cost per client you can put on an invoice, and caps that keep fixed-fee projects fixed-fee.
We ran GLM and Claude Opus 4.8 through 9 coding tasks. The frontier model was the cheapest.
Per-token price is the most misleading number in AI. We ran 9 verifiable tasks through three models on Cerver and tested every answer by running the code — and the "expensive" model came out cheapest and fastest, because the cheap ones burn 10–20× more reasoning tokens.
The loops era has a cost problem. Open-source models are 40% of the answer.
Gemma 4 runs on a 16GB laptop and Mistral keeps shipping startlingly good small models — while agents burn frontier tokens on grep-shaped turns. Route the inner loop to open weights and the bill collapses.
Your AI app's first six months, on a schedule.
Six problems every new AI app builder meets — the key leak, the whale user, the overnight reprice, the weird refusal, the outage, the roadmap hiding in transcripts — each with the day it arrives and the boring fix.
Lovable + Cerver: see what your users' AI actually costs.
Your Lovable app has AI features and no idea what they cost. One prompt pasted into Lovable wires a metered, capped backend — per-user usage, spend caps, model swaps.
Replit + Cerver: see what your users' AI actually costs.
Replit Agent built your app; cerver runs its AI. Secret key in Replit Secrets or publishable key in the client — either way, every user's AI spend becomes queryable.
Base44 + Cerver: see what your users' AI actually costs.
No-code app, real AI backend: point Base44's backend functions at two Cerver endpoints and get per-user costs, caps, and model swaps without seeing a line of code.
Bolt + Cerver: see what your users' AI actually costs.
Bolt apps are client-heavy — exactly what publishable keys are for. Ship a capped pk_ key, stream replies, and read every user's usage in one dashboard.
Build an AI iOS app that can't leak your key.
282 iOS apps just got caught leaking their LLM keys. A complete SwiftUI tutorial for the architecture that can't: an espresso dial-in coach with publishable keys, per-user costs, and spend caps — zero backend code.
Cerver vs. OpenRouter
Both promise "one API, any model" — but one meters every token through a request proxy, and the other runs full agent sessions on your own machines through the flat-rate subscriptions you already pay for.
Wild agents: the billing surprise no one's watching.
Cron jobs and autonomous agents quietly running from laptops and servers are the next AI billing surprise. The fix isn't to kill them — it's one place to see, permit, and bill them all.
The session is the unit. — interactive
Per-seat and per-month hide the truth about AI spend. See one live session in 3D, then the cost distribution behind ten thousand — and why the session is the unit that makes the bill make sense.
GitHub Copilot went token-based. Your $29 bill just became unpredictable.
Copilot moved to usage-based billing and heavy agent users are seeing $750+. The real shift isn't the price — it's that the bill is now unforecastable. Why per-session visibility beats a flat plan.
Uber capped AI coding at $1,500 a developer. The cap is the wrong fix.
A flat budget ceiling treats the symptom — unpredictable, unattributable agent token spend — not the cause. Why per-session cost visibility beats a ceiling.
Series — same brief, two harnesses
Same prompt, two harnesses, two answers
I asked Claude and Codex the same question. Both replied. Both were right — about different things.
What Claude missed: live data
Claude said Clay starts at $349. The actual price is $185. Approximate truth is still wrong.
What Codex missed: a worldview
Codex gave me 23 ranked items. Claude wrote the one sentence that reordered my plan.
$185 vs $349 — small numbers, big consequences
A 47% pricing error doesn't sound like much. Multiplied across decisions, it's six months of runway.
When only one harness sees the deadline
One reviewer surfaced a deadline closing in seven days. The other did not.
Curation is the product
Picking the right model is a 2024 problem. The 2026 move is running all of them and curating the winner.