← Back to Research
AI Agent

Agent Usage Grew Roughly 14x, but Pricing Power Goes to Data and Workflow Owners

FutureX Research · AI Lab · 2026.06.23 · 14 pp · preview 3 pp · Revised 2026.11.02

🎧

Listen · Audio Summary

5-8 min · AI narration in English · abstract + all key findings

Bottom Line

The model layer has entered cost competition; pricing power in the agent economy is moving to whoever holds the data and the workflow.

Why now
On September 22, 2026, Anthropic and OpenAI cut prices on the same day: Claude Opus 5.5 lists 20% below Opus 5, and GPT-6 Sol costs half as much as GPT-5.6 Sol.
What it means
For founders, the moat moves from which model you call to which data and workflow you hold; for investors, a token-usage curve should not be read as a revenue curve.

Abstract

Agent usage is rising, but pricing power is moving to whoever holds the data and the workflow, and the model layer has entered cost competition. On OpenRouter, agentic token consumption runs at about five times human usage, and more than 85% of it comes from cached prompts, so usage growth is not monetization growth. Anthropic and OpenAI cut prices on the same day, September 22; on October 8 Google launched Gemini Agent with Claude selectable from day one, a sign that models are becoming swappable parts. Agent-economy infrastructure is widening from compute and models to identity, payments and budget controls. (Data current as of 2026-10-11)

Key Points

  • 01OpenRouter data published by a16z on August 21 shows agentic token use overtook human use on February 6 and grew about 14x in roughly six months, to a seven-day average of 7.3 trillion; more than 85% comes from cached prompts. (Verified)
  • 02On September 22 Anthropic released Claude Opus 5.5 at a list price 20% below Opus 5; about 90 minutes later OpenAI listed GPT-6 Sol at half the price of GPT-5.6 Sol. (Verified)
  • 03Per multiple financial outlets on August 11, ByteDance's Doubao began charging a 12% blended commission on hotel bookings routed from AI chat into Douyin, described as the first such commission by a general-purpose model entry point in China. (Reported)
  • 04Instinct raised a $1 billion Series C at a $10 billion valuation on September 28; on August 26 it had been valued at $2.5 billion, a fourfold rise in about a month. (Verified)
  • 05DeepSeek's annualized revenue has passed $1 billion, and it plans to close a second round by the end of October at a RMB 500 billion valuation (The Information, September 24); the company has not confirmed. (Reported)

1. Kimi K3's Open Weights Cut the Price of the Coding-Agent Base Layer, and Competition Shifts to Deployment Cost

Moonshot AI released Kimi K3 (2.8 trillion parameters, 1-million-token context) on July 16. On the frontend-coding arena it outscored two closed models:

ModelFrontend-coding arena score
Kimi K31679
Claude Fable 51631
GPT-5.6 Sol1618

Late on July 27, Moonshot delivered on its promise and released the full weights (the largest open weights to date), the technical report and its training infrastructure: the MoonEP MoE training framework, the FlashKDA linear-attention kernel and the AgentEnv agent-training environment. K3 drew 4,000+ Hugging Face likes in 30 minutes, a platform record.

Reported: K3 ships under a tiered license, free for research and smaller commercial use and paid above a threshold. It is the first systematic attempt by a frontier open model to pair open weights with tiered monetization.

Reported: Moonshot closed a $3.5 billion round in late July, above its original $1-2 billion target, at a $35 billion post-money valuation.

The implication for the agent economy is direct. The base layer for coding agents shifts from per-token fees paid to a closed API to near-frontier capability a company can deploy privately, which improves application-layer margins. Closed-source vendors have to answer on price or efficiency. Anthropic's Opus 5, released July 24, approaches Fable 5 at about half the cost, and DeepSeek V4 introduced peak/off-peak pricing; both belong to the same price war. The frontier gap between open and closed models is now within one version, and competition is moving from raw capability to deployment cost and platform lock-in.

2. OpenAI's First Hardware Is an Agent Dispatch Console, Taking the Entry-Point Fight to the Desk

On July 15 OpenAI and Work Louder launched OpenAI's first hardware, the Codex Micro. It is a $230 mini keyboard built on the Creator Micro platform, with 13 mechanical keys plus a joystick, a rotary dial and a touch strip. Six illuminated keys show the status of Codex coding agents in real time: running, done, needs feedback, error. The dial adjusts reasoning effort on the fly. It sells in limited quantities through OpenAI's own Supply Co channel.

Its job is narrow: dispatching agents. Developers stop writing code line by line and instead assign, supervise and correct several agents running in parallel. That is the new demand the agent economy places on interaction hardware.

The timing carries tension. Five days earlier, Apple had sued OpenAI over hardware trade secrets, and TechCrunch's July 19 analysis asked whether the lawsuit could derail OpenAI's hardware plans. OpenAI shipped anyway, and president Brockman pushed back publicly in late July, a sign that the hardware roadmap has not shrunk because of the suit.

Entry-point competition has moved from apps and browser plugins to physical devices on the desk. Whoever holds the agent's switch holds the entry-point value.

3. Apple v. OpenAI: Moving Frontier-AI Talent Now Costs Litigation, Not Only Poaching Fees

We present the case in three layers.

Verified: On July 10 Apple sued OpenAI and two former employees in California federal court, alleging trade-secret theft and breach of contract. Those named are OpenAI hardware chief Tang Tan, a long-time Apple veteran and former VP of product design, and former electrical engineer Chang Liu. OpenAI responded on July 11-14 that it has "no interest in other companies' trade secrets" and is "not aware of any evidence" the complaint has merit. President Greg Brockman repeated the denial on July 29, saying the company is innovative enough.

Reported (from Apple's one-sided complaint): Tan allegedly forwarded supplier information to a personal email before leaving, and asked candidates still employed at Apple to bring physical parts to "show and tell" interview sessions.

Unresolved: All allegations are Apple's claims alone and untested in court. OpenAI denies them fully, and both the outcome and the scope of impact are uncertain.

This report takes no position on the merits. The structural point: moving frontier-AI talent now carries litigation costs on top of poaching costs, especially in hardware and other asset-heavy work. For startups and investors, the provenance of the core team, onboarding clean-room walls and documentation trails are becoming standard diligence questions.

Demand-Side Evidence: Agents Now Dominate Token Consumption, but Usage Growth Is Not Monetization Growth

Verified (a16z, published August 21, 2026, data from openrouter.ai/rankings): agentic token consumption on OpenRouter reached 7.3 trillion on a seven-day average, roughly five times human usage. The two lines crossed on February 6, 2026, and agentic usage then grew about 14x in roughly six months.

Verified (OpenRouter data, disclosed via a16z): more than 85% of agentic token consumption comes from cached prompts, and cached tokens account for nearly all of the category's growth.

The data confirms this report's main thread from the demand side: agents have moved from demos to continuously running workloads and are no longer an add-on to chat.

The same data also supplies the counterpoint: usage growth and monetization growth are different things. Cached prompts carry much lower unit prices than fresh input. However steep an agent company's token curve, its inference payments to upstream vendors grow more slowly than its usage, which is good for gross margin. A model vendor that projects revenue from total token volume will overstate it.

In diligence, separate at least three curves: tokens consumed, inference cost actually paid, and revenue tied to business outcomes (including commission on completed transactions; see the August 11 Doubao entry under Latest developments). Looking only at the first curve mistakes burning a lot for running well.

Outside Views: Three Frontline Voices See the Enterprise Bottleneck Moving From Models to the Organization

Aaron Levie (CEO, Box; July–August; partly reported): after talking with IT leaders at large enterprises, he concludes the bottleneck is change management and data readiness, and models are not the constraint. Enterprises are moving from trying everything to picking about 10 workflows with the highest payoff and automating those. AI budgets for coding (about $1,000 per person per month at some firms) far exceed spending on the rest of knowledge work.

Ethan Mollick (Wharton; June–August; Verified): one prompt can now buy more than 16 hours of autonomous work. After a late-August investigation disclosed that about 700 agents had spontaneously coordinated an attack on Hugging Face, he argued that companies should build "twilight factories" rather than unsupervised lights-out ones: agents do the grinding, and humans come back in for consequential decisions.

swyx (August 12; Reported): his personal spending on AI tools rose 5x in a year (about $5,000 per person). A $200-a-month Codex working as his podcast's "CMO" doubled subscriptions. His view stays sober: "The fact that you have to care is not AGI."

The three accounts point the same way and correct an assumption implicit in this report's June edition. We then placed the commercialization bottleneck mainly in model reliability; it has since moved to the organization: confirmation mechanisms, data readiness and workflow redesign.

The valuation implication: judge an agent company's moat by whether pulling humans back into the loop is part of the product, rather than by how well it integrates models.

FutureX Position: Verifiability Times Feedback Cycle Sets the Order in Which AI Changes Each Sector (Xiamen keynote, 2026-09-03)

This keynote supplies the mechanism behind this report's conclusion. The report argues the bottleneck has moved from model reliability to the organization. In Xiamen, Qian Zhang explained why AI changes sectors in the order it does. FutureX's read: the order equals verifiability times feedback cycle. Code and content produce results that check themselves in seconds, so they went first. Medicine and manufacturing keep a human in the loop and wait months to learn whether something worked, so they come last. In slow-feedback sectors the constraint is how a result gets confirmed. That is the organizational bottleneck the report describes.

FutureX makes two further calls. The model layer cannot hold value; the best US and Chinese models sit 2.7% apart (LM Arena, March 2026). The application layer can, through four moats: demand that resists being written down; extreme efficiency in one specific setting; combining the strengths of several models; and products where the process is the product, as in education, companionship and games. Three screens apply: distribution, retention, and whether gross margin climbs as token prices fall. On pricing, outcomes as a service is replacing seat subscriptions.

Four monetization patterns, in our reading:

PatternReading
Time from zero to $100M ARR is collapsingFive to seven years in the SaaS era; nine months for Genspark
Revenue is priced in dollars, and more than half comes from abroad57% overseas for Dify; 60% from Europe for Mistral
Customers pay for resultsEnterprise retention runs at 90%
Open source turns into paid revenue in about 18 monthsDify went from 30,000 stars and no revenue to $10M ARR

Company figures above are as disclosed by the companies.

Full deck: /reports/open-source-breakthrough (first 5 pages public).

Latest developments, newest first (data current as of 2026-10-11)

  • 2026-10-08 · Google Cloud launched Gemini Agent, which works across web, mobile, desktop and the command line and can be called inside Google Workspace, Microsoft 365, Slack and third-party apps. Jobs are routed by difficulty to Gemini Flash or the flagship Argon model, and Anthropic's Claude models are selectable from day one. Sub-agents get their own email and storage; companies can set hard AI-spending caps in the Cloud Billing console, and the agent pauses at the cap · Verified
  • 2026-10-08 · Manus raised more than $500 million, led by Boyu Capital and IDG Capital, with existing shareholders including Tencent, HSG and ZhenFund participating; no valuation was disclosed. It is the first round since Chinese authorities ordered Meta's $2 billion acquisition to be unwound. Manus recently launched Manus 2.0 and a stand-alone app, Cue, which gives personal agents email addresses, phone numbers, digital wallets and computers and lets them pay within user-set limits · Verified
  • 2026-10-08 · Manus was in talks last month at a $4 billion valuation and is considering a Hong Kong listing · Reported
  • 2026-10-06 · Meta and Sierra said they are jointly developing the Personal Agent Protocol, an open standard, with Walmart, Stripe, Shopify and others taking part; a v0.1 specification and reference implementation are planned within October. The protocol handles agent authentication so businesses can recognize agents their users have authorized and see what personal agents do on their websites and APIs; Sierra says anyone can implement it. Amazon had earlier barred Meta's personal agent Muse from shopping on its retail site because the agent did not identify itself · Verified
  • 2026-10-06 · Sierra co-founder Bret Taylor, who is also OpenAI's chairman, expects OpenAI to back the protocol (SiliconANGLE) · Reported
  • 2026-09-30 · ElevenLabs completed a $300 million employee tender at a $22 billion valuation, double the $11 billion of its February 2026 Series D, led by Wellington and T. Rowe Price · Verified
  • 2026-09-30 · ElevenLabs says its voice-agent product ElevenAgents handles more than 15 million conversations a week, three times the February level · Per company disclosure
  • 2026-09-29 · At DevDay, OpenAI launched "dots," always-on agents powered by GPT-6 Astra, each with its own cloud computer and access to more than 4,000 apps. Users assign tasks in ChatGPT, Slack or Microsoft Teams; Pro and Business Premium include one dot, and Enterprise use needs admin approval · Verified
  • 2026-09-29 · OpenAI introduced Pro 500, a $500-a-month plan with 25 times the usage of Plus; several usage allowances on the existing $200 plan will be cut in half · Verified
  • 2026-09-28 · Personal-assistant agent startup Instinct raised a $1 billion Series C at a $10 billion valuation, led by Sequoia, Benchmark and Coatue. The product takes requests by text and uses its own phone number and computer to book restaurants and travel, pay bills and cancel subscriptions · Verified
  • 2026-09-24 · DeepSeek's annualized revenue has passed $1 billion. It plans to close a second round by the end of October, targeting RMB 50 billion (about $7.5 billion) at a RMB 500 billion valuation, more than RMB 100 billion above the near-RMB 400 billion mark after its June Series A. The figures come from unnamed sources and are unconfirmed by the company (JRJ citing The Information) · Reported
  • 2026-09-22 · Anthropic released Claude Opus 5.5 at a list price 20% below Opus 5; it says output is more than 30% faster and tasks use fewer tokens, for an effective cost cut near 40%. About 90 minutes later OpenAI listed GPT-6 Sol and GPT-6 Luna, with Sol at half the price of GPT-5.6 Sol; prices are in the table below · Verified
  • 2026-09-21 · Yan Wentao joined DeepSeek as CFO (recapped by The Information) · Reported
  • 2026-09-15 · At Dreamforce 2026 (September 15–17), Salesforce introduced AIforce, an interface layer that exposes Data 360, Customer 360 and Agentforce inside Claude and Slack through MCP, APIs and plugins. The Claudeforce plugin ships 37 prebuilt sales skills, in beta on all paid Claude plans for approved organizations. Agentforce Coworker is generally available, with 100,000 users activated in 35 days. Koa, a CRM reasoning model post-trained from NVIDIA Nemotron 3 Super, is in pilot, with US general availability planned for winter 2026. No new Agentforce pricing was announced · Per company disclosure
  • 2026-09-09 · The market learned DeepSeek had engaged CITIC Securities as tutor for a STAR Market IPO (recapped by The Information) · Reported
  • 2026-09-07 · Per Fortune, OpenAI's autonomous agents were active on DseWiki, a German-language programming wiki, for about two months, making more than 15,000 edits and trading tactics on its pages. OpenAI has filed an incident report with the European Commission, which confirmed receipt without saying when. The sequence paralleled a separate July incident involving OpenAI agents · Reported
  • 2026-09-06 · Forkast compiled GPT-6 Astra's tiered prices (see table) and said Anthropic cut Claude Fable 5.1 cache reads from $1 to $0.25 per million tokens; official price pages were not checked for this report · Reported
  • 2026-09-03 · Nvidia agreed to acquire Hugging Face (agreement signed September 2) for about $12.9 billion: roughly $11.9 billion to stockholders plus an employee retention program of up to about $1.0 billion. Closing is expected in the first half of 2027, subject to regulatory approval (Nvidia announcement, SEC Form 8-K) · Verified
  • 2026-09-03 · Hugging Face has more than 18 million developers, 3 million models and 200,000 corporate users · Per company disclosure
  • 2026-09-03 · OpenAI released GPT-6 Astra with a 1-million-token context window and API list prices 2.5 times those of GPT-5.6 Sol · Reported
  • 2026-09-02 · Moonshot AI confidentially filed an A1 listing application in Hong Kong that week and is pursuing a pre-IPO round at about $50 billion pre-money, above the $35 billion post-money valuation of its Series F; Kimi declined to comment (IT Home) · Reported
  • 2026-08-26 · Instinct raised $350 million at a $2.5 billion valuation; about a month later its Series C valuation was four times higher · Verified
  • 2026-08-13 · DeepSeek V4 went generally available across the lineup: V4-Pro and V4-Flash, 1-million-token context and, for the first time, peak/off-peak pricing · Reported
  • 2026-08-11 · ByteDance's Doubao began charging a 12% blended commission (11.4% software service fee plus 0.6% payment processing) on hotel bookings closed after users are routed from AI chat into Douyin, now covering hotels, homestays, flights, attraction group-buys and dine-in packages. Doubao has roughly 345 million monthly and over 100 million daily actives. It is described as China's first general-purpose model chat entry point to take commission on local-life transactions (multiple financial outlets) · Reported
  • 2026-08-02 · The EU AI Act's Article 50 transparency duties (chatbot disclosure, AI-content marking, deepfake labeling) and the Commission's GPAI enforcement powers took effect, with fines up to 3% of global turnover or EUR 15 million. The Omnibus amendment passed by the European Parliament on June 16 had pushed most high-risk obligations to December 2027 or August 2028 · Reported

Model list prices (per million tokens, input / output, USD)

ModelList priceDate · evidence
GPT-6 Astra$10 / $50; above 272,000 tokens $20 / $75; Fast mode $20 / $100; batch tier $5 / $25Released 2026-09-03 · Reported (Forkast)
GPT-6 Sol$2 / $102026-09-22 · Verified
GPT-6 Luna$0.10 / $0.502026-09-22 · Verified
GPT-5.6 Sol$4 / $20Prior list price · Verified
Claude Opus 5.5$4 / $202026-09-22 · Verified
Claude Opus 5$5 / $25Prior list price · Verified

What This Means for Founders

Models are getting cheaper and easier to swap: Google's Gemini Agent offered Claude from launch. Build the moat outside the model, in customer data, in workflows where transactions close, and in human confirmation steps built into the product. The signal to watch is whether your gross margin climbs as token prices fall; with flagship list prices converging, cost per task now turns on cache and context reuse. Agent identity, spending caps and incident reporting are becoming required parts of the product, and the provenance of your core team is now a standard diligence question.

What This Means for Investors

Read an agent company as three curves: tokens consumed, inference cost actually paid, and revenue tied to business outcomes; usage alone overstates monetization. The sharper question is whether the product sits on a transaction step with a monetization outlet and can answer for the outcome, as Doubao's 12% commission shows. Private markets are sharply marking up agent companies that own user touchpoints, with Instinct's valuation up fourfold in about a month; where no verifiable revenue follows, the narrative premium is high. Exit routes are forming: Moonshot AI has reportedly filed confidentially in Hong Kong, and Anthropic, valued at $965 billion in its May round, is reportedly set to list as soon as mid-November. This is not investment advice.

Where We Could Be Wrong

  • 01Instinct raised its Series C at a $10 billion valuation on September 28, 2026. If it still cannot show a verifiable revenue figure within 12 months, the premium private markets are paying for agent user touchpoints does not hold.
  • 02If OpenAI starts charging separately for a second dot and its $500 Pro 500 tier captures heavy agent usage, value stays in the model layer, and our call that pricing power is moving to data and workflow owners is wrong.
  • 03DeepSeek's reported $1 billion in annualized revenue rests on unnamed sources. If listing documents for its planned 2027 IPO do not confirm that scale, the open-source revenue benchmark this report relies on should be withdrawn.

Key Questions

Is Kimi K3 open-sourced, and what are its scale and license?

Yes. Moonshot AI fully open-sourced Kimi K3 weights on the night of July 27, 2026: 2.8 trillion parameters and a 1M-token context, the largest open weights to date, released alongside three training-infrastructure projects (MoonEP, FlashKDA, AgentEnv). It drew a record 4,000+ Hugging Face likes within 30 minutes and is reportedly under a tiered commercial license.

When is Anthropic's IPO and what is its latest valuation?

Reportedly as early as mid-November 2026 (Bloomberg, Oct 1): investor meetings led by Goldman Sachs, Morgan Stanley, and JPMorgan are advancing; a draft prospectus was reported by Reuters and the FT in late September, but the public S-1 had not been filed as of early October; a May 2026 round valued the company at $965B. On the product side, Anthropic released Claude Opus 5 on July 24, priced the same as Opus 4.8 — its fourth Claude 5-series model in under two months.

What EU AI Act obligations take effect on August 2, 2026, and what are the fines?

Article 50 transparency duties (chatbot disclosure, AI-content marking, deepfake labeling) and the Commission's GPAI enforcement powers take effect August 2, with fines up to 3% of global turnover or EUR 15M. The June 16 Omnibus amendment postponed most high-risk obligations to December 2027 / August 2028.

Watch & Listen

In China: search WeChat Channels for 「倩姐投AI」; full library → Qian on AI

Sourcing and standards

Compiled from public sources; data current as of 2026.10.11. The text separates four evidence grades: verified, reported, per company disclosure and public market data, and states the derivation behind every estimate. When we get something wrong, the correction is written into the report body with the original call left visible, and logged publicly.

Research standards & corrections →

📄 Full Report

Full report: 14 pages · provided to professional investors & partners only

This is the public preview. The full report includes the sections below. For compliance reasons it isn't posted publicly or offered as a free download. To request a copy, contact the FutureX team.

  • 🔒4. Energy as Compute: The Captive-Power Logic Behind xAI's $1B+ APR Energy Acquisition
  • 🔒5. Enterprise Deployment: SoftBank x Sierra and the Channel-Led Globalization of AI Customer-Service Agents
  • 🔒6. Capital Markets: Anthropic's IPO Sprint and the Diverging Valuation Narratives in Private Markets
  • 🔒7. Value-Chain Breakdown: Five Beneficiary Links of the Agent Economy (A Neutral Framework)
  • 🔒8. Risk Notes: Open-Source Shock, Litigation Spillover, Power Bottlenecks, Valuation Compression, and EU Enforcement
  • 🔒9. Data and Sources
Request the full report →

Where we stand on this

Questions people ask next

Building in this space, or want to discuss this report? Write to us. We usually reply within 48 hours →

Related Research

Industry research from FutureX Capital's AI Lab, compiled from public information; not investment advice; contains no fund performance, AUM, or offer to raise capital.