The AI model business, explained: who builds, who gives away, who pays

Mistral AI put Mistral Large 4 into public preview on Oct. 6, describing it as a model of about 1 trillion parameters and promising the weights "by the end of the month." The model's page on Hugging Face, the developer platform where open models are posted, gives an expected release date of Oct. 31. The Paris company has not said which license the weights will carry. Until then, developers can reach the model only through Mistral's own API, where it is listed at a sale price of $0.68 per million input tokens and $2.09 per million output tokens, half the standard rate. The release capped a crowded two weeks for the companies that keep their models to themselves. Anthropic released Claude Sonnet 5.5 on Sept. 28 and Claude Haiku 5.5 on Oct. 7. OpenAI introduced GPT-6.1 Sol on Sept. 29 and pushed an update to its GPT-6 Sol and Luna models on Oct. 7. Google unveiled Gemini 4 Argon on Sept. 30 but is rolling it out only to vetted cyber defenders through what it calls the Fairwind Program. None of those companies lets outsiders download its newest models; all of them sell access by the token.

For bankers, the line between those two approaches matters more than the model names. A closed model runs only where its developer decides, which ties its compute bill to a short list of cloud providers and chip contracts. An open-weight model can run in any data center whose owner downloads it, which spreads spending across many more buyers of servers and sellers of capacity. Which camp wins a given workload decides who orders the chips, who rents out the machines and who keeps the margin on each answer a model gives. This guide relies only on what the companies themselves have published: model cards, launch posts, pricing pages, license texts, securities filings and earnings calls. Prices and model versions are a snapshot as of Oct. 11, 2026.

The closed camp is the larger by revenue and the more familiar. OpenAI, Anthropic, Google DeepMind and xAI keep their trained parameters, known as weights, on their own servers. Customers send text, images or code to an application programming interface, or API, or use a chat app, and pay per million tokens processed, a token being a word fragment a few characters long. OpenAI calls GPT-6 Astra, whose system card was published Sept. 3, its most intelligent model. Anthropic's top tier is Claude Fable 5.1, released in September, with Claude Opus 5.5 following on Sept. 22. Google's newest model, Gemini 4 Argon, remains in limited release, so its generally available workhorse is Gemini 3.8 Flash, whose model card is dated Sept. 2. xAI, which SpaceX acquired in February, made Grok 4.7 available on its API on Sept. 21. The open-weight camp publishes the weights for anyone to download, run and modify, under license terms that vary a great deal. DeepSeek, the Chinese lab, released DeepSeek-V4.1-Flash on Sept. 10. Alibaba's Qwen team began selling Qwen3.8-Max through its cloud on Aug. 3 and posted the weights of the underlying 2.4 trillion-parameter model on Aug. 12. Mistral's most recent general-purpose open release, Mistral Medium 3.5, came out in March, with Large 4 due this month.

Several of the biggest developers work both sides. Meta released Muse Glimmer, a 30 billion-parameter open model, on Aug. 10, while offering Muse Spark 1.1, a closed model, through a public preview of its Meta Model API; its Llama 4 models date from April 2025. Meta's annual report says it expects "to continue training a combination of open and closed models going forward." Google publishes its Gemma 4 family alongside closed Gemini. OpenAI released two open-weight models, gpt-oss-120b and gpt-oss-20b, in August 2025, and xAI has published the weights of its older Grok 2. "Open weight" is not the same as "open source," and the distinction matters in diligence. Open-source software traditionally means the full recipe is public. None of the labs covered here publishes the data its models were trained on. Some describe it in general terms: Google's Gemma 4 card cites web documents, code, images and audio with a January 2025 cutoff, and Meta's Muse Glimmer card cites "publicly available data, data provided by third parties and information from Meta's products and services." DeepSeek goes further than most by posting a technical report and inference code alongside the weights. A buyer still gets a finished model it can run and adapt, not the means to rebuild it.

The license is where the commercial terms sit, and they are far from uniform. The most permissive releases use standard software licenses that allow commercial use with few conditions. OpenAI's gpt-oss models, Google's Gemma 4, Meta's Muse Glimmer and Qwen's smaller Qwen3.8-27B ship under Apache 2.0, and DeepSeek-V4.1-Flash ships under the MIT license. Meta's Llama 4 license is the most prominent custom alternative. Any company whose products had more than 700 million monthly active users in the calendar month before the release must "request a license from Meta, which Meta may grant to you in its sole discretion." Anyone distributing a product built on the models must "prominently display 'Built with Llama,'" and anyone using Llama to build another model that is distributed must "include 'Llama' at the beginning of any such AI model name." The acceptable use policy that the license incorporates bars uses such as military applications, and it withholds the license grant for Llama 4's multimodal models from individuals and companies based in the European Union, though not from end users of products that contain them.

Alibaba's flagship weights carry a custom "Qwen3.8-Max License" that reads like MIT with two thresholds. A commercial product with more than 100 million monthly active users or more than $20 million in monthly revenue must "prominently" display the model name. A company that runs a model-as-a-service or "AI Work Assistant" business, and whose group revenue tops $50 million over any 12 months, must "obtain a separate license from Qwen" before using the model commercially, although purely internal use is exempt. Mistral sets its line lower. The "Modified MIT License" on Medium 3.5 withholds all rights from any company whose "global consolidated monthly revenue" exceeds $20 million, and directs such companies to buy a commercial license or use Mistral's hosted service. xAI's license for Grok 2 is "revocable," allows commercial use only within xAI's acceptable use policy, and bars using the model or its outputs "to train, create, or improve any foundational, large language, or general-purpose AI models." The pattern is consistent: small companies get permissive terms, while the largest platforms, which could turn a model into a rival product, face thresholds or a separate negotiation. In an acquisition that depends on a model, a user count or revenue line buried in a license can matter as much as any benchmark score.

Every lab publishes scores on standard tests when it launches a model, and every lab chooses which tests to publish. The figures below come from each lab's own model card or launch materials. Labs test under different settings, including how long a model may "think," which software harness runs it, whether it may use tools and whether every task or only a subset was attempted, so these scores are not a like-for-like contest. Rackline charted only benchmarks that at least three labs report for their own models. Two coding-agent tests clear that bar comfortably. Terminal-Bench 4.0, in which a model works at a computer's command line to complete multistep tasks, is reported by six labs, as is DeepSWE v1.1, a test of autonomous software engineering. Gemini 4 Argon has no model card, so its scores come from Google's Sept. 30 launch post and the Gemini models page on Google DeepMind's site.

On reasoning, the common reference points are GPQA Diamond, a set of graduate-level science questions, and Humanity's Last Exam, or HLE, a broad test of expert-level questions. OpenAI's launch page for GPT-6 Astra reports 96.0% on GPQA Diamond and 57.2% on HLE with tools, figures it says are the maximum at any effort setting. Qwen's model card for Qwen3.8-Max reports 92.6% on GPQA Diamond and 43.6% on HLE, rising to 56.2% with tools. DeepSeek's card for V4.1-Flash, at its highest reasoning setting, reports 90.9% on GPQA Diamond and 36.8% on HLE, or 63.9% with tools. Anthropic does not list GPQA results on the launch pages for its newest models. Its Sept. 22 page for Claude Opus 5.5 reports 67.7% on HLE with tools at maximum effort, and its Claude Fable 5.1 page reports 60.9% without tools and 65.0% with them; the Opus 5.5 page cites Fable's with-tools figure as 65.6%. Smaller open models report many of the same tests. Google's card for Gemma 4 31B lists 84.3% on GPQA Diamond and 89.2% on the 2026 AIME math competition, and Meta's card for Muse Glimmer lists 94.7% on AIME 2026. Among the labs' own model pages, only OpenAI's reports ARC-AGI results, giving Astra 99.9% on ARC-AGI-3 when run through OpenAI's own API harness with two settings changed; Rackline has not charted the test.

Coding is where the labs publish the most comparable figures. On Terminal-Bench 4.0, Anthropic reports 66.4% for Opus 5.5 at its "xhigh" effort setting and 55.8% for Fable 5.1. OpenAI reports 57.9% for Astra, and Google's Gemini models page reports 57.4% for Argon. xAI reports 37.6% for Grok 4.7, DeepSeek 31.2% for V4.1-Flash and Mistral 28.3% for the Large 4 preview. On DeepSWE v1.1, Google's launch post reports 77.9% for Argon, DeepSeek 74.2% for V4.1-Flash using the mini-SWE harness the benchmark's official setup calls for, OpenAI 74.1% for Astra, xAI 71.0% for Grok 4.7 at "high" rather than maximum effort, Mistral 61.7% for Large 4 and Qwen 56.6% for Qwen3.8-Max. Google's own comparison table lists other labs' Terminal-Bench 4.0 figures that differ from those labs' pages, putting Astra at 58.2% and Fable 5.1 at 57.9%; Rackline uses each lab's figure for its own model. SWE-bench Verified, once the standard coding test, is absent from the closed flagships' launch pages. It now appears mainly on open-model cards: Meta reports 76.0% for Muse Glimmer and Mistral reports 77.6% for Medium 3.5.

For agents, meaning models that operate software, browse or complete multistep business tasks on their own, the labs have converged on newer tests. AutomationBench, which simulates workflows across business applications, is reported by six labs. Mistral lists 59.9% for Large 4, DeepSeek 54.8% for V4.1-Flash, Google 51.3% for Argon and OpenAI 41.4% for Astra. Anthropic lists 40.0% for Opus 5.5 in a run it says Zapier conducted, and Qwen lists 27.3% for Qwen3.8-Max on a 600-task subset. OSWorld, which tests control of a desktop computer, now comes in more than one version and scoring mode, which makes the figures harder to set side by side. OpenAI reports 72.6% for Astra on version 2.0 in an offline, partial-credit setup, and Google reports 69.2% for Argon on the same basis. Anthropic reports 77.9% partial and 41.7% strict for Fable 5.1 on version 2.0, using a later task release that it says is not directly comparable with earlier results, and 81.8% partial for Opus 5.5 on version 2.1. Terminal-Bench Science, which applies the command-line format to scientific work, is reported by three labs: 64.6% for Astra on OpenAI's Astra page, 58.7% for Opus 5.5 and 57.6% for Argon. OpenAI's later post on GPT-6.1 Sol cites 68.1% for Astra on the same test.

On multimodal work, which covers image, chart and video understanding, the labs share fewer tests and mostly report ones of their own choosing. Google reports 91.7% for Argon on LVBench, a long-video test. On Chartography, a chart-reading test, Anthropic reports 89.0% for Opus 5.5 with tools and Google 71.6% for Argon. Three developers report MMMU-Pro, a college-level test of images and text, but only for open models: Google lists 76.9% for Gemma 4 31B, Meta a score of 74 for Muse Glimmer and DeepSeek 56.5% for the V4.1-Flash base model with a four-example prompt.

Each company's own pricing page gives list prices for its main models. As of Oct. 11, 2026, in dollars per million input and output tokens, they were as follows. Anthropic charges $10 and $50 for Claude Fable 5.1, with cache reads at $0.25; $4 and $20 for Claude Opus 5.5, with cache reads at $0.20; $2 and $10 for Claude Sonnet 5.5; and from $0.10 and $0.50 for Claude Haiku 5.5. OpenAI charges $10 and $50 for GPT-6 Astra at short context, rising to $20 and $75 for long context; $2 and $10 for GPT-6.1 Sol, with cached input at $0.10; and $0.10 and $0.50 for GPT-6 Luna. Google lists Gemini 3.8 Flash at $0.75 and $3.75 through Dec. 31, 2026, rising to $1.50 and $7.50 from Jan. 1, 2027. xAI charges $2 and $6 for Grok 4.7 under 200,000 prompt tokens and $4 and $12 above that. Qwen3.8-Max costs $2 and $6, with a 1 million-token context. Mistral lists the Large 4 preview at $1.36 and $4.18, on sale at $0.68 and $2.09. DeepSeek-V4.1-Flash costs $0.30 and $1.20 at peak, and $0.15 and $0.60 off-peak.

Some models are missing from that list for want of a published list price. Gemini 4 Argon is not yet on Google's API price list, although Google's Sept. 30 announcement gives an introductory price of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward. Meta has not published a price for Muse Spark 1.1 that Rackline could find. OpenAI does not serve its gpt-oss models through its own API, and Google offers Gemma 4 on its Gemini API only in a free tier. The spread is wide. OpenAI's and Anthropic's top-tier models, GPT-6 Astra and Claude Fable 5.1, carry the same $50 output price. DeepSeek charges $1.20 at peak and half that off-peak. Grok 4.7 and Qwen3.8-Max sit at $6, and Mistral's preview lists at $4.18 before its discount. The labs are also experimenting with how they charge. Cache discounts reward customers who resend the same long prompt, DeepSeek's time-of-day rates reward those who can shift work to quiet hours, and introductory prices such as Google's on Gemini 3.8 Flash are scheduled to rise.

Those prices matter to the chip trade because every token a customer pays for is a unit of inference, the work of running a trained model to answer a request. Unlike training, inference runs continuously and grows with use. Nvidia's latest quarterly report states that "AI clouds and AI model makers have significant demand for training and inference compute." Oracle said in its Sept. 10 results release that customer demand "for AI Cloud Training and Inferencing Services continues to grow faster than supply." Anthropic's own disclosures show the scale of paid usage at a single closed lab: it put its annualized run-rate revenue at $14 billion in February, more than $30 billion in April and more than $47 billion in May. Open weights change where that inference runs and who pays for the hardware. A company that downloads DeepSeek's or Qwen's weights buys or rents its own servers instead of paying a per-token price to the developer. The developers' own price lists keep moving in the meantime. DeepSeek is phasing out its larger V4 Pro model in favor of V4.1-Flash, which uses 8 billion active parameters to read input and 16 billion to write output out of 552 billion in total. Since Sept. 14 it has routed all V4 Pro requests to V4.1-Flash at the cheaper Flash rates, "until V4.1-Pro launches."

Building and serving frontier models has turned the labs into prolific private fundraisers, and much of the money returns to the hardware market as orders for chips and data centers. OpenAI closed its latest round on March 31 with $122 billion of committed capital at an $852 billion post-money valuation. That completed a raise the company had announced on Feb. 27 at $110 billion, made up of $30 billion from SoftBank, $30 billion from Nvidia and $50 billion from Amazon, at a $730 billion pre-money valuation. Amazon's $50 billion is a commitment. Its quarterly report for the period ended June 30 shows it invested $15 billion in OpenAI Series C preferred stock in the first quarter, agreed to buy $35 billion more and put $13.7 billion of that in during the second quarter, for $28.7 billion on its books at June 30. SoftBank co-led the final round with a16z, D. E. Shaw Ventures, MGX, TPG and accounts advised by T. Rowe Price, and Microsoft continued to participate. OpenAI said it raised more than $3 billion from individual investors through bank channels and expanded its undrawn revolving credit facility to about $4.7 billion.

Anthropic has raised twice this year. A $30 billion Series G on Feb. 12, led by GIC and Coatue, valued it at $380 billion post-money. A $65 billion Series H on May 28, led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital, valued it at $965 billion. Anthropic said the Series H "includes $15 billion of previously committed investments from hyperscalers, including $5 billion from Amazon." Amazon's second-quarter filing records two $5 billion purchases in the quarter: one in Anthropic Series G nonvoting preferred stock, and one in Series H nonvoting preferred stock made under a financing arrangement of up to $20 billion, which reduced the amount still available to $15 billion. xAI raised a $20 billion Series E on Jan. 6, above its $15 billion target, with Nvidia and Cisco Investments as strategic investors alongside Valor Equity Partners, StepStone, Fidelity, Qatar Investment Authority, MGX and Baron Capital. It did not state a valuation. On Feb. 2, SpaceX announced that it had acquired xAI.

Mistral is the only open-weight developer covered here with an announced outside round this year. It raised €3 billion in a Series D on Sept. 8 at a post-money valuation of more than €21 billion. Samsung Electronics led the round, with EQT's Scaleup Europe Fund and PSG Equity co-leading, and existing investors including Nvidia and ASML took part. In July, Microsoft expanded its partnership with Mistral, including what it called "a multibillion dollar commitment" to use Mistral's European GPU capacity. Meta, Alibaba and DeepSeek have not announced outside venture rounds for their model work.

Much of this capital is already spoken for by multiyear chip and cloud agreements, which makes the labs' fundraising an early signal of hardware demand. In September 2025, OpenAI agreed to deploy at least 10 gigawatts of Nvidia systems, and Nvidia said it intends to invest up to $100 billion in OpenAI as that capacity comes online. The following month, OpenAI agreed to deploy 6 gigawatts of AMD graphics processors in exchange for a warrant on up to 160 million AMD shares that vests with deployment milestones. Also in October 2025, it agreed with Broadcom on 10 gigawatts of accelerators of its own design, and in January it signed for 750 megawatts of capacity from Cerebras. On April 20, Anthropic agreed to spend more than $100 billion over 10 years on Amazon Web Services technologies for up to 5 gigawatts of new capacity on Amazon's Trainium chips; Amazon invested $5 billion that day, with up to $20 billion more possible, on top of the $8 billion it had already put in. On April 6, Anthropic agreed with Google and Broadcom on "multiple gigawatts" of tensor processing unit capacity expected from 2027, which its Series H announcement later put at five gigawatts.

The closed labs, in other words, are financing their own capacity at gigawatt scale. Mistral's round shows money reaching open-weight developers too, with chip and equipment makers among the backers on both sides.

The suppliers' filings show how each one's revenue depends on which camp wins the work. Nvidia's latest quarterly report, for the period ended July 26, ties its demand to both. "The demand for open-source foundation models and applications promotes use of our products worldwide," it says, and it warns that limits on supporting "applications built on foundation models originating in China such as DeepSeek, Qwen, or Kimi, could have a material impact on our business." Nvidia also develops open models of its own, including Nemotron. On the closed side, it agreed in August to guarantees capped at $105 billion covering leases for about 4.25 gigawatts of data centers in Ohio that will host its systems for OpenAI. A single direct customer accounted for 16% of second-quarter revenue, while H200 shipments under U.S. licenses were "less than 1% of Data Center revenue."

Microsoft is bound to the closed camp through equity and contracts. Its fiscal 2026 annual report says it holds "an approximate 25% interest on an as-converted basis" in OpenAI and recorded $24.1 billion of revenue "from commercial arrangements with OpenAI, inclusive of revenue-sharing payments." Yet Azure sells models from both camps: on the July 29 earnings call, Chief Executive Satya Nadella said it offers "over 11,000 models, including the latest from OpenAI, Anthropic, Mistral, xAI, as well as our own MAI family." Chief Financial Officer Amy Hood added that "All sequential commercial RPO growth was driven by commitments from customers outside of frontier model companies," referring to remaining performance obligations, or contracted revenue not yet recognized. Alphabet sells both kinds of model and now sells its chips, too. On its July 22 call, Alphabet said the Gemma open models "have been downloaded over 900 million times." Chief Financial Officer Anat Ashkenazi raised 2026 capital spending guidance to "$195-205 billion" and said Google had delivered TPU systems to customers' own data centers for the first time in the second quarter, adding that investors would "see the vast majority of the revenue from that agreement come through in 2027."

Amazon's exposure runs through both closed labs. Its second-quarter filing says it invested $8 billion in Anthropic convertible notes from 2023 through 2025, part of which has since converted to preferred stock, and records the two $5 billion preferred purchases and an Anthropic commitment expanded "by more than $100.0 billion over 10.0 years, which includes contractual obligations related to the performance of AWS chips." It also records the OpenAI preferred stake and a $100 billion, eight-year expansion of OpenAI's AWS commitment. Amazon's Bedrock service carries Google's open Gemma 4 models alongside the closed ones. Meta builds compute for its own use rather than selling it, and expects 2026 capital spending, including finance-lease principal, of "$130-145 billion." On the July 29 call, Chief Executive Mark Zuckerberg said open source "creates its own feedback loops that are positive for us." Asked whether Meta could rely on outside open models, he said "right now the open source models are not as strong as the frontier models. So no is the basic answer."

AMD's disclosed commitments sit with the closed labs and Meta. The OpenAI warrant vests in tranches starting "after the delivery of the initial one (1) gigawatt" of its MI450 chips, with stock-price targets that "escalate to $600 per share for the final tranche," and a February filing gives Meta a 160 million-share warrant on similar terms. As of June 27, "no warrant tranches had vested or become exercisable." AMD's second-quarter release added a partnership with Anthropic to deploy up to 2 gigawatts of MI450 chips.

Broadcom, which designs custom accelerators, cites "unprecedented demand" from customers that it says include frontier AI labs, without naming them or mentioning open-weight models. It estimates its five largest end customers made up about 55% of third-quarter revenue, and Chief Executive Hock Tan said Broadcom expects fourth-quarter AI semiconductor revenue "to accelerate to $21.7 billion, up 236% year-over-year." Oracle booked more than $30 billion of additional AI cloud contracts in its fiscal first quarter, lifting remaining performance obligations to $664 billion, though its quarterly filing names no model developer as a customer. CoreWeave's quarterly filing names Microsoft and OpenAI among its significant customers and discloses a Meta commitment of up to about $21 billion; its three largest customers accounted for roughly 36%, 26% and 10% of second-quarter revenue.

Read together, the filings describe a concentrated set of buyers. The closed labs and a few hyperscalers sign the multigigawatt contracts that fill order books at Nvidia, AMD, Broadcom, Oracle and CoreWeave. The open-weight camp shows up less in contracts than in risk factors and download counts, such as Nvidia's China language and Alphabet's Gemma figure. A shift of workloads toward open models would spread inference spending across more, smaller buyers, the broad demand Nvidia's filing ties to its products. Suppliers that disclose heavy reliance on a handful of customers, Broadcom and CoreWeave among them, are more tied to the closed labs' build-out.

Next on the calendar: Oracle hosts its AI World conference in Las Vegas from Oct. 25 to Oct. 28. Microsoft reports results for its fiscal first quarter on Oct. 28; on the July call, Hood put calendar 2026 capital spending at "approximately $175 billion." Oct. 31 is the expected date, according to Mistral's Hugging Face page, for the Large 4 weights, and the license that comes with them will show whether Mistral's largest model follows the revenue cutoff on Medium 3.5 or a different set of terms. AMD reports third-quarter results on Nov. 3, its first set of numbers since disclosing the Anthropic partnership, and any vesting of the OpenAI or Meta warrants would surface in its filings. On Jan. 1, 2027, Google's introductory price on Gemini 3.8 Flash ends, lifting the output price from $3.75 to $7.50 per million tokens. Several other milestones have been announced without dates. Google has not said when Gemini 4 Argon will open beyond the Fairwind Program; its introductory price lasts for an unspecified period. DeepSeek has not dated the V4.1-Pro launch that will end its current routing arrangement. Alphabet has said its capital spending will "increase significantly in 2027," with details to come, and Meta has yet to announce the date of its third-quarter results, when it is expected to update its spending range.

Top image: Nvidia H100 data-center accelerators. Photo: 极客湾Geekerwan / Wikimedia Commons, CC BY 3.0.