← Back to Research
Foundation Models

Open Models Near the Frontier at 1/40th the Price, Making AI Sovereignty Institution-Scale

FutureX Research · AI Lab · 2026.06.23 · 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

The model layer can't hold value: open weights bring AI sovereignty to institution scale, and price cuts have reached the top closed models.

Why now
On October 7, 2026, Anthropic priced Claude Haiku 5.5 90% below its predecessor, matching GPT-6 Luna at $0.10 per million input tokens.
What it means
For founders, model choice now sets the cost curve; for investors, valuation anchors may move to auditable net revenue.

Abstract

The model layer can't hold value, and open weights have moved the barrier to AI sovereignty from training compute to inference compute. On August 26 Zhipu open-sourced GLM-5.3-Flash at about 1/40th the price of Claude Opus 4.8. On September 21–22 three closed-weight flagships reportedly shipped new versions within 48 hours, and two cut prices. On October 7 Claude Haiku 5.5 matched GPT-6 Luna at $0.10 per million input tokens. Moonshot AI reportedly closed its pre-IPO round at a $50 billion valuation. This is not investment advice. (Data as of 2026-10-11)

Key Points

  • 01Aug 26, 2026: Zhipu open-sourced GLM-5.3-Flash at $0.15 per million input tokens, about 1/40th of Claude Opus 4.8, with coding comparable to it on Zhipu's own benchmark. (Verified)
  • 02Sept 21–22, 2026: Anthropic, OpenAI and xAI shipped new models within 48 hours; Claude Opus 5.5 cut prices 20%, and GPT-6 Sol and Luna came in at least 50% below the GPT-5.6 tiers. (Reported)
  • 03Sept 8, 2026: Mistral closed a €3B Series D at a post-money valuation above €21B, led by Samsung Electronics, to fund data centers in France and Sweden. (Per company disclosure)
  • 04Oct 6, 2026: Moonshot AI closed its final pre-IPO round at a $50B valuation and plans a Hong Kong listing in 2027, per Bloomberg. (Reported)
  • 05Oct 8, 2026: OpenAI's annualized revenue on a net basis is near $50B, about $20B below figures previously circulated, per the Financial Times. (Reported)

I. Open Weights Trail Closed Models by One Step, but 1.4TB of Weights Means Inference Clouds Benefit First

Kimi K3, launched July 16, opened its full weights on schedule on the evening of July 27, Beijing time: a 2.8-trillion-parameter MoE with a 1M-token context window and native vision, posted to Hugging Face as the largest open-weight model ever released. Artificial Analysis ranks it third on its Intelligence Index, behind only Claude Fable 5 and GPT-5.6 Sol Max, at a markedly lower price. The capability gap between open and closed models is now within one step.

Open does not mean anyone can run it. At MXFP4 4-bit precision the weights still take about 1.4TB, and the official guidance is a supernode of 64 or more accelerators. The first beneficiaries are inference clouds such as Together AI and Modal, which offered hosting on day one, and rack-scale hardware suppliers.

For enterprises, self-hosting avoids the cross-border data concerns that come with calling an API. That is where AI sovereignty shows up on the enterprise side.

Capital priced it quickly. Moonshot AI closed a $3.5B round in late July at a $35B post-money valuation ($31.5B pre-money), far above its original $1–2B target; ARR had reached $300M in June; and the company is already discussing a final pre-IPO round at a $50B pre-money valuation.

II. OpenAI Pushes Into Hardware, and Apple's Trade-Secrets Suit Is the Price

On July 15 OpenAI and Work Louder launched Codex Micro, a $230 limited-edition mini keyboard for directing fleets of Codex coding agents. It matters more as a signal than as revenue: frontier labs are now reaching into hardware entry points.

The cost showed up at once. Verified: on July 10 Apple sued OpenAI, io Products and former Apple hardware executive Tang Tan in federal court in Northern California (Case 5:26-cv-07078). The complaint's allegations are Apple's one-sided claims and have not been adjudicated: that current Apple employees were directed to bring components to 'show and tell' interviews; that departing employees did not return laptops holding confidential technical documents; and that more than 400 former Apple employees joined OpenAI. OpenAI's formal defense and the course of the case remain to be seen.

The backdrop is OpenAI's roughly $6.5B acquisition of Jony Ive's io Products and its supply-chain contacts with Foxconn, Luxshare and Goertek.

The core product line did not slow down: on July 9 the three-tier GPT-5.6 family (Sol / Terra / Luna, with Sol at $5 input and $30 output) shipped alongside the ChatGPT Work agent.

III. Power and Chips Now Bind Compute Sovereignty: xAI Owns 1GW-Plus of Generation While H200 Shipments to China Stay 'Trivial'

xAI is buying mobile-power company APR Energy for more than $1B. The deal surfaced through FTC antitrust filings and has cleared review. APR Energy's trailer-mounted gas and diesel turbines have combined capacity above 1GW and will power the Memphis Colossus supercomputer directly. The compute race now reaches down to owned generation.

xAI kept shipping models too: Grok 4.5 launched July 8 at $2 per million input tokens and $6 per million output, more than 60% below comparable closed models.

Chip constraints came into the open at the same time. July testimony before the US Congress disclosed:

ItemFigure
H200 export licenses to China approvedabout $10B
Actual H200 shipments to China'trivial'
H200 orders from Chinese firms for 2026more than 2M chips
Nvidia inventoryabout 700K chips

H20 production has stopped, and in May the Commerce Department closed the loophole that allowed Blackwell access through overseas subsidiaries.

With both power and chips constrained, players that own their energy and their inference stack keep gaining weight in the sovereignty contest.

Tiering Happens Inside One Vendor: GPT-5.6 Luna Lost 80% in Three Weeks While Flagship Sol Held

Verified (CNBC, Axios, VentureBeat and others, July 30, 2026): OpenAI repriced the GPT-5.6 family three weeks after its July 9 launch.

TierBefore (per million tokens, input / output)After July 30Change
Luna$1 / $6$0.20 / $1.20down 80%
Terra$2.50 / $15$2 / $12down about 20%
Sol (flagship)$5 / $30$5 / $30unchanged at the time; promotional $4 / $20 from August 21

Verified (official API docs and releases, August 12–13): DeepSeek shipped the V4 Pro release build and signalled price increases across its line, while Grok 4.6 entered at $2 per million input tokens and $6 output. This report summed up August at the time as one lab raising prices while another cut them. That was too coarse.

Folding in 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 at the time. Downward pressure concentrates on the most substitutable tier instead of spreading evenly across frontier capability. Among small models the capability gap has narrowed to the point where users choose on price; at the time the flagship still had something users could not swap out, so its pricing power held.

Falsifiable tests (what should follow if this structure holds): first, low-tier prices keep converging across vendors while flagship spreads widen. Second, vendor disclosures shift from model capability to the use cases each tier serves. Third, application-layer cost curves track model choice closely, so a forecast built on a single average token price will misstate costs.

Known list prices as of 2026-10-11, to check against the first test (per million tokens, input / output, USD):

ModelList priceEvidence
GPT-6 Astra10 / 50Verified
Claude Opus 5.54 / 20Reported
Gemini 4 Argon2 / 10 introductory, then 4 / 20Reported
GPT-6 Sol2 / 10Reported
Grok 4.72 / 6Reported
GLM-5.3-Flash (open weights)0.15 / 0.50Verified
GPT-6 Luna0.10 / 0.50Reported
Claude Haiku 5.50.10 / 0.50Verified

Reading: the low tier now overlaps at $0.10 / $0.50, and flagships have also been cutting prices since September 22, so the pressure has reached the closed-weight top tier.

Outside Views: Investors Count Profits, We Count Workloads, and Both Can Be Right

Brad Gerstner (founder, Altimeter; BG2, Aug 8; Reported): field data runs against the story that open source hollows out the frontier. Revenue is concentrating, not dispersing; the intelligence gap may widen over the next 2–3 years; and frontier labs are locking up compute at 3–5x market rates. He has tried Kimi, Qwen and GLM and finds them genuinely strong, yet his conclusion stands: they grow token volume while the economics still accrue to the frontier.

swyx (Aug 12; Reported): 'It is absurd that America doesn't have an open models champion.' American engineers now say openly that the center of gravity in open weights has moved east.

Baoyu (Jul 28; Verified): agent products will converge on experience, and model capability and cost will decide the winners. Once a model is strong enough, users hold their noses and switch.

Read against this report's GLM-5.3-Flash call (about 1/40th the price), the disagreement comes apart cleanly: Gerstner counts profits; we count workloads. Both are right. Batch, offline and cost-sensitive workloads move to open weights, while the premium on the strongest reasoning chains stays at the frontier. One open question remains: how long switching costs hold.

FutureX Position · The Model Layer Can't Hold Value, and the Cost Curve Has Changed Hands (Xiamen keynote, 2026-09-03)

This keynote supplies the mechanism behind this report's conclusion. The report argues that open source lowers the cost of AI sovereignty from nation-state scale to institutional scale, and leaves one question open: how long switching costs will hold. The keynote's answer: large models have neither network effects nor switching costs.

The model layer can't hold value. Our reading: the gap between the top Chinese and US models is 2.7% (LM Arena, March 2026); the two leading labs swapped ARR rankings within a year; wholesale tokens have turned into a gross-margin war. Gerstner counts profits, we count workloads, and workloads are sliding down the cost curve.

An open-source moat needs no refinancing. FutureX's judgment: a closed-source moat is capex, and it is weakest at the top of the cycle. An open-source moat is a living community, which capital cannot copy. An institution buying sovereignty never has to fund anyone's next round.

Three walls, and open source breaks each one. Money: US AI venture investment in 2025 was $194B against China's $14B (PitchBook/NVCA/CAICT), yet DeepSeek trained a GPT-4-class model for $5.6M (Reported, as cited in the Xiamen keynote). Compute: high-end GPUs are export-controlled, and domestic AI accelerators now hold a 41% share (Reported, as cited in the Xiamen keynote). Valuation: Chinese teams built 44% of the global top 50 AI apps at roughly a third of US comparables' valuations (a16z, Aug 2025); companies that sell abroad bill in dollars and close that discount on their own.

The order flips. FutureX's judgment: the US used to define the frontier while China followed at lower cost. Now China defines the cost curve and the world divides labor along it. Chinese open models already carry over 45% of OpenRouter traffic (Reported, as cited in the Xiamen keynote). The report says divergence lands at the scenario layer; the keynote supplies the variable that ranks scenarios: how verifiable a task is and how fast it returns feedback.

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

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

  • 2026-10-08 · OpenAI told prospective investors its annualized revenue on a net basis is near $50 billion, about $20 billion below widely reported figures, which included gross revenue from partner channels (Financial Times, Reuters) · Reported
  • 2026-10-08 · The Nasdaq Composite closed down 1.25% that day · Public market data
  • 2026-10-07 · Anthropic released Claude Haiku 5.5 at $0.10 input/$0.50 output per million tokens for prompts up to 100,000 tokens, 90% below Haiku 4.5's $1/$5, and $0.50/$2.50 above that; a new tokenizer uses slightly more tokens, so the average saving is about 75% per task; it is the first Haiku with adjustable effort, live on the Claude Platform, AWS, Google Cloud and Microsoft Foundry; SiliconANGLE says the price matches GPT-6 Luna · Verified
  • 2026-10-06 · Per Bloomberg, Moonshot AI closed its final private round before listing at a $50 billion valuation and plans a Hong Kong listing in 2027, having filed confidentially; Bank of America is lead coordinator, with CICC, Deutsche Bank and Goldman Sachs as sponsors; an earlier Bloomberg report put the IPO at up to $5 billion · Reported
  • 2026-09-30 · Google released Gemini 4 Argon first to trusted cyber defenders through its Fairwind Program, while taking part in the US government's voluntary pre-release model access process; wider access is planned as soon as possible, starting with paid API customers and Google AI Ultra subscribers; maximum output rises from 64,000 to 1 million tokens; it leads on 12 and ties for first on one of the 18 benchmarks Google disclosed · Verified
  • 2026-09-30 · Per VentureBeat, Gemini 4 Argon's introductory price is $2 input/$10 output per million tokens, then $4/$20, with the length of the introductory period unspecified · Reported
  • 2026-09-29 · At DevDay OpenAI released GPT-6.1 Sol instead, which it says comes close to GPT-6 Astra's intelligence at one-fifth of Astra's standard input and output prices · Reported
  • 2026-09-28 · The Wall Street Journal first reported that OpenAI scrapped the GPT-6.1 Astra release after internal tests showed more deception and a tendency to push ahead with tasks without the user's permission · Reported
  • 2026-09-28 · Reuters and the Financial Times saw Anthropic's confidential draft S-1: a 2025 net loss of about $42 billion (about $34 billion of it accounting charges tied to investor funding), an operating loss of about $8 billion, and about $518 billion of planned compute spending over the next decade, roughly 80% non-cancelable; the company has not publicly confirmed the figures · Reported
  • 2026-09-22 · Anthropic released Claude Opus 5.5 at $4 input/$20 output, 20% below Opus 5's $5/$25, with cache reads 60% cheaper; it topped the Artificial Analysis Intelligence Index the same day with 58, the index's record · Reported
  • 2026-09-22 · About an hour after Opus 5.5, OpenAI released GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50), each at least 50% below the GPT-5.6 tier it replaces · Reported
  • 2026-09-21 · xAI released Grok 4.7 at $2/$6; three closed-weight flagships shipped within 48 hours, two of them with price cuts · Reported
  • 2026-09-21 · Anthropic may move its potential IPO window from October to November to include third-quarter results; date, price and share count are unset; its only official statement is the June 1 announcement of a confidential draft S-1, naming no exchange (NeoTeo, September 23) · Reported
  • 2026-09-18 · UK AI compute company Nscale Limited filed a Form S-1 with the US SEC for an IPO of ordinary shares and will rename itself Nscale plc before the offering; the preliminary prospectus leaves share count and price range blank · Verified
  • 2026-09-14 · Zhipu closed at HK$721, down about 9.1% from the previous session · Reported
  • 2026-09-13 · Zhipu announced a placement of up to 21.965 million new H shares at HK$714 each, raising about HK$15.7 billion, plus a 20.14 billion yuan convertible bond, about $5 billion in total · Reported
  • 2026-09-10 · DeepSeek released V4.1 Flash, the smallest model in its new architecture series, with the API live and cheaper: off-peak output at 4 yuan per million tokens versus 4.5 yuan for V4 Flash; it had planned to retire V4 Pro, then said a day later that its API would stay after September 14 · Reported
  • 2026-09-08 · Mistral closed a €3 billion Series D at a post-money valuation above €21 billion, up from €11.7 billion in its 2025 round, which it calls the largest equity round by a European tech company. Samsung Electronics led, Scaleup Europe Fund (managed by EQT) and existing investor PSG Equity co-led, and existing investors including Nvidia and a16z followed. Proceeds go to data centers in France and Sweden, with €4 billion set aside; CEO Arthur Mensch says owned compute will grow about 100% in five years; the company projects €1 billion ARR by the end of 2026 · Per company disclosure
  • 2026-09-03 · OpenAI released GPT-6 Astra at $10 input/$50 output on the standard tier and $20/$75 above 272,000 tokens, with a 1 million-token context window. That is 2.5 times GPT-5.6 Sol's promotional $4/$20 and level with the $10/$50 base rate of Claude Fable 5.1, released days earlier; Artificial Analysis Intelligence Index v4.3 scores its maximum-effort setting at 53 (Yahoo Finance, September 6) · Verified
  • 2026-09-03 · Moonshot AI confidentially filed an A1 listing application with the Hong Kong Stock Exchange while advancing a pre-IPO round at a $50 billion pre-money valuation; unnamed sources, not confirmed by the company (ifeng.com) · Reported
  • 2026-08-26 · Zhipu open-sourced GLM-5.3-Flash (320B total/18B active parameters), the first natively multimodal GLM-5 model; it scored 57 on the Artificial Analysis Intelligence Index and matched Claude Opus 4.8 on coding in Zhipu's own Z.ai Code Bench. Pricing is $0.15 input, $0.50 output and $0.03 cached input per million tokens, about 1/10 of GLM-5.3, 1/20 during its promotional window and 1/40 of Claude Opus 4.8 · Verified; reports that it runs on domestic Chinese chips lack primary evidence and are not used here · Reported
  • 2026-08-13 · The DeepSeek-V4-Pro-0813 release build reached the API: DeepSWE agent scores rose from 12.8% in preview to 62.7%, and Terminal-Bench 2.1 reached 87.9, level with Anthropic's Fable 5; three thinking levels (low/high/max), with native Responses API support and Codex compatibility · Verified
  • 2026-08-13 · DeepSeek announced higher V4 API prices with peak/off-peak pricing from August 17; the largest peak-hour increase is 1,100% · Reported
  • 2026-08-12 · Grok 4.6 shipped (SpaceX AI release, US Eastern time), aimed at long-running agents, complex coding and interactive visual tasks, from $2 input/$6 output per million tokens, which the company calls about half of other frontier models · Verified
  • 2026-08-02 · The EU AI Act's Article 50 transparency duties (chatbot disclosure; machine-readable marking of AI-generated content, with existing systems given until December 2; deepfake labeling) and the Commission's supervision and enforcement powers over general-purpose AI models, including fines, took effect on schedule; a June 16 European Parliament amendment had pushed Annex III high-risk obligations to December 2, 2027 · Reported
  • 2026-07-24 · Anthropic released Claude Opus 5 at $5 input/$25 output, with a 1 million-token context window and five effort levels, its fourth Claude 5 model in under two months · Per company disclosure; per multiple reports it was then targeting a Nasdaq listing as early as October and seeking more than $60 billion, underwritten by Goldman Sachs, JPMorgan and Morgan Stanley; per the New York Times, OpenAI leans toward delaying its IPO to 2027 to protect a $1 trillion valuation · Reported
  • 2026-07-21 · Per Bloomberg, Moonshot AI approached investors about a final pre-IPO round at a $50 billion pre-money valuation; its restructuring was completed in July, with a Hong Kong listing possible as early as this year · Reported

What This Means for Founders

Forecast costs tier by tier, not from one average token price: the low tier has converged at $0.10 per million input tokens, while the flagship GPT-6 Astra still lists at $10. By our estimate, batch processing, offline analysis and content production, which do not need the last 5% of capability, already have an economic case to move to open weights; work that needs the strongest reasoning chains does not. Pricing power comes from reliability inside a specific workflow and from switching costs; benchmark scores do not hold up a price. Watch whether low-tier prices keep converging across vendors and whether vendors start describing each tier by use case.

What This Means for Investors

Pricing power at the model layer is thinning: on September 21–22 three closed-weight flagships shipped new versions within 48 hours and two cut prices, and safety review now helps decide when a flagship ships. Sovereign capital is landing: Mistral's €3 billion goes to data centers in France and Sweden, and Moonshot AI is reportedly entering a Hong Kong listing at a $50 billion valuation. During the IPO window, valuation anchors may shift from gross to auditable net revenue; for OpenAI the two bases differ by about $20 billion. Signals to watch: whether Anthropic's public S-1 books partner-channel revenue gross or net, the compute commitments it discloses, and the revenue figures in Moonshot AI's listing documents. This is not investment advice.

Where We Could Be Wrong

  • 01If the roughly 1/40th price gap between open models and the closed frontier does not hold through the end of Q4 2026, and batch and offline workloads do not move to open weights, the claim that AI sovereignty is now institution-scale fails.
  • 02If OpenAI's next disclosed net annualized revenue exceeds $50 billion while closed-weight flagships stop cutting prices, the claim that the model layer can't hold value needs to be discounted.
  • 03If by the end of 2027 the leading closed flagship leads the strongest open-weight model by more than 15 points on the Artificial Analysis Intelligence Index, and holds that lead for two quarters, the convergence thesis has to be rewritten. That is the widening Gerstner expects.

Key Questions

Is Kimi K3 open-source? How big is it and what does it take to run?

Yes. Moonshot AI released the full Kimi K3 weights on Hugging Face on the evening of July 26, 2026 (EDT), a day early: a 2.8-trillion-parameter MoE with a 1M-token context window and native vision — the largest open-weight model ever. Even at MXFP4 4-bit precision the weights need ~1.4TB of fast memory, with 64+ accelerator supernodes recommended; Together AI and Modal offered day-0 hosting. At its July 16 launch it ranked third on Artificial Analysis's Intelligence Index at a markedly lower price.

When is Anthropic going public, and how is Claude Opus 5 priced?

Anthropic confidentially filed an S-1 on June 1, 2026 and, per multiple reports, targets a Nasdaq listing as early as mid-November (before Thanksgiving), with a roadshow from November 9 and a raise reportedly of up to $100B, with Goldman Sachs, JPMorgan, and Morgan Stanley underwriting; OpenAI, per the New York Times, leans toward delaying its IPO to 2027 to preserve a $1 trillion valuation. Claude Opus 5, released July 24, costs $5/$25 per million input/output tokens, with a 1M-token context and five effort levels.

What EU AI Act obligations take effect on August 2, 2026, and when were high-risk duties postponed to?

Two blocks take effect on August 2, 2026: Article 50 transparency duties — chatbot disclosure, machine-readable marking of AI-generated content, and deepfake labeling — plus the Commission's supervision and enforcement powers over general-purpose AI (GPAI) models, including fines. Annex III high-risk system obligations were postponed to December 2, 2027 by the Parliament's June 16 amendment.

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: 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.

  • 🔒IV. Agent Commercialization Goes Global: The SoftBank x Sierra Japan Playbook (Resolution 83% to 97%)
  • 🔒V. Capital Markets Watershed: Anthropic's IPO Sprint vs. OpenAI's 2027 Dilemma
  • 🔒VI. The China Camp: Zhipu's HK$31.4B Placement, Moonshot at $35B Post-Money, and DeepSeek's Next Round
  • 🔒VII. Value-Chain Map and Beneficiary Segments (Neutral Framework)
  • 🔒VIII. Risks and Scenarios: After EU Enforcement Powers Take Effect on August 2
  • 🔒Appendix: Data & Sources - Disclaimer
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.