Will One AI Company Take It All? Premium frontier models, open-weight challengers four months behind, the chip war, and the sovereignty wall: the 2026 AI market in pictures

When OpenAI opened GPT-4 to developers in March 2023, a million input tokens cost $30.^[1] GPT-6 Luna, which OpenAI released in September 2026, charges $0.10 for the same amount.^[2] The two models are not equally capable, but the gap shows how much the price of using AI has moved in three and a half years.

In the early days of the generative AI boom, many people expected one or two labs with the biggest computers and the most data to take nearly the whole market. That is not quite how it played out. The most expensive frontier models still sell well, but open-weight models follow a few months behind, cloud companies are designing their own chips instead of buying only from NVIDIA, and European public bodies are wary of handing their data to US-owned servers.

One caution before we start: the numbers here come from different sources, periods, and scopes, so two similar-looking figures are not necessarily comparable. We keep company claims apart from independent measurements, and results apart from plans. For model-by-model performance and pricing, see our 2026 AI model map and our open-weight models guide.

01. The AI market became a food court

Think of the AI market as a place to eat. A few years ago it looked like one restaurant with a star chef would win every customer in town. In 2026 it looks more like a food court. You order the chef’s tasting menu for an important occasion, grab lunch from a cheap counter on an ordinary day, and if you are not allowed to take your ingredients outside the building, you use a private kitchen. We will stay in this food court for the rest of the article.

The crowd is growing fast. Venture firm Menlo Ventures estimates that enterprises spent $1.7 billion on generative AI in 2023, $11.5 billion in 2024, and $37 billion in 2025.^[3] The figure covers model APIs, model training infrastructure, and AI applications. It does not include NVIDIA chip purchases.

02. Where the leader makes money: expensive mistakes

The striking part of the same Menlo survey is how the ranking flipped. OpenAI’s share of enterprise LLM API spend fell from 50% in 2023 to 27% in 2025, while Anthropic’s rose from 12% to 40%. Google climbed from 7% to 21%.^[3] These are spend estimates based on API usage in production, which is a different measure from how many people use a chat app.

Anthropic’s own numbers point the same way. When it announced a funding round in September 2025, the company said its run-rate revenue had grown from about $1 billion at the start of that year to more than $5 billion in August.^[4] According to the IPO prospectus Reuters reviewed in late September, actual 2025 revenue was nearly $4.6 billion, up about 12-fold in a year, with an operating loss of more than $8 billion.^[5] One reading is that for work such as coding tools and business automation, where a wrong answer is expensive, more companies are willing to pay for the model that makes the fewest mistakes. In food court terms, that is the tasting menu you order when an important guest is at the table.

Not every meal has to be a tasting menu, though. You can put a router in front of your models that reads each request and sends it to the right one. Like a food court’s front desk, it sends high-stakes work such as contract review to a model proven on that task, with a person checking the result. Bulk work such as receipt tagging goes to a cheap model, and data that cannot leave the building stays on an in-house model. A new kind of model is now competing for the front-desk job: decision models, which return options and probabilities instead of prose. We covered TypeSafe’s Jev, released September 15, and the OpenAI and AWS models that followed in our decision-model race article.

03. How the leader stays ahead: think longer, try more

So how does the leader stay ahead? Bigger models trained on more text are not the only route. Test-time compute means spending more computation at the moment of answering: letting the model reason longer, or generating several candidates and checking which one holds up. Whether that pays off depends on the model, the task, and whether there is a way to check the answer.

In April 2026, researchers at Wisconsin–Madison and Stanford published a Train-to-Test, or T², scaling study that jointly considers training and inference computation.^[6] It sets separate budgets for training and for answering, then tunes model size, training data, and the number of attempts together. The experiments centered on models under one billion parameters and eight tasks, so it is too early to read them as proof of how frontier labs stay ahead.

The paper counts attempts as k = C_inf ÷ (2N). C_inf is the inference budget in FLOPs per token served, N is the model size, and one forward pass is approximated as 2N FLOPs. Under those assumptions, halving the model buys twice as many attempts on the same budget. That does not mean full responses cost the same regardless of their length or serving setup. The slider shows only this proportional relationship.

There is a catch. Studies like this often report pass@k, which counts a success if any one of k attempts is correct. A user, though, needs the one right answer picked out of, say, twenty drafts. To benefit from repeated attempts in a real product, you need a grader or verifier that can pick it. Even in math and coding, where grading is easier, the benefit is not guaranteed on every workload.

04. Olympiad gold, but it can’t read a clock

Stanford HAI’s 2026 AI Index report contains a telling contrast. Google’s Gemini Deep Think scored 35 points at the 2025 International Mathematical Olympiad, a gold-medal result. Yet on ClockBench, a test of reading the time from analog clock images, the best model (GPT-5.4 High) got only 50.6% right. Humans got 90.1%.^[7] The report calls this “jagged intelligence.” Try it yourself.

AI has reached Olympiad-gold level in math, yet even the best model trails people badly at reading a clock. A high benchmark score does not guarantee good results on your own work. It also helps explain why expensive frontier models keep selling. The work companies want to hand off, such as tax paperwork, loan files, and financial analysis, looks nothing like public test questions. If one model makes fewer mistakes on a company’s own data even when public scores are similar, paying more for it can make sense. Our field notes explain how to build a small test set from your own tasks.

05. Open-weight models, four months behind

Now let’s visit another counter. An open-weight model is one whose trained weights anyone can download and run on their own servers. The weights are the numbers that hold what a model learned in training. In food court terms, you get the finished recipe. Companies rarely share where the ingredients came from or how they practiced, meaning the training data and code, so these models are called open-weight rather than open source.

Epoch AI’s May 29, 2026 analysis found that since January 2026, the best open-weight models have reached the closed frontier’s level about four months later on average. The yardstick is ECI, Epoch’s composite capability index built from many benchmarks, and the average gap of 8 points is similar to the gap between GPT-5 and GPT-5.5.^[8] That does not mean a four-month lag on every workload. Epoch also noted the gap is slightly larger than in its earlier analysis, which put it at about three months over 2023–2025.^[9] Open models are close behind, but there is no sign yet that the distance is shrinking.

Individual models make the chase feel real. China’s DeepSeek released the weights of DeepSeek-V4-Pro under the MIT license. In the comparison table DeepSeek published on the model card, V4-Pro’s maximum reasoning mode scored 80.6% on SWE-bench Verified, a test of fixing real code issues, close to Anthropic’s Claude Opus 4.6 at 80.8% in the same table. On LiveCodeBench it scored 93.5%, the highest in the table.^[10] DeepSeek chose the tests and settings itself, so treat these as vendor-reported results rather than independent ones. In Europe, Mistral released a public API preview of Mistral Large 4, a model of about one trillion parameters, on October 6, and said the weights will follow by the end of October.^[20]

When DeepSeek launched its R1 reasoning model in January 2025, the price was the headline. R1 cost $0.55 per million input tokens and $2.19 per million output tokens,^[11] about 4% of what OpenAI charged for o1 ($15 and $60), the model it was most often compared with.^[12] Some argue this rattled investors. In a working paper released in August 2026, David Krause, an emeritus professor at Marquette University, calculated that the ratio of the tech-heavy volatility index to the broad market’s (VXN/VIX) peaked at 1.73 in July 2026, higher than during the 2008 financial crisis.^[13] It is one researcher’s calculation and has not yet been peer reviewed.

06. 8 of 128 switched on: mixture of experts

Many open models use mixture of experts, or MoE. An MoE model splits part of its network into expert sub-networks, and a learned router inside the model picks only a few of them for each token. Picture a kitchen with 128 cooks where only 8 work on any given order. That cuts computation sharply for the model’s size, but architecture alone does not explain how fast open models are catching up or what they cost to serve.

Alibaba’s Qwen3-235B-A22B works exactly this way. Of its 128 experts, only 8 switch on for each token, so it uses 22 billion of its 235 billion parameters, about 9.4%, per step.^[19] DeepSeek-V4-Pro uses 49 billion of 1.6 trillion, about 3.1%.^[10]

There is a catch: even if only 8 cooks work at a time, all 128 still need room in the kitchen. Sparse activation does not remove the need to store the full weights, though they need not all sit in GPU memory at once. Quantization (storing numbers at lower precision) and offloading part of the model to the CPU can ease GPU memory needs, at the cost of data movement and speed.^[58] Self-hosting also depends on weight precision, context length, batching, and the server setup.

07. 280 times cheaper, yet bigger bills

Efficiency tricks like MoE, plus competition, have pushed down the price of a given level of AI quickly. Stanford’s AI Index 2025 found that the price of a model performing at GPT-3.5’s level fell from $20 per million tokens in November 2022 to $0.07 in October 2024, a drop of more than 280 times.^[21] Epoch AI found that, depending on the task, prices fell by anywhere from 9 to 900 times per year.^[22]

Price competition was fierce too. When Anthropic launched Claude Opus 4.5 in November 2025, it cut Opus pricing from $15 input and $75 output to $5 and $25, a two-thirds reduction.^[23]

Yet bills often go up. Economists call this the Jevons paradox. When steam engines became more efficient in 19th-century Britain, coal use rose instead of falling, and when AI gets cheaper, people hand it jobs that used to be too expensive to automate. A Fortune article in June 2026 summed it up: “The models get cheaper. The usage gets heavier. The bill stays stubbornly high.”^[24] Try the calculator.

The newest models show the same pattern. With Claude Fable 5.1 in September, Anthropic cut the price of cache reads, meaning reuse of a long input it has already stored, to $0.25 per million tokens, 75% below the previous model. The base price stays at $10 input and $50 output.^[25] Yet the company’s estimate and an independent measurement point in opposite directions.

Which one is right? The two did not measure the same workload with the same settings, so they cannot be compared head to head. How often the cache is reused, how long the answers run, how hard the model reasons, and what tasks it gets all change the cost per task. The independent cost rose because the model wrote more output tokens, which on its own does not disprove Anthropic’s discount estimate. As of September 2026, all four models below charged five times as much for output as for input, with cached input priced separately.^{[25][2][27][28]} For task costs including tool calls and retries, see our Claude Opus 5.5 vs GPT-6 Sol cost comparison.

08. The chip war: building instead of buying

A food court also runs on its kitchen equipment. Most of the chips that train and run AI today are NVIDIA GPUs. In the second quarter of its 2027 fiscal year, reported in August 2026, NVIDIA’s gross margin was 75.0%.^[29] For every $100 of sales, about $75 is left after the cost of making the product. That gives cloud companies buying chips by the truckload every reason to make their own.

The jobs are splitting, too. Training new frontier models still leans heavily on NVIDIA GPUs and the CUDA software ecosystem. Inference, meaning running finished models to answer requests, happens nonstop every day, so chips tuned to run specific models cheaply have an edge there. Deloitte expects inference to account for roughly two-thirds of all AI compute in 2026.^[39]

Lower kitchen-equipment costs can create room for cheaper food, but menu prices do not fall automatically. In-house chips work the same way. They can lower the cost of particular workloads and push AI prices down, but the real effect depends on how busy the chips are kept, on memory, networking, and software, and on contract terms. Advertised speed and cost advantages are mostly vendors’ own measurements and hard to compare directly. Google’s 8th-generation TPU specs and the three clouds’ chip strategies are covered in our cloud AI war article.

09. The border wall: data sovereignty and Cohere

Some walls are legal, not technical. The 2018 US CLOUD Act can require a provider subject to US jurisdiction to disclose responsive data in its possession, custody, or control under valid legal process, even when stored abroad.^[40] Nor is coverage decided by a company’s nationality or server location alone. So with sensitive data, check who you contract with, who runs the system, and who can access the data on what terms. Sovereign AI is the name for deployment and operating models built for customers who want that control. In food court terms, it is the private kitchen where ingredients never leave the building.

Toronto-based Cohere is a leading example. Rather than compete in consumer chatbots, Cohere focused on installing models on customers’ own servers or private clouds. According to Reuters, about 85% of its business came from these private deployments as of May 2025.^[41] CNBC reported that Cohere’s annual recurring revenue was about $240 million at the end of 2025.^[42]

On September 16, 2026, Cohere and Germany’s Aleph Alpha signed a definitive agreement to combine.^[43] The two had announced the plan in April, when Germany’s Schwarz Group, the retail giant, agreed to lead an investment in Cohere and provide its STACKIT sovereign cloud. After the deal, Cohere’s shareholders will own about 90% and Aleph Alpha’s about 10%.^[44] The deal still awaits final regulatory approvals, and Cohere expects it to close later this year; no completion had been announced as of October 6. Nor does the combination mean exclusive use of STACKIT or a blanket exemption from the CLOUD Act. On September 11, Bloomberg reported that Cohere was in talks to raise money at a valuation of about $20 billion.^[45]

10. The rulebook threshold: 10²⁶ FLOPs

AI safety laws shape the market too. Some frontier and general-purpose model rules use training computation, but not every AI regulation uses the same test. FLOPs counts floating-point operations, the basic arithmetic steps behind AI computing. 10²⁶ is a one followed by 26 zeros: a computer doing a trillion operations per second would need about 3.17 million years of nonstop work to reach it.

The EU AI Act presumes high-impact capabilities above 10²⁵ training FLOPs and applies its systemic-risk classification procedure. Notification is due without delay and within two weeks of meeting, or knowing the model will meet, the condition. General-purpose AI duties applied from August 2, 2025, with a transition to August 2, 2027 for models already on the market before that date.^[46][47]

California SB 53’s model threshold is more than 10²⁶ integer or floating-point training operations. The large-developer test is more than $500 million in prior-year collective affiliate gross revenue. In force since January 1, 2026, it requires covered safety incidents to be reported within 15 days of discovery, with a 24-hour disclosure when death or serious injury is imminent.^[48][49]

New York’s March 2026 RAISE amendments likewise use more than 10²⁶ operations and more than $500 million in prior-year collective affiliate gross revenue. The law takes effect January 1, 2027. A critical incident must be reported within 72 hours of determination or sufficient supporting facts, with a separate 24-hour imminent-death or serious-injury disclosure.^[50][51][61]

So do these rules favor big companies? Some worry that compliance costs make it harder for small developers to enter; others argue that tying duties to size spares them. California and New York add duties for large developers, but that design alone does not settle the competitive effect. Actual entry and operating costs, and enforcement results, will tell. This summary is not legal advice or a compliance guarantee for any particular service.

11. New security holes in the agent era

When agents read email, access systems, and decide their next action, cost and permission controls become important. In June 2025 Gartner forecast that more than 40% of agentic AI projects would be canceled by the end of 2027 because of rising costs, unclear business value, or inadequate risk controls. That is a forecast, not a cancellation rate already observed or a failure rate for every company.^[52]

The security risks are concrete. OWASP, the software security community, kept prompt injection as the number one risk in the 2026 edition of its top 10 for large language model applications, released in August 2026.^[53][54] It published a separate top 10 for agentic applications in December 2025.^[55] The trickiest version is indirect prompt injection. Instead of instructing the agent directly, an attacker hides instructions in an email or web page the agent will read. If the agent cannot tell outside content from the developer’s instructions, it may simply follow the hidden command. Try it below.

The second risk is too much permission. Give an agent that only summarizes reports database administrator rights, and one hidden instruction can put all of your data at risk. OWASP calls this “excessive agency” and recommends granting only the minimum permissions a task needs.^[56] In the 2026 edition, it moved up from sixth to third.^[57]

Teams also use tools that preserve progress. LangGraph saves state with a checkpointer, while Temporal uses workflow history and retry policies to support recovery.^[59][60] Being able to resume does not make the model’s answers correct or its outside actions safe. An unfinished step may run twice, so payments and email sends need idempotency keys (which stop the same action from running twice) or result checks, plus separate permissions and approvals. For real incidents where agents crossed prohibited lines, see our agent control-loss files. Before rolling out an agent, run through this checklist.

12. So what should you use?

So will one company take it all? Not on today’s evidence. What we see is competition across several layers. Frontier models compete on making fewer mistakes, open weights on price and the freedom to self-host, sovereign deployments on control, and in-house chips on cost. The lines overlap, though. A closed model is not automatically safest for high-stakes work, and an open model is not automatically cheapest. Test on your own workload and check operating conditions first.

Individuals and small teams can mix models to fit the job, too. The 30-second check below is a starting point, not a performance ranking or a safety guarantee for any named model. Test your candidates on real work, and count the cost of checking and fixing their mistakes.

The top spot in AI may change hands again, just as OpenAI’s half of enterprise API spend shrank to 27% in two years. Today’s layered competition gives you more options, but it does not prove a future monopoly is impossible. So instead of chasing the leaderboard, pick on three things: how few mistakes a model makes on your work, what it costs in total, and who controls your data.

Frequently asked questions

Is “open-weight” the same as “open source”?

No. An open-weight model publishes the trained model files, but training data and training code are rarely released. Licenses also differ from model to model, so check whether commercial use is allowed.

If open models are only four months behind, should I just use them?

Four months is Epoch AI’s May 2026 composite-index average for how far the best open models trail the closed frontier. It is not the gap on every workload, and it does not promise parity four months later. Evaluate your own tasks and include self-hosting, security, and maintenance costs.

Why are AI bills rising when token prices are falling?

Lower unit prices can coexist with higher bills when call volume, output length, reasoning effort, or retries rise. The four models shown had base output prices five times input prices in the September 2026 snapshot; cache rates are separate. Compare full task costs under matched settings.

What is sovereign AI?

A way of deploying AI that lets the customer control which laws apply, who can access the data, and who runs the system. Server location or company nationality alone does not decide legal exposure or exemptions. The CLOUD Act can reach overseas data held or controlled by providers subject to US jurisdiction under valid legal process, so contracts and operating conditions matter.

Keep reading