SAVRN Insights · Research · Agent economics

The Inference Paradox

and the enterprise agent trap

Token prices are falling by more than 90 percent. The cost of a finished agent workflow is going up fivefold. Both are true at once. Here is the arithmetic, the evidence, and the five decisions that decide which side of it you land on.

Chad Everett Harris·Founder, SAVRN ·August 23, 2026 ·17 min read ·20 sources

What you see−90%+price per token by 2030

What you pay5×cost per agentic workflow by 2028

The paradox in one picture. The unit price you can see keeps shrinking. The stack you actually pay for keeps growing.

5×

rise in inference cost per agentic workflow through 2028, per Gartner

Gartner, August 17, 2026

−90%+

fall in the price of frontier inference by 2030, also per Gartner

Gartner, March 25, 2026

15× · 1,000×

tokens a multi-agent system and an agentic coding task burn versus chat

Anthropic · Microsoft Research

42%

of enterprises now abandon most AI initiatives before production, up from 17%

S&P Global, 2025

01 · Summary

Cheaper tokens, costlier work

On August 17, Gartner put a number on something I have been watching in our own work for a year. The cost of running an agentic AI workflow will rise more than fivefold through 2028 (Gartner, Aug 17, 2026). In the same breath, Gartner expects the per-token price of frontier inference to fall by more than 90 percent by 2030 (Gartner, Mar 25, 2026). Gartner calls the combination the inference paradox. The visible unit price goes down. The cost of the thing you actually buy, a completed workflow, goes up.

Will Sommer, the Gartner analyst behind the forecast, said it plainly: routing a task to an agentic reasoning model increases inference costs at least fivefold, and potentially by much more as the task becomes more complex (The Register). Tokens are getting cheaper. The savings are not keeping pace with what the newer capabilities cost.

I want to do three things in this piece. First, write the paradox down as an equation, because once you see the shape of it the rest follows. Second, show that every term in that equation is backed by someone other than Gartner: NVIDIA's earnings calls, Anthropic's and Microsoft Research's own token studies, a16z and Epoch AI on deflation, MIT and S&P Global on what happens to projects. Third, reframe the paradox as a trap with five gates, because it is not a weather forecast. It is the sum of five decisions most enterprises are already making by default.

The core finding is simple. The paradox is not a forecasting artifact. It is what you get when four things compound upward and one thing compounds downward, and nobody in the building owns the ratio.

02 · The paradox in one equation

Four terms up, one term down

Gartner built its forecast on a Tokenomics Model run across twelve model archetypes (Computerworld). Strip it to the bone and the cost of one completed agent workflow is five things multiplied together.

Cost per workflow = A × S × T × M × P

▲ up

Agents

How many agents take part in one completed workflow. Swarms push this up.

▲ up

Steps

Reasoning steps per agent per task. Planning, verification and tool loops push this up.

▲ up

Tokens per step

Longer contexts, richer prompts, more retrieval. This goes up too.

▲ up

Tier multiplier

The premium for running on a reasoning tier instead of a chat tier. Gartner puts planning work at 8 to 10 times a basic workflow and reasoning agents at up to 150 times a chatbot on a single task.

▼ down

Price per token

The one term falling, driven by supplier competition and hardware. Down more than 90 percent by 2030.

Four of the five terms are trending up. They are set by architecture choices inside the enterprise. One term is trending down. It is set by suppliers. The paradox is that four compounding multipliers beat one compounding divisor, and they beat it badly.

To make that concrete I tuned a version of the model so the base case lands exactly on Gartner's headline, fivefold from 2025 to 2028. The numbers below are mine, not Gartner's, and they are one reasonable way to get there. The point is the structure, not the decimals.

The 5.0× base case, as a waterfall

Index = 100 at the 2025 baseline, log scale. Four complexity terms compound up; one price term compounds down.

2.35 × 2.60 × 1.95 × 2.47 = 29.4. Multiply by 0.17 and you get 5.0. Complexity grows 29 times. Deflation has to run at 83 percent cumulative just to hold the net at five. Anything less and the multiple grows.

Driver	3-year multiplier	Annual	Read
Agent swarms	2.35×	+33%/yr	modest swarm adoption inside workflows
Reasoning steps	2.60×	+38%/yr	more planning, verification and tool loops per task
Tokens per step	1.95×	+25%/yr	longer contexts, richer prompts, more retrieval
Tier-mix shift	2.47×	+35%/yr	routine work migrating onto reasoning tiers
Cumulative complexity	29.4×		product of the four up-terms
Token price	0.17×	−45%/yr	deflation consistent with more than 90% by 2030
Net cost multiple	5.0×	+71%/yr	29.4 × 0.17 = 5.0

That is the whole story in one line: four compounding multipliers versus one compounding divisor. The tier-mix shift is the largest single bar, and it is the one most people never see, because it is a routing habit rather than an engineering decision.

03 · Independent evidence for each driver

Everyone else is measuring the same thing

Gartner's forecast rests on four things you can observe without Gartner. Each one has been measured by a frontier lab, a chip supplier, an academic group or an enterprise-survey firm. I went and read them.

How many times more than chat

Measured multipliers from first-party and independent sources. Log scale.

Gartner's working range is 5 to 30 times. Anthropic's own multi-agent system sits at 15 times. Microsoft Research measured agentic coding at 1,000 times code chat. The range is the low end.

Reasoning tiers cost an order of magnitude more. Sometimes two.

Jensen Huang said on NVIDIA's fourth-quarter fiscal 2025 call that reasoning can demand 100 times more compute than a simple query, that the vast majority of NVIDIA's compute today is actually inference, and that future reasoning models could be hundreds of thousands, millions of times more compute than today (NVIDIA Q4 FY2025 call, transcript). That is the dominant compute supplier saying, on the record, that reasoning is already an order of magnitude above chat.

The benchmarking data agrees. Across seven benchmarks, OpenAI's o1 produced more than 44 million tokens where GPT-4o produced around 5.5 million. Benchmarking o1 cost $2,767. Benchmarking GPT-4o cost $109. o1-pro and GPT-4.5 were priced at $600 per million output tokens against $15 for Claude 3.5 Sonnet (DeepLearning.AI, The Batch). On a medium-complexity coding prompt, o1 ran at $0.450 per request and o3 at $0.364 against $0.011 for GPT-4, roughly 33 to 40 times more (ValueAddVC). Even a cheap reasoning model loses its headline advantage once the reasoning tokens are counted: a single DeepSeek R1 request can generate 5 to 50 times the tokens of V3.1 (DeployBase).

Agents burn more tokens than chat. Far more than Gartner's range.

Gartner says agents use 5 to 30 times more tokens than a chatbot for an equivalent task (Computerworld). First-party measurements exceed that. Anthropic, writing about its own multi-agent research system, found that agents typically use about 4 times more tokens than chat interactions, and multi-agent systems use about 15 times more tokens as chats. Anthropic also found that token usage by itself explains 80 percent of the variance in performance on the BrowseComp evaluation (Anthropic Engineering). Performance is, to a large degree, bought with tokens.

Microsoft Research measured agentic coding on SWE-bench Verified and found agentic tasks consuming 1,000 times more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost. Runs on the same task differed by up to 30 times in total tokens, and higher token usage did not translate into higher accuracy (Microsoft Research).

At the workflow level the same thing shows up in ordinary examples. A basic chat request runs around 800 tokens. One fraud-check task, with a transaction lookup, risk scoring, a retry loop and case comparison, runs about 13,500 tokens across its tool calls. Seventeen times more for what the user experiences as one question answered (Spheron). And the architecture can make it worse on its own: AutoGen's default context appends every message to every agent's history, so token cost grows with the square of the turn count. At 30 turns the actual cost is 15.5 times the naive linear estimate (n1n.ai).

Deflation is real, fast, and resets at every new tier.

The deflation half of the paradox is well documented and in places runs faster than Gartner says. a16z found that for an LLM of equivalent performance, the cost is decreasing by 10 times every year: GPT-3 cost $60 per million tokens at an MMLU of 42 in late 2021, and an equivalent model cost $0.06 per million by late 2024, a thousandfold drop in three years. The same piece notes that OpenAI's o1 launched at the same $60 per million output tokens that GPT-3 charged at launch (a16z, LLMflation).

Epoch AI tracked the price of reaching a fixed capability across six benchmarks and found declines ranging from 9 times to 900 times per year depending on the milestone (Epoch AI). On FrontierMath, reaching roughly 27 percent accuracy took about 43 million output tokens with o4-mini in April 2025 and about 5 million with GPT-5.2 in December 2025. Epoch's read of the trend is very roughly a 5 to 10 times cost reduction per year for reaching a given capability level (Epoch AI, Feb 2026).

So the price of yesterday's capability collapses. The price of today's capability resets to the top of the range with each release. If your workflows keep migrating to the newest tier, you ride the reset, not the collapse.

Deflation is real. So is the reset.

Dollars per million output tokens, log scale. Sources: a16z, DeepLearning.AI.

The price of yesterday's capability falls about 10× a year. The price of today's capability resets with each release. Workflows that keep migrating upward ride the reset, not the discount.

The tier-mix shift is the quiet driver, and the scaling law pushes it.

The biggest bar in the waterfall is not agents, steps or tokens. It is work drifting onto reasoning tiers. Nobody decides this. It happens one routing choice at a time, because more thinking reliably buys better output. OpenAI wrote it down in the o1 announcement: the performance of o1 consistently improves with more reinforcement learning and with more time spent thinking (OpenAI). Snell and colleagues at UC Berkeley and Google DeepMind formalized it as a scaling law: scaling test-time compute can be more effective than scaling model parameters (Snell et al., 2024).

When the law says more inference compute buys quality and the product owner wants quality, the default is always to escalate. Absent a governance layer, the tier mix only moves one way.

At the industry level, demand is outrunning efficiency.

If cheaper chips absorbed this, you would not see what the supply chain is doing. SemiAnalysis models AI demand at 86 percent of TSMC's 2027 N3 wafer output, nearly entirely squeezing out smartphone and CPU wafers, with 2026 hyperscaler capex consensus moving materially higher and Google's roughly doubling (SemiAnalysis, Mar 2026). Meta raised 2026 capex guidance to $125 to $145 billion, nearly double its $72.2 billion in 2025; combined 2026 capex for Meta, Microsoft, Alphabet and Amazon is estimated around $725 billion, up 77 percent; Meta signed more than 5 GW of capacity in the first half of 2026 alone (SemiAnalysis via Bitget, Jul 2026). IDC expects AI spending to grow 31.9 percent a year to $1.3 trillion in 2029, more than 26 percent of worldwide IT spending (IDC, Aug 2025).

The chips are getting cheaper per operation. The operations per workflow are growing faster. The hyperscalers are voting with about $725 billion.

The failure signature is already in the survey data.

This is not a 2028 problem. Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, naming escalating costs, unclear business value and inadequate risk controls (Gartner, Jun 25, 2025). MIT's NANDA group found that only about 5 percent of enterprise generative AI pilots achieve rapid revenue acceleration, based on 150 interviews, a 350-person survey and 300 public deployments; purchased tools succeeded about 67 percent of the time and internal builds about a third as often (MIT NANDA, via Fortune). S&P Global surveyed more than 1,000 enterprises and found the share abandoning most of their AI initiatives before production jumped from 17 percent to 42 percent in a year, with 46 percent of proofs of concept scrapped before broad adoption (S&P Global Market Intelligence).

Read those three together and you are looking at the enterprise-level trace of the multiplication above. Cost goes up. Value is hard to show. Projects die.

40%+

of agentic AI projects canceled by end of 2027

Gartner

~5%

of enterprise GenAI pilots reach rapid revenue impact

MIT NANDA

17% → 42%

of enterprises abandoning most AI initiatives before production, one year

S&P Global

04 · The enterprise agent trap

Five gates, each closed by default

Gartner's number is not a forecast you wait for. It is a description of a trap that closes on an enterprise the moment it starts buying agentic AI the way it buys SaaS. The trap has five gates. Each one is a decision most enterprises are making right now, by default, without a meeting.

Gate 101

Renting the model layer

The first gate closes when you decide to consume frontier reasoning through metered APIs. Every reasoning-tier release resets pricing to the top of the range. o1 launched at the same $60 per million tokens GPT-3 charged in 2021, which erased three years of deflation in one product cycle (a16z). Rented capability means the vendor captures every efficiency gain and passes on what competition forces it to pass on. Your cost curve follows their roadmap, not your workload.

Beaten byOwnership discipline

Gate 202

Letting complexity grow on its own

The second gate closes when workflow complexity grows bottom-up, team by team. Microsoft Research measured agentic coding at 1,000 times the tokens of code chat with 30 times variance on identical inputs (Microsoft Research). Without a platform layer that caps agents per workflow, steps per task and tokens per step, complexity is set by whoever wants the most impressive demo. It compounds annually into the three red bars of the waterfall.

Beaten byGovernance discipline

Gate 303

No tier discipline

The third gate closes when every workflow defaults to the most capable model available. The tier-mix shift is the largest single driver in the base case. Reasoning tiers run tens of times more per equivalent request than chat tiers (DeepLearning.AI). An enterprise without an inference-tiering rule pays reasoning prices for chat work, and pays them on every one of the millions of calls a swarm generates in a day.

Beaten byGovernance discipline

Gate 404

Ungoverned swarms

The fourth gate closes when multi-agent frameworks ship without orchestration rules. Anthropic's own data puts multi-agent systems at 15 times the tokens of chat (Anthropic), and AutoGen's default context handling scales with the square of the turn count because every added agent re-sends the whole conversation to every peer (n1n.ai). A swarm is the compute equivalent of unbounded recursion. It needs a runtime that enforces handoff rules, memory scopes and escalation caps, or the cost curve goes vertical.

Beaten byOrchestration discipline

Gate 505

No decision provenance

The fifth gate turns cost into cancellation. Gartner's 40 percent cancellation forecast and MIT's 95 percent pilot failure rate share a cause: the enterprise cannot show which agent decisions produced business value. Without a traceable chain from prompt to output to outcome, finance sees spend and no return. When cost triples and value is invisible, the project dies regardless of how good the technology was.

Beaten byProvenance discipline

05 · Getting through the trap

Four disciplines, one stack

The enterprises that come through this are treating agentic AI as industrial infrastructure, not a shopping list of vendor APIs. The pattern has four discipline layers, and each one maps to a gate.

beats Gate 1

Ownership discipline

Own the model layer end to end so token-price resets do not cascade into the P&L. Fine-tuned open-weight models, in-house serving or dedicated capacity contracts turn every efficiency gain into margin instead of surrendering it to metered pricing. It is why the hyperscalers are building capacity rather than renting it, and why roughly $725 billion of 2026 capex is going into owned infrastructure rather than API line items.

beats Gates 2 and 3

Governance discipline

Govern complexity at the platform, not the workflow. Cap agents per workflow, steps per task and tier routing as policy, not preference. Route each request to the cheapest model that can serve it, which is Gartner's inference-tiering advice, but enforce it in the runtime so no team can escalate its way onto a reasoning tier by default. The tier-mix bar is the largest and the most controllable.

beats Gate 4

Orchestration discipline

Run a runtime that treats handoffs, memory scopes and escalation as governed operations rather than emergent behavior of an unbounded chat loop. The 15 times multi-agent multiplier is not physics. It is the observed cost of ungoverned context passing. Bound the memory each agent carries, bound the escalation path, and the overhead comes down hard.

beats Gate 5

Provenance discipline

Instrument every agent decision with the chain that connects prompt, intermediate reasoning, output and downstream outcome. Buy capacity as prepaid units of work whose consumption you control, not metered tokens whose consumption the workflow sets. When finance can see spend and return in one view, projects survive scrutiny.

Same deflation, three disciplines, three futures

Cost multiple per workflow, 2025 to 2028, under the same 83% cumulative token deflation.

Flooroptimized enterprise

15% / 15% / 10% a year, tier mix held flat

Cost roughly halves despite continued adoption.

BaseGartner headline

33% / 38% / 25% a year, tier shift 2.47×

The default for enterprises that make no deliberate architectural choices.

Ceilingunbounded swarm

60% / 55% / 30% a year, tier shift 2.47×

The mode implied by a 95% pilot failure rate and 42% abandonment.

Each scenario applies the same 83 percent cumulative token deflation. Only the four up-terms change. Floor: 1.52 × 1.52 × 1.33 × 1.0 × 0.17 ≈ 0.5. Base: 29.4 × 0.17 = 5.0. Ceiling: 4.10 × 3.72 × 2.20 × 2.47 × 0.17 ≈ 14. The difference is not the technology. It is the operating discipline.

06 · The read

Not the technology. The discipline.

Gartner's inference paradox is arithmetically built in given today's agent defaults, and it is confirmed by independent sources at every layer: the chip supplier, the frontier labs, the academic scaling law, the enterprise surveys. The fivefold multiple is not a worst case. It is the base case for enterprises that make no deliberate choices.

The ceiling is much higher. Let agents, steps and tokens grow at 60, 55 and 30 percent a year and the same deflation still leaves you near 14 times. That is the mode behind the 95 percent pilot failure rate and the 42 percent abandonment number.

The floor is much lower. Hold complexity growth to 15, 15 and 10 percent, keep the tier mix flat, and cost holds roughly flat or falls while adoption keeps climbing. That is the failure-avoidance mode.

The difference between the two is not the technology. It is the operating discipline. The enterprises that come through the 2028 cost curve are the ones building an industrial AI stack, owned infrastructure, governed complexity, orchestrated swarms and audited provenance, before the trap closes. The ones that treat agentic AI as software will be in the 40 percent cancellation cohort with nothing to show for the spend.

The waterfall shows the trap. The four disciplines are the way out.

Where SAVRN stands

This is the problem SAVRN is built around. We own our model layer, we govern complexity and tier routing in the platform rather than per team, we run swarms inside a runtime that bounds memory and escalation, and every deliverable carries its provenance. I wrote up what that looks like in practice in Three Levers, One Loop and The Ownership Divide. If you are trying to keep an agent program out of the 40 percent, the four disciplines above are the checklist.

Start with a working session or see the platform this runs on →

Want the next one?

When a new piece publishes on SAVRN Insights, you get one email with what it covers and a link to read it. No digests, no promotions.

One email when it publishes. Unsubscribe in one click. Privacy

You're on the list. You'll hear from us the next time something publishes.

07 · Questions

Frequently asked questions

What is the inference paradox?

Gartner's name for a situation where the price per token keeps falling while the cost of a completed agentic workflow keeps rising. Gartner forecasts per-token prices down more than 90 percent by 2030 and per-workflow costs up more than fivefold through 2028. Both are true at the same time because a workflow is tokens multiplied by agents, steps and tier, and those multipliers grow faster than the price shrinks.

Where does the fivefold number come from?

From Gartner's August 17, 2026 forecast, built on its Tokenomics Model across twelve model archetypes. The decomposition in this essay, 2.35 × 2.60 × 1.95 × 2.47 × 0.17 = 5.0, is our own tuning that lands on Gartner's headline. It is one reasonable way to get there, shown so the structure is visible.

Why do agents use so many more tokens than chat?

Because an agent re-reads its context on every step: instructions, tool definitions, prior reasoning, tool outputs. Anthropic measured agents at about 4 times chat and multi-agent systems at about 15 times. Microsoft Research measured agentic coding at 1,000 times code chat, with input tokens, not output, driving the cost.

Is token deflation real?

Yes, and fast. a16z finds roughly 10 times per year for equivalent performance; Epoch AI finds 9 to 900 times per year depending on the benchmark. The catch is that deflation applies to yesterday's capability. Each new reasoning tier launches at the top of the price range, so workflows that keep migrating upward ride the reset instead of the discount.

What is the tier-mix shift and why does it matter most?

It is routine work drifting onto reasoning tiers that cost tens of times more per request, one routing decision at a time, because more thinking reliably buys better output. In the base case it is the single largest cost driver. It is also the most controllable, because it is a platform routing policy rather than an engineering fact.

What are the five gates of the enterprise agent trap?

Renting the model layer, letting complexity grow team by team, running without tier discipline, deploying swarms without orchestration rules, and having no decision provenance. Each is a default decision. Together they produce the fivefold curve and the cancellation statistics that follow it.

What do the surviving enterprises do differently?

Four disciplines: own the model layer, govern complexity and tier routing at the platform, orchestrate swarms in a runtime that bounds memory and escalation, and instrument every decision with provenance so finance can see spend and return together. Each maps to a gate.

What are the floor and ceiling scenarios?

Same 83 percent token deflation, different complexity growth. Floor: 15, 15 and 10 percent a year with a flat tier mix, cost roughly halves. Base: Gartner's fivefold. Ceiling: 60, 55 and 30 percent a year with the tier shift, about 14 times. The technology is the same in all three. The discipline is not.

Does owning the model layer really change the economics?

It changes who captures the efficiency gains. A rented reasoning tier resets to list price with each release. An owned stack, fine-tuned open-weight models on infrastructure you control, turns hardware and software efficiency into margin. The hyperscalers' roughly $725 billion of 2026 capex is the same bet at a different scale.

How does this connect to SAVRN's own work?

Directly. Three Levers, One Loop shows our own model program, where the harness and evaluation layers moved production more than fine-tuning did. The Ownership Divide documents what happens to enterprises that skip ownership. This essay is the cost curve that makes both urgent.

08 · Sources

Every source behind this essay

20 outside sources, linked in place, plus three SAVRN pages. The decomposition, the scenarios and the five-gate framing are ours.

We publish the receiptsEvery source behind this essay, grouped by topic, on one page

Gartner

Reasoning-tier premium

Agent token consumption

Deflation and scaling laws

Compute demand and enterprise outcomes

SAVRN

Every figure above was re-checked against its source on August 22, 2026. Where a draft figure could not be confirmed at the source it was removed. Gartner's press releases, OpenAI's o1 post, S&P Global and IDC are cited as the primary sources even though several of those organizations appear elsewhere on savrn.com; the secondary sources were chosen from publications not otherwise used on this site. The fivefold decomposition, the floor and ceiling scenarios, and the five-gate framing are SAVRN's own.