Tokens No Longer Have One Price: Cheap Tier Cut 90%, Fast Tier Costs 6x
FutureX Research · AI Lab · 2026.03.15 · 26 pp · preview 7 pp · Revised 2026.11.02
Listen · Audio Summary
5-8 min · AI narration in English · abstract + all key findings
Bottom Line
Token volume is surging, prices are splitting by tier, speed and caching, and compute is locked up in long contracts and possibly debt.
- Why now
- On October 7, 2026, Anthropic cut Haiku 5.5's short-prompt list price by 90%; the next day OpenAI priced GPT-6.1 Sol's Ultrafast mode at 6x its standard rate.
- What it means
- For founders, costs must be modelled by tier, speed and caching; for investors, rising token volume does not mean vendor revenue rises with it.
Abstract
Token volume is still surging, but prices now split by tier, speed and caching, and compute is being locked up in long contracts, possibly in debt too. Inference is about two-thirds of global AI compute, and Goldman Sachs expects monthly token consumption to grow 24x by 2030. On October 7, Haiku 5.5 cut short-prompt list prices by 90%; a day later GPT-6.1 Sol's Ultrafast mode was priced at 6x its standard rate. BenchLM's frontier price index held flat from July to September. Anthropic reportedly carries about $518 billion in compute commitments, and SpaceX is reportedly seeking about $40 billion to buy Nvidia chips. This is not investment advice. (Data updated through 2026-10-11)
Key Points
- 01Demand hit another record: in the week to September 27, 2026, OpenRouter processed 146.0 trillion tokens, up 13.2% week on week, with Chinese models at 63.0% of the top-20 total (Reported).
- 02Agents are now the main consumer: a16z reported on August 21, 2026 that agentic use on OpenRouter hit 7.3 trillion tokens on a seven-day average, about 5x direct human use, with over 85% from cached prompts (Verified).
- 03Prices split both ways: on October 7, 2026 Anthropic cut Haiku 5.5's short-prompt list price 90% to $0.10/$0.50; the next day OpenAI priced GPT-6.1 Sol's Ultrafast mode at 6x standard (Verified).
- 04Frontier prices stopped falling: BenchLM's frontier token price index read 16 in September 2026 (March 2023 = 100), flat with July and August, a third straight month without a decline (Public market data).
- 05Contracts and debt: Anthropic's draft prospectus (September 28-29, 2026) shows about $518B in compute and related commitments; in October SpaceX sought about $40B of debt financing to buy Nvidia chips (Reported).
1. Inference Is About Two-Thirds of AI Compute: Token Volume Rises While Unit Prices Fall
The center of gravity in AI compute has moved from training to inference. Deloitte's TMT Predictions 2026 puts inference's share of global AI compute as follows:
| Year | Inference share of global AI compute |
|---|---|
| 2023 | About one-third |
| 2025 | About half |
| 2026 | About two-thirds |
Goldman Sachs' May report, Decoding the Agentic Economy, draws a steeper curve: monthly token consumption reaching 120 quadrillion by 2030, 24x today's level. Consumer agents contribute 12x growth by 2030, and enterprise agents 55x by 2040.
The driver is that agents run continuously. On OpenRouter, agentic token flow overtook direct human usage in early February 2026, and a single agent request consumes about 15x the tokens of a human chat.
While demand surges, unit prices are collapsing. Per-token inference cost at the chip level is falling 60-70% a year. The price war of late July turned price cuts from an annual event into a weekly one: - July 24: Anthropic released Opus 5 at $5 per million input tokens and $25 output, with an adjustable effort dial. - July 30: OpenAI cut GPT-5.6 Luna by 80%. - July 31: DeepSeek shipped V4-Flash 0731 at $0.14/$0.28. - Google's Gemini 3.6 Flash (July 21) and Alibaba's Qwen3.7 Flash (July 27) launched in the same window.
Volume up, unit price down: that gap is the thread running through our value-chain analysis.
2. Chinese Models Take 58% of U.S. Companies' OpenRouter Usage, and Open Weights Become a Valuation Engine
Kimi K3, released July 16, is a 2.8-trillion-parameter MoE model with a 1M-token context window. On the evening of July 27, Moonshot AI released the weights on the date it had promised, making Kimi K3 the largest open-weight model in history. Together AI and Modal offered day-0 hosting.
Share also hit a record:
| Period | Chinese models' share of U.S. companies' OpenRouter token usage |
|---|---|
| Early 2025 | Under 10% |
| Early July 2026 | Briefly 63% |
| July 2026, full month | 58%, a record |
All five of July's most-used models were Chinese: Xiaomi's MiMo-V2.5 first, followed by the DeepSeek, MiniMax, Qwen and Kimi families.
Price is the main driver. Open Chinese models typically cost 60-90% less than leading closed U.S. APIs. Agentic workloads multiply usage 15x, and the price gap grows by the same factor.
Capital followed. Bloomberg reported on July 29 that Moonshot closed a $3.5 billion round at a $35 billion post-money valuation, far above its initial $1-2 billion target, and plans a pre-IPO round in August at up to $50 billion ahead of a Hong Kong listing (Reported). Open weights have turned from a technical gesture into a valuation engine.
Capital moves at two other Chinese model companies (both Reported): - Zhipu placed new H-shares at HK$1,588 each on July 9, raising about HK$31.4 billion, and launched its "Touch High" plan on July 11. - DeepSeek closed a record $7 billion first round in June. Bloomberg reported that its second round (about $71 billion valuation) was paused on July 25 and restarted in early August; the company is also preparing a STAR Market listing.
3. The Bottleneck Moves from Chips to Power: Big Four Cloud Capex Reaches About $725B in 2026, Up About 77%
The bottleneck in compute is moving from chips to power. SpaceX, which acquired xAI in February, bought mobile-generation company APR Energy for more than $1 billion to supply power for Musk's AI compute. APR Energy runs over 1GW of gas turbines, and the deal has won final clearance.
Compliance risk is rising alongside. The NAACP and environmental groups have filed a Clean Air Act suit over up to 35 gas turbines alleged to be unpermitted; the allegations are pending in court and xAI has not accepted them. TechCrunch reported on July 31 that SpaceX will not remove all of the turbines for another year (Reported).
The other face of the power constraint is capex, and July earnings put it on the income and cash-flow statements (Alphabet and Meta figures per company disclosure):
| Company | Capex | Notes |
|---|---|---|
| Alphabet (July 22 results) | Q2 $44.9B; full-year guidance raised to $195-205B | Quarterly free cash flow turned negative at -$5.9B; stock fell about 7% the next day |
| Meta (July 29) | Full-year guidance raised to $130-145B | Stock also under pressure |
| Big four hyperscalers combined | About $725B in 2026 | About $410B in 2025, up about 77% |
Revenue is growing too: Microsoft's AI business within Azure reached a $37 billion annualized run rate, up 123% year on year. 'Power is compute' is turning from a slogan into a real constraint on the financial statements.
People Are No Longer the Main Token Consumers: Agents Use About 5x as Much, and Over 85% Is Cached
Verified (a16z, published August 21, 2026, using data from openrouter.ai/rankings): on OpenRouter, agentic token consumption reached 7.3 trillion on a seven-day average, roughly five times what human users consume directly. The agentic line first crossed above human usage on February 6, 2026, and grew about 14x over the following six months.
| Use | Token consumption | Basis |
|---|---|---|
| Agentic | 7.3 trillion, seven-day average | Verified |
| Human | About 1.3-1.4 trillion over the same period | Our estimate |
| Mixed | About 1.3-1.4 trillion over the same period | Our estimate |
Our estimate is read off the scale of the a16z chart, not a disclosed figure. It shows relative magnitude only and should not be cited as precise data.
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 relative growth. An agent runs a read-write-execute loop again and again and keeps context across steps, so the initial prompt, loaded with a codebase style guide, company policy or task brief, gets counted over and over.
Three revisions to this report's model
First, decouple inference demand from user counts. This report modelled token demand from DAU and calls per user. With agents now the main consumer, the real drivers are agent runtime × loop count × context length. A user-growth model will systematically understate compute demand.
Second, token growth does not turn into vendor revenue one for one. Cached prompts are priced well below fresh input tokens. When more than 85% of the increment is cached, backing revenue out of total token volume overstates it badly. Monetization analysis should split cached and non-cached volume.
Third, the bottleneck moves from compute to memory. This matters most for infrastructure. Cache-heavy workloads keep the KV cache resident for long periods, shifting the constraint from FLOPs to memory capacity and bandwidth. That runs against the instinct to answer a compute crunch with more GPUs; the binding constraint is more likely on the memory side. Read this together with this report's compute-infrastructure conclusions.
Cuts Land on the Most Substitutable Tier: GPT-5.6 Luna Fell 80% in Three Weeks While Flagship Sol Held
Verified (CNBC, Axios, VentureBeat and others, July 30, 2026): OpenAI repriced the GPT-5.6 family just three weeks after its July 9 launch.
| Tier | Before (input/output, $ per million tokens) | After | Change |
|---|---|---|---|
| Luna | 1 / 6 | 0.20 / 1.20 | Down 80% |
| Terra | 2.50 / 15 | 2 / 12 | Down about 20% |
| Sol (flagship) | 5 / 30 | Unchanged at the time; promotional 4 / 20 from August 21 | Unchanged at the time |
Verified (official API docs and releases, August 12-13): DeepSeek shipped the V4 Pro release build and signalled price increases across its lineup; Grok 4.6 entered at $2 per million input tokens and $6 output.
The split runs inside a single vendor
This report earlier summed up August as one lab raising prices while another cut them. That was too coarse. Adding the July 30 repricing, the split runs inside a single vendor: the cheap tier lost 80% of its price in three weeks while the flagship did not move. Price pressure lands on the most substitutable tier. Capability gaps among small models have narrowed to the point where users pick on price; the flagship tier is still hard to replace and keeps its pricing power.
Three falsifiable tests
If this structure holds, three things should follow:
- Low-tier prices keep converging across vendors, while flagship spreads widen.
- Vendors shift disclosure from model capability to per-tier use cases, to justify the tiers.
- Application-layer cost curves track model choice closely, so a single average token price misstates cost.
September data revised this section's wording: the frontier tier as a whole has plateaued, and same-price upgrades and selective increases have replaced cuts. The divergence moved inside the frontier tier, to selective increases and high prices for new entries, plus expiry dates on workhorse-model promotions; time became a new pricing tool. In October, the unit of price competition started to split from per million tokens into three dimensions: capability, speed and caching (see Latest developments).
Latest developments, newest first (data current as of 2026-10-11)
- 2026-10-08 · OpenAI launched an Ultrafast mode for GPT-6.1 Sol, which it says runs up to 8x faster than standard mode. API pricing is $12 per million input tokens and $60 output, 6x the model's standard rate ($2/$10) and above the standard rate of flagship GPT-6 Astra ($10/$50) · Verified (OpenAI API pricing page and launch announcement, reported by iFeng Tech and others)
- 2026-10-07 · Anthropic released Claude Haiku 5.5. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 output, versus $1/$5 for Haiku 4.5, a 90% cut in list price; the company says average running cost falls about 75%. The price matches the $0.10/$0.50 listed for GPT-6 Luna on OpenAI's pricing page · Verified (Anthropic website)
- 2026-10-07 · The same day, Anthropic halved Sonnet 5.5's cache-read price to $0.10 per million tokens and said most agentic tasks get about 20% cheaper as a result · Verified (Anthropic website)
- 2026-10-06 · SpaceX is seeking about $40 billion of financing to buy Nvidia chips, including about $10 billion in bank loans and $30 billion in investment-grade debt, with Apollo and several banks in talks · Reported (Financial Times, Reuters, CNBC and The Wall Street Journal, October 6-7, relayed by TrendForce and Yahoo Finance)
- 2026-10-06 · Broadcom is arranging more than $50 billion of financing for OpenAI's custom AI chips and has approached Apollo and Blackstone; talks are early and the size could change · Reported (Financial Times and others, October 6-7)
- 2026-09-28 · A draft prospectus seen by Reuters and the Financial Times shows Anthropic has about $518 billion in commitments for cloud services, compute and infrastructure over the coming years; its 2025 compute and infrastructure spending was $7.33 billion, roughly triple the prior year · Reported (September 28-29, relayed by Fortune and The Decoder)
- 2026-09-28 · Anthropic released Sonnet 5.5 at $2 per million input tokens and $10 output, half the price of Opus 5.5 ($4/$20). The company says total cost per task falls by up to 30%, mainly because the model uses fewer tokens and tool calls · Per company disclosure (Anthropic website, VentureBeat)
- 2026-09-27 · OpenRouter weekly token consumption reached 146.0 trillion, up 13.2% week on week and another record. Chinese models took 10 of the top 20 slots with 71.87 trillion tokens, 63.0% of the top-20 total. The top three were DeepSeek V4.1 Flash (19.60 trillion), GLM 5.3 Flash (16.30 trillion) and the anonymous "Space Bunny" (13.90 trillion; tests suggest it may come from MiniMax, unconfirmed) · Reported (Yilan Business, September 28)
- 2026-09-24 · Akamai signed a seven-year, $11.6 billion compute agreement with Anthropic for its growing CPU workloads, expandable by $9 billion to about $20 billion. Akamai plans about $5.5 billion of capex for it and issued Anthropic a warrant for up to about 5% of its outstanding shares at $111.33 per share; about 2% vests with the current commitment and about 3% on expansion milestones · Per company disclosure (Akamai press release)
- 2026-09-10 · DeepSeek released V4.1-Flash, a 552-billion-parameter MoE model on a new Causal Encoder-Decoder architecture that activates 8 billion parameters on input and 16 billion on output. KV-cache demand on HBM falls to one quarter of the previous generation and on SSD to one eighth. It cut API prices and added peak/off-peak pricing, with off-peak at half the peak rate · Verified (DeepSeek official announcement)
- 2026-09-10 · Microsoft plans to raise global data-center capacity from about 12 gigawatts to more than 38 gigawatts by 2032, counting owned and leased sites and excluding capacity rented from neoclouds such as CoreWeave. The report cites people familiar with the plan; Microsoft has not confirmed the figure in any filing or announcement, and the roadmap could change · Reported (Bloomberg; relayed by Dataconomy on September 11)
- 2026-09-10 · As of that day, weekly token calls on OpenRouter were up 1,863% from the start of the year; the three most-called models were Hy4-preview, GPT-5.6 Luna and GLM 5.3 Flash · Reported (Zhongtai Securities media and internet team, September 11)
- 2026-09-03 · BenchLM's frontier token price index read 16 for September (March 2023 = 100), flat with July and August, a third straight month without a decline. The median blended price across 21 frontier models is $6.00 per million tokens, 84% below the base period. The index covers 40 models (21 frontier, 11 mid-tier, 8 budget). In the period, Claude Sonnet 5's blended price rose from $4 to $6, up 50%, and Claude Fable 5.1 entered the frontier tier at $20 blended · Public market data (BenchLM Token Price Index)
- 2026-09-02 · Google released Gemini 3.8 Flash and 3.8 Flash Cyber, a restricted-access variant for cybersecurity. API pricing keeps 3.7 Flash's introductory rate of $0.75 per million input tokens and $3.75 output through December 31, 2026; from January 1, 2027, the standard rate of $1.50/$7.50 applies · Verified (Google blog)
- 2026-09-01 · Anthropic signed a $35 billion, six-year compute agreement with Nvidia-backed cloud provider Lambda for about 350 MW at Hut 8's Beacon Point campus in Nueces County, Texas, expected online in the first quarter of 2027. Nvidia holds the 15-year site lease; Lambda buys and installs the GPUs and resells the compute to Anthropic. Hut 8 has disclosed that the two-phase lease totals 704 MW with a base contract value of $19.6 billion. With the earlier $45 billion, six-year Nscale agreement (460 MW), Anthropic's named compute contracts total at least $80 billion. The September 1 report said term and capacity were undisclosed; the six-year term and 350 MW come from the September 2 report · Reported (first by The Wall Street Journal, with Yahoo Finance follow-ups on September 1 and 2)
- 2026-08-27 · Lambda closed a $926 million senior secured term loan · Reported (The Wall Street Journal, via Yahoo Finance)
- 2026-08-12 to 13 · DeepSeek shipped the V4 Pro release build and signalled price increases across its lineup (Verified, official API docs and releases). Financial media reported in mid-August that DeepSeek announced an across-the-board API price increase, the first time a leading Chinese lab raised prices at the same moment its capability jumped · Reported (financial media)
- 2026-08-12 · SpaceX AI's Grok 4.6 entered at $2 per million input tokens and $6 output, roughly half of comparable frontier models · Verified (SpaceX AI)
- 2026-08-10 · Nvidia signed memoranda of understanding with Apollo Global Management, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to build an independent financing platform meant to mobilize more than $500 billion of third-party capital over time for data-center construction and hardware purchases by cloud providers and AI companies. Jensen Huang said AI-factory compute is becoming an investable asset class, with demand from real commercial workloads and each project diligenced independently. Nvidia shares closed down 2.86% that day · Verified (Nvidia company statement)
- 2026-07-15 · OpenAI's first hardware product, the $230 Codex Micro, went on sale and sold out in about 12 hours; first deliveries began July 24. eBay resales ran around $800, with listings up to $1,850 · Reported
- 2026-07-14 · SoftBank became Sierra's exclusive partner in Japan; pilot resolution rates rose from 83% to 97%, as enterprise token spend moves toward outcome-based pricing · Reported
- 2026-07-10 · Apple sued OpenAI and io for trade-secret theft; the allegations are one-sided and OpenAI denies them · Reported
- 2026-06-01 · Anthropic confidentially filed its S-1, led by Goldman Sachs, JPMorgan and Morgan Stanley, with more than $60 billion expected to be raised; its May Series H valued it at about $965 billion. Its Nasdaq listing has reportedly been pushed to November · Reported
What This Means for Founders
Model costs by tier, speed and caching: in one October week Haiku 5.5 cut short-prompt list prices 90% while GPT-6.1 Sol's Ultrafast mode cost 6x standard. Do not build gross margin on upstream prices always falling; BenchLM's frontier index was flat from July to September, and Gemini 3.8 Flash's introductory price expires on December 31, 2026. Design agent products around caching: over 85% of agentic tokens come from cached prompts, and Sonnet 5.5's cache-read price has been halved. The signal to watch is whether vendors start quoting total cost per task; Anthropic says Sonnet 5.5 cuts it by up to 30%.
What This Means for Investors
Rising token volume does not mean vendor revenue rises with it: over 85% of agentic tokens come from cached prompts, which are priced lower. Anthropic reportedly carries about $518 billion in compute commitments, and chip purchases are reportedly turning to large debt financings, so financing costs may feed into token costs. Inference costs now need separate GPU and CPU lines: Akamai has signed a seven-year, $11.6 billion compute deal with Anthropic for CPU workloads, and the driver of lower cost is shifting from price wars to architecture. Watch whether SpaceX's roughly $40 billion chip debt financing closes at full size, and whether OpenRouter's weekly growth holds above 5%. This is not investment advice.
Where We Could Be Wrong
- 01If OpenRouter's weekly token consumption grows less than 5% week over week for four straight weeks, or the Chinese share of the top 20 falls below 50%, our calls on token growth and open-model share need to be revisited.
- 02If SpaceX's roughly $40 billion and Broadcom's $50 billion-plus chip financings end up materially smaller or are shelved, our view that compute demand is inelastic needs to be revisited.
- 03Gemini 3.8 Flash's introductory price runs through December 31, 2026. If Google does not restore the $1.50/$7.50 standard rate from January 1, 2027 and cuts the price further instead, our call that workhorse models now upgrade at the same price without further cuts should be withdrawn.
Key Questions
What share of global AI compute does inference take in 2026, and how fast will token consumption grow?
Inference now accounts for roughly two-thirds of global AI compute (Deloitte TMT Predictions 2026; about 1/3 in 2023, 1/2 in 2025). Goldman Sachs' May report projects monthly token consumption growing 24x to 120 quadrillion by 2030. Agentic traffic on OpenRouter surpassed direct human usage around early February 2026, with agent requests using about 15x the tokens of human chat. Data through 2026-08-01.
Which models cut prices in the July 2026 token price war, and to what levels?
The price war erupted in the last week of July 2026: OpenAI cut GPT-5.6 Luna by 80% to $0.20/$1.20 per 1M tokens and Terra by 20% to $2/$12 on July 30; DeepSeek shipped V4-Flash 0731 ($0.14/$0.28) on July 31; Anthropic released Opus 5 ($5/$25, adjustable effort dial) on July 24; Alibaba's Qwen3.7 Flash (July 27) and Google's Gemini 3.6 Flash (July 21) landed in the same window.
How are Chinese models performing in global open-source and token-share terms, and what did Kimi K3 open-source?
Moonshot AI open-sourced Kimi K3 weights on July 27, 2026: 2.8 trillion parameters and 1M-token context, the largest open-weight release in history, on its promised date. Chinese models hit a record 58% of U.S. companies' OpenRouter token usage in July (briefly 63% in the first week), holding all top-five spots. Bloomberg reported July 29 that Moonshot closed a $3.5B round at a $35B post-money valuation, planning an August pre-IPO round at up to $50B.
Watch & Listen
▶ A 'Barbell' Investment Logic Under AI Token DeflationYouTube · needs VPN in China
▶ What's Really Behind China's Free Open-Source Models?YouTube · needs VPN in China
▶ Where Should AI-Compute Money Go?YouTube · needs VPN in China
▶ How Tech Giants Monetize AIYouTube · needs VPN in ChinaIn 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: 26 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. The Endpoint War: Codex Micro's 12-Hour Sellout, 8x Resale Premium, and Apple's Trade-Secret Lawsuit
- 🔒5. Agent Commercialization: The SoftBank-Sierra Exclusive and the Outcome-Based Pricing Revolution
- 🔒6. Capital Markets: Anthropic's IPO Sprint with Opus 5 and the Near-Trillion-Dollar Benchmark
- 🔒7. Private Markets: Moonshot at $35B Post-Money, Zhipu's 'Touch High', and What DeepSeek's Paused Round Signals
- 🔒8. Value-Chain Map: Winners and Pressured Segments of the Inference Era (A Neutral Framework)
- 🔒9. Risks and Uncertainties: Power Litigation, Price Wars, Capex, and Valuation Reversal
Where we stand on this
Four moats, three screens, and outcomes-as-a-service pricing for AI application companies
Profit in this AI cycle lands in the application layer; the model layer cannot hold it. Application companies have four moats: taste, extreme efficiency in one workflow, several models combined, and process as product. Three screens decide whether they earn: distribution, retention, and gross margin that climbs as token prices fall. Outcomes-as-a-service is replacing seat subscriptions.
Full argument and falsification tests →The bubble sits in the middle layer that must keep refinancing: the FutureX seven-dimension bubble scorecard
The AI bubble sits in the middle layer that must keep refinancing: compute leasing on borrowed money, circular financing, GPU-backed debt. FutureX Capital has scored seven dimensions daily since March 2026; the August reading lit six, and liquidity triggered on August 17. The scorecard answers one question: which layer cannot be bought at the consensus price.
Full argument and falsification tests →Questions people ask next
Related Research
Artificial Intelligence / Macro & Capital
AGI Is Here, Capital Is Losing Its Bearings: Valuations Doubled Before Any Public Test
AI Coding
Cursor Sold for $60 Billion, but Coding Models Are Becoming Commodities
China AI Going Global
China AI Going Global: 23 Weeks Ahead on Usage, Not Yet on Pricing Power
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.