Analysis · 2026-10-17 · 7 min read

Why Insurtech AI Token Bills Keep Climbing (and How to Cap Them)

At ITC Vegas, Blitz and Shift Technology said rising AI token costs are a real worry. Gartner says per-token prices fall 90% by 2030 but agentic workflow costs rise 5x by 2028. Here is the TokenCost math for insurers.

!OpenAI logo !Anthropic logo !Google Gemini logo !DeepSeek logo

Insurers Stopped Asking What AI Can Do. Now They Ask What It Costs.

At ITC Vegas, insurance technology leaders told Insurance Business that AI token bills have become a board-level worry.

Sergio Zuniga, CEO of Blitz (a commission management platform expanding from Latin America into the US), was blunt: *"We are pretty scared because we know those companies are losing money. So at some point they will start charging more."* His fear is that the AI work Blitz runs for clients could suddenly cost "double or triple." And the ground keeps moving: *"every 3 months is something completely different."*

CNBC's Contessa Brewer, moderating a fraud panel, shared a story every CFO should hear: one employee experimenting on an enterprise AI platform cost their company $70,000 in a single month.

The Paradox: Cheaper Tokens, Bigger Bills

Two Gartner forecasts explain why both things are true at once:

Price per token is collapsing. Gartner predicts that by 2030, inference on a 1-trillion-parameter LLM will cost providers more than 90% less than in 2025 — but warns those savings won't be fully passed on.
Consumption is exploding faster. Gartner also predicts inference costs per agentic AI workflow will rise more than 5x through 2028, because agents call models over and over in multi-step loops.

Unit prices go down; units consumed go up faster. That is the whole insurtech cost story in one line — and it matches the OpenRouter data showing agents now use 5x more tokens than humans.

The TokenCost Math: A Claims-Triage Workload

Take a mid-size claims operation: 10,000 claims a month, each sending ~20K input tokens (policy, notes, photos described as text) and getting ~2K output tokens back. That is 200M input + 20M output tokens per month, at today's list prices from our pricing table:

Now apply Zuniga's nightmare — a provider triples prices on the frontier model. Sonnet-only goes from ~$900 to ~$2,700 a month. Add a 5x jump in calls as the workflow turns agentic and the same pipeline is heading toward $13,500/month with no change in claim volume.

Shift Technology's Answer: Benchmark Weekly, Route Ruthlessly

Shift Technology CEO Jeremy Jawish described the playbook that works: *"Our job every week is to benchmark these models... Their price is changing, their efficiency is changing, their models are changing."* Simple tasks go to cheap models; only the hardest go to *"the most expensive one."*

Here is that idea applied to the workload above — 80% of claims routed to DeepSeek V4.1, 20% escalated to Claude Sonnet 5:

0.8 × $118 + 0.2 × $900 ≈ $274/month
That is ~70% cheaper than Sonnet-only, and if Sonnet's price triples, the routed bill only rises to ~$634 — not $2,700.

Routing turns a price hike from a crisis into a rounding error.

Five Cost Controls Every Insurtech Should Ship

1. Set hard spend caps per key and per team. The $70,000 experiment happens when nobody owns a limit. Put budgets in an AI gateway.

2. Route by task difficulty. Extraction, classification, and summaries go to Tier-1 models; fraud reasoning and coverage decisions earn frontier pricing. See the three-tier market guide.

3. Cache the static parts. Policy wording, system prompts, and tool schemas repeat on every call — prompt caching cuts that input by 80–90%.

4. Batch anything not real-time. Overnight claims review and renewals scoring can run at ~50% off through batch APIs.

5. Re-benchmark on a schedule. Prices change every quarter. Track value-per-dollar on the LLM leaderboard and switch when the math moves.

Two Speeds, One Budget

Former health insurance CEO Cain Hayes told the panel that companies need *"more than one speed"* with AI: move fast on experiments, slow down on decisions that touch customers' health or finances. The cost version of that rule: give experiments cheap models and tight caps, and spend frontier dollars only where accuracy is worth it.

Bottom Line

Token prices will keep falling, and insurtech AI bills will keep rising anyway — because agents use far more tokens than the price cuts save. The teams that win, like Shift, treat model choice as a weekly buying decision, not a one-time integration. Compare live rates on the TokenCost pricing table.