The fire of open source is spreading to Silicon Valley

Original Author: Su Yang

Original Editor: Xu Qingyang

Original Source: Tencent Technology

A subtle shift is taking place in Silicon Valley.

Over the past two years, when enterprises talked about AI, the question they cared about most was often "whose model is the strongest." Now, as AI applications scale up, another question is becoming increasingly practical: is it really necessary to use the most expensive model to complete the same task?

As IT spending continues to rise, U.S. enterprises are beginning to re-examine AI costs and are increasingly turning to cheaper "open-source" AI models. More precisely, a considerable portion of these are "open-weight" models.

According to data from research platform AlphaSense, in August and September of this year, mentions of "open-weight" or "open-source" models during U.S. corporate earnings calls and investor meetings were six times higher than the same period last year.

From tech companies to finance, logistics, industrial, and telecommunications sectors, enterprises are beginning to experiment with combining different models based on task difficulty: the most expensive frontier models handle complex tasks, while cheaper open-source models take on a large volume of routine work.

Frontier Models Can't "Sweep Everything"

Two Chinese AI startups, Zhipu AI and Moonshot AI, released new models whose performance rivals that of American lab products

Over the past few years, OpenAI and Anthropic have continuously pushed the boundaries of frontier model capabilities, establishing a clear business logic: enterprises are willing to pay higher prices for stronger models, because greater intelligence means better code, more complex reasoning, and stronger automation capabilities.

But as AI moves from small-scale experimentation into large-scale production environments, enterprises are recalculating the math.

A large number of routine tasks—information extraction, text classification, simple code modifications, customer service replies—do not require calling the most expensive frontier models.

Reuters reported in June of this year that executives including Microsoft CEO Satya Nadella, Palo Alto Networks CEO Nikesh Arora, and Coinbase CEO Brian Armstrong have all begun to emphasize that smaller, cheaper models can meet a considerable portion of enterprise needs.

Uber is a typical case.

Due to heavy employee use of AI coding tools, the company burned through its entire AI budget originally covering all of 2026 in just four months, ultimately having to set caps on AI tool usage.

More complex tasks further drive up costs. As a single request involves more steps, longer context, and more tool calls, even if the per-token price declines, the enterprise's final bill can still grow rapidly.

Thus, "which model is the strongest" has gradually become "which model is best suited for this task." Open-source models have gained an opportunity as a result.

Unlike companies such as OpenAI and Anthropic, which primarily provide model services through the cloud, open-source models allow enterprises to download model parameters, run them on their own servers or cloud infrastructure, and fine-tune them according to business needs—thereby reducing some inference costs while gaining greater deployment and data control.

The Financial Times, citing Vercel data, reported that in August of this year, open-source models accounted for 56% of all tokens processed by the company's AI Gateway, up from just 7% in December last year. Reuters, citing Citibank data, reported that the share of tokens processed by open-source models on OpenRouter also rose from 34% in January to 65% in June of this year.

The "Model Mix"

What is truly reshaping the AI business model is not some open-source model defeating closed-source giants, but enterprises beginning to abandon the idea of "one model to solve all problems." This shift is particularly evident in the software industry.

Match Group's Tinder, in order to control surging AI costs, has begun routing some non-technical queries to open-source models. Its CTO Vinay Kuruvila revealed that the company's annualized AI spending surged from $1 million in January to $10 million in July. Although top-tier models can handle about 90% of tasks, as open-source models continue to catch up, a large volume of work no longer requires calling the most expensive systems.

Telecom giant AT&T has gone even further. Its Chief Data and AI Officer Andy Markus revealed that approximately 40% of AI workloads currently run on open-source models, with plans to increase that to 70% within a year. Facing a daily processing volume of 45 billion tokens, AT&T also uses proprietary data to fine-tune open-source models, enabling them to match or even exceed closed-source model performance on specific tasks.

As a result, the "model mix" is becoming a common architecture for AI applications: frontier models handle complex reasoning and planning, while low-cost models take on a large volume of execution work.

This shift is not limited to consumer applications—tools and developer software are undergoing the same transition.

Testing by the Cursor team also demonstrates this cost difference. Building a browser from scratch using OpenAI GPT-5.5 throughout would cost over $10,000; switching to a self-developed Composer model paired with Anthropic Opus 4.8 reduces the cost to $1,339—a difference of about 7.5x.

This model also makes "Model Agnostic" an important capability for AI applications. Enterprises can flexibly route between OpenAI, Anthropic, Google, and open-source models based on task difficulty and cost, rather than being locked into a single vendor.

Telnyx previously used Anthropic's top-tier model to run 1,000 bot agents, which under the original billing method could cost up to $100,000 per day on average. After switching to open-source models, its bot agents increased to 1,400, yet the average daily cost per agent dropped to about $100.

Legal AI startup Harvey has adopted a similar strategy: routine tasks are handled by open-source models, and only extremely difficult work calls for top-tier models.

Redistributing Value

This shift poses a profound challenge to the commercial foundations of OpenAI and Anthropic.

In the past, frontier model companies drove enterprises to increase AI spending by continuously improving model capabilities. The stronger the model, the higher the premium, and that premium was converted into compute investment for the next generation of models. But open-source models are breaking this closed loop.

Citibank data shows that some open-source models were already priced as low as 18 cents per million tokens, while top-tier closed-source models average around $4. Even if open-source models have performance gaps on some extreme tasks, such a stark price difference is enough to make enterprises reassess whether it is necessary to call top-tier models for all tasks.

More importantly, the open-source camp is catching up on hardcore capabilities far faster than expected.

DeepSeek released its V4 Pro model to general availability (GA) without any pre-heating, and the model scored a top-tier 80.6% on SWE-bench Verified, setting the highest Codeforces rating among tested models. Its overall token rate is only about 1/46 that of Claude Opus 4.7.

When an open-source model costing just 1/46 of a closed-source top performer can not only do the job but even directly match frontier top-tier models on key metrics like deep coding and algorithmic reasoning, the balance beam of buyer procurement strategy is completely broken.

WEKA Chief AI Officer Val Bercovici once summarized this trend as "achieving 90% of the results at 10% of the price," and the emergence of products like DeepSeek V4 Pro is even breaking through the "90% of results" limitation, moving toward parity with top-tier performance.

Beyond cost, data control is also a significant draw of open-source models.

Scott Wallace, Senior Global Director of Solutions Architecture at Digital Realty, said the company has established an internal chat interface running on open-source models, and combines open-source and proprietary models based on the sensitivity and complexity of the task.

For tasks involving sensitive medical, financial records, or trade secrets, open-source models allow enterprises to deploy code and parameters within their own infrastructure for fine-tuning and auditing—this is not just an economic calculation, but also a compliance and governance one.

Software Companies Are Back in the Game

Over the past two years, there has been a persistent concern in the AI industry: foundation model companies might directly eat into the traditional software market. If users can directly ask ChatGPT or Claude, why would they still need to buy traditional software? But as model costs drop rapidly, this logic is reversing.

For software companies, AI is essentially an input cost. The cheaper the model, the more software companies can embed AI capabilities into their products at extremely low cost, while boosting their own profit margins.

William Blair analysts recently pointed out that cheaper, more open AI models constitute a "clear positive" for software companies. Software companies can both expand the reach of AI features and improve profit margins. Fierce competition among OpenAI, Anthropic, Google, and open-source models has given software buyers unprecedented bargaining power and choice.

Echoing this trend is the strong rally in enterprise software stocks.

Over the past month, HubSpot rose nearly 30%, Adobe gained over 20%, and Intuit, Salesforce, ServiceNow, and Asana all rose at least 12%.

Former Benchmark partner Bill Gurley believes that if powerful AI models can remain open and affordable, startups, cloud service providers, chip companies, enterprises, and researchers will all benefit.

For a large number of AI startups, they no longer need to invest billions of dollars in competing over foundation model pre-training. Instead, they can directly integrate existing high-performance models into their products, using customer relationships, industry knowledge, proprietary data, and workflows to build competitive moats. The cheaper the models, the less cost pressure this model faces.

The Era of "Good Enough"?

NVIDIA CEO Jensen Huang said in a post on X that he supports the coexistence of closed-source and open-source models.

This shift will ultimately force giants to adjust their long-term strategies.

Previously, OpenAI launched GPT-5.6 Sol to improve token efficiency, and Anthropic introduced the cost-effective Claude Opus 5; subsequently, both companies accelerated the release of their latest frontier and lightweight models, attempting to lock in enterprise customers with different compute budgets within their own ecosystems by offering a richer "high-low" product lineup.

The metrics for model competition are also being reshaped.

Beyond benchmark scores and pure reasoning capability, cost per token, token efficiency, cost per task, and the ability to flexibly route within multi-model architectures are becoming more important than ever.

The Financial Times reported that Anthropic is preparing for a potential IPO, with market expectations that its valuation could reach $2 trillion or even higher; OpenAI, after postponing its IPO plans to 2027, is also in early negotiations for a private funding round at a valuation of approximately $1.2 trillion. The flow of enterprise AI spending is directly tied to whether both companies' massive compute investments can continue to translate into commercial returns.

Meanwhile, the open-source model camp continues to expand at an accelerating pace.

Teams including Mistral, NVIDIA, Reflection AI, Thinking Machines Lab, and DeepSeek are all continuously pushing the upper limits of open-source models, and companies such as Microsoft, NVIDIA, Meta, Palantir, and IBM have publicly supported an open AI ecosystem. NVIDIA CEO Jensen Huang once said: "The world needs both frontier closed-source models and frontier open-source models."

Against this backdrop, the blind inertia of indiscriminately paying a "closed-model tax" on every prompt is being thoroughly broken. From startups' flexible architectural pivots to telecom giants' meticulous accounting of dynamically routing tens of billions of tokens daily, decision-making power is shifting from pure "brand trust" to "architecture design and contract terms."

Or rather, the logic of future choices lies in: for this task, who can deliver good enough results at a lower cost?