Model Prices Are Collapsing. Your AI CRM Bill Isn't. Where Is the Money Going?

On June 30, Anthropic launched Claude Sonnet 5 and positioned it explicitly as a cheaper way to run agents. The launch price is $2 per million input tokens and $10 per million output tokens through August 31, then $3 and $15 after that. For agentic workloads, that undercuts the previous generation by a wide margin.

Anthropic wasn't alone. In the same window, Google cut prices roughly 20% on its AI Ultra tier and made Gemini 3.5 Flash generally available, putting another fast, cheap model into general circulation. OpenAI has been expanding its agentic Codex platform. Three frontier labs, all pushing the price of a unit of machine intelligence down at the same time.

Zoom out and the pattern is unmistakable. The cost of intelligence is falling by roughly half every few quarters. It's one of the steepest sustained price declines in the history of enterprise inputs. Faster than compute in the 90s. Faster than bandwidth in the 2000s.

Now open your AI CRM invoice. Is your AI line item down 50% year over year? Is it down at all? For almost every SaaS buyer we talk to, the answer is no. The line item is flat or growing. Which raises an uncomfortable question: if the underlying input just got dramatically cheaper, and your bill didn't move, where is the money going?

The steepest deflation in enterprise software, invisible on your invoice

The buyer side of the market has already noticed. CNBC reported in late June that enterprise AI buying is shifting away from what insiders call "tokenmaxxing," burning maximum tokens to show maximum activity, toward efficiency. The metric sophisticated buyers now optimize is cost per completed task. Not tokens consumed. Not credits burned. Tasks finished.

That shift matters because tokens were never the thing you were buying. Nobody's board asks how many tokens the sales team processed last quarter. They ask how many deals closed, how many leads got enriched, how many calls got summarized and acted on. Tokens are an input, and the price of that input just fell again.

So the input deflates, the buyer's priorities shift to efficiency, and yet the invoices stay flat. That's not an accident. It's a direct consequence of how AI-augmented CRM vendors chose to bill.

The token meter is a one-way valve

Look at how the largest incumbent talks about tokens. Salesforce's Q1 FY27 disclosures celebrated 28.6 trillion tokens processed, up 152% quarter over quarter. Presented as a growth metric. To a vendor billing on consumption, more tokens means more revenue, so token volume is something to celebrate. To a buyer paying that bill, the same number reads very differently: the meter is spinning 152% faster than it was three months ago.

Then follow the infrastructure investments. On July 1, Salesforce closed its acquisition of m3ter, a metering company that exists to power consumption billing. You do not acquire a metering company because you plan to send smaller invoices. And the new Help Agent is priced at $2 per resolved issue, a work-unit meter layered on top of the token meter. We covered where per-outcome pricing helps and where it hides costs in our breakdown of pay-per-resolution pricing, and the short version is: every meter, whatever its unit, has the same structural property.

A meter converts a deflationary input into an inflationary invoice. Here's the mechanism. When a vendor bills per token, per credit, or per work unit, your price is fixed by contract while their cost floats with the model market. Every time Anthropic, Google, or OpenAI cuts prices, the vendor's cost of goods falls. Your price doesn't. The gap between what you pay and what the inference actually costs widens with every price cut, and the widening accrues entirely to the vendor.

None of this requires bad faith. It's just what meters do. But it means the single most important question in AI procurement right now is not "what does this cost today?" It's "who captures the price declines that everyone knows are coming?" And the answer is determined by architecture. Three choices, specifically.

Choice one: model-agnostic orchestration

The first question to ask any AI CRM vendor: can your platform swap models? Not "do you use multiple models" as a marketing line, but is the orchestration layer genuinely model-agnostic, so that when a cheaper capable model ships, every task that model can handle moves to it?

A platform built this way captures every price cut automatically. Sonnet 5 launches at $2/$10 on June 30; within days, the routing layer sends eligible workloads to it, and the platform's cost per task drops. Gemini 3.5 Flash goes GA; same thing. The platform arbitrages the model market continuously, on the buyer's behalf if the pricing structure passes savings through, or on its own behalf if it doesn't.

A platform locked to a single vendor's models cannot do this. Its costs fall only when its one supplier cuts prices, and its capabilities are capped by that supplier's roadmap. Worse, single-vendor lock-in at the platform layer usually means the platform itself is a captive buyer, and captive buyers pay more. Those costs flow downhill to you. If you've read our Salesforce AI pricing breakdown, you've seen what happens when a locked-in AI stack meets consumption billing: the buyer ends up funding both the lock-in premium and the meter margin.

Choice two: route routine work to cheap models

The second architecture question: does the platform distinguish between routine work and reasoning work?

Here's the thing most vendors won't say out loud. The overwhelming majority of AI work inside a CRM is routine. Extracting a phone number from an email signature. Classifying an inbound lead. Pulling action items out of a call transcript. Normalizing a company name. Tagging sentiment. These tasks do not need a frontier model. A fast, cheap model like Gemini 3.5 Flash or a small Claude variant handles them at a tiny fraction of the cost, at equal or better latency.

Frontier models earn their price on the minority of tasks that genuinely require reasoning: multi-step deal strategy, forecast reconciliation, drafting a nuanced renewal negotiation email. A well-built platform routes each task to the cheapest model that can complete it reliably, and escalates only when the task demands it. We wrote up the full architecture in Dual-Mode AI: Why One Model Isn't Enough for Sales, but the economic point is simple: dual-mode routing is how a platform's blended cost per task falls even faster than any single model's price.

A vendor running everything through one large model, and billing you per token for it, has a cost structure 5 to 20 times heavier than it needs to be for routine work. On a meter, that inefficiency is your problem. You're paying frontier prices for extraction and classification that should cost pennies per thousand operations.

Choice three: a pricing structure that puts the vendor on your side

The third choice isn't technical at all. It's who bears inference cost risk.

Under per-token or per-credit pricing, the buyer bears it. The vendor has no reason to optimize inference cost, because inefficiency is billable. Every wasted token is revenue. This is the core dynamic we described in Usage-Based AI Pricing Is a Trap, and the 2026 price collapse has made it worse, not better, because now the vendor is also pocketing the deflation.

Under flat per-seat pricing, the vendor bears it. Every token the platform wastes comes out of the vendor's margin, not the buyer's budget. That flips the incentive completely. The vendor is now the party hunting for cheaper models, building smarter routing, caching aggressively, and trimming prompts, because efficiency is profit. The buyer gets a predictable bill, and the vendor's engineering roadmap is permanently aligned with driving inference cost toward zero.

Notice that flat pricing only works if the vendor actually controls its inference economics, which loops back to choices one and two. Model-agnostic orchestration and dual-mode routing are what make a flat price sustainable. The three choices aren't independent. They're one architecture, or they're one meter.

Gartner's 40% and the "escalating costs" excuse

Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027, and escalating costs is one of the three named causes. That figure gets quoted as if it were a verdict on AI. It isn't. It's a verdict on procurement.

Think about what "escalating costs" means in a year when Sonnet 5 launched at $2 per million input tokens and Google cut Ultra by 20%. The input got cheaper. The projects got more expensive. The only way both of those are true at once is if the pricing structure between the model and the buyer is inflating on the way through. Escalating cost is not a property of agentic AI. It's a property of metered agentic AI bought without asking who captures the decline.

The projects that survive to 2028 will mostly be the ones where cost per completed task falls every quarter, because the platform underneath them is riding the model price curve down instead of standing on top of it collecting tolls. If you're budgeting a stack right now, our guide to the true cost of your sales tech stack walks through how to model this, but the one-line version is: project your usage forward at success-case volumes, not pilot volumes, and see which pricing structures survive contact with your own growth.

How PipeLance approaches this

PipeLance made all three choices in one direction, and made them early.

The AI layer is multi-provider LLM orchestration, model-agnostic by design. No task in the platform is hard-wired to a single lab's model. When Sonnet 5 shipped on June 30 at launch pricing, eligible workloads could move to it without a rewrite, a migration, or a contract amendment. The same will be true of whatever ships cheaper next quarter.

Routing is dual-mode. Routine extraction, classification, enrichment, and summarization run on fast, cheap models. Frontier models are reserved for the reasoning tasks that justify them: deal strategy, forecast analysis, complex multi-step agent plans. Blended cost per task falls with every generation on both sides of that split.

And pricing is flat per seat: Core at $69 per user per month, Pro at $149, with all 33 capabilities and all 119 AI tools included. No token meters, no credit packs, no per-conversation charges. Call intelligence is native, with zero variable cost per call, rather than a per-minute meter bolted onto a third-party transcription bill. When model prices fall, our costs fall, and our incentive is to keep pushing them down, because we're the ones paying them. That is the entire point.

Step back far enough and this is a familiar story. Every foundational input that has ever deflated this fast, transistors, storage, bandwidth, eventually forced a reckoning between the companies that passed the decline through to customers and the companies that pocketed it. The pocketing works for a few quarters. Then buyers learn to read the curve. Intelligence is now the fastest-deflating input in your entire stack, and the vendors betting you won't notice are betting against the most predictable price trend in software. Escalating AI cost is an architecture choice, not a law of nature. Choose vendors accordingly.