2 News This Week: China AI Tokens Usage Through the Roof ; DeepSeek New Coding Model to be Released Soon. Implications?

Two AI stories this week look minor at first glance, but together they hint at much bigger changes in who wins and loses in the AI economy.

1. Chinese‑built AI models now account for about 61% of all usage (measured in “tokens”) on OpenRouter, a major platform where developers can choose among many AI models through one API.

2. DeepSeek, a fast‑growing Chinese AI startup, is expected to launch DeepSeek V4, a new model focused on coding — aiming directly at one of the most valuable parts of the AI market.

On the surface this is just “usage statistics” and “yet another model launch.” Underneath, it is about two things:

• how AI companies charge for what they do, and

• who will control the tools that write the world’s software.

Tokens: The AI “meter” — and why falling prices matter

When you or a company use an AI model, you pay per “token.” A token is a chunk of text — part of a word — that the AI reads or writes. Every time you ask a question, upload a document, or generate code, the system counts how many tokens were processed and bills you accordingly.

Think of it as the electricity meter of AI. More tokens = more work done = more money charged.

The price of tokens has collapsed

Over the last two years, token prices have dropped dramatically, especially once low‑cost Chinese models entered the global market.

Some examples (prices per 1 million tokens, USD, indicative ranges):

In 2023–early 2024, paying USD 20–30 per 1 million tokens for a top model was normal. By 2025–26, you can buy capable models — especially from Chinese providers — for well under USD 1 per 1 million tokens in many cases.

For a start‑up or enterprise processing, say, 1 billion tokens per day (not unusual for a serious application), this difference is huge:

• At USD 20 per 1M tokens: about USD 20,000 per day in model costs.

• At USD 0.20 per 1M tokens: about USD 200 per day for the same volume.

That is a 100x cost difference for doing the same amount of AI “work.”

Open source and low‑cost models shift the balance toward users

At the same time that prices have fallen, more high‑quality open‑source and low‑cost models have become available, many of them originating in China.

• DeepSeek, GLM (Zhipu), MiniMax, and Kimi / Moonshot all offer competitive models at a fraction of the cost of older US flagships.

• Some versions can be run locally or in private clouds, giving companies more bargaining power with the big US platforms like OpenAI, Anthropic, Google and Microsoft Azure AI.

Put simply: for end users — whether consumers or enterprises — this is good news. The trend is toward:

• More competition between US and Chinese models.

• Lower cost per token across the board.

• More choice, including open‑source options that can be self‑hosted.

The 61% token‑usage share for Chinese models on OpenRouter is a sign that usage is already starting to follow price and value, not just brand.

Price isn’t everything: quality and the coding‑AI gold rush

Cheap tokens only matter if the model can actually do what you want. One of the clearest tests of quality today is coding — using AI to help write and maintain software.

Why coding is the “hot” use case

Coding is one of the most economically valuable uses of AI because:

• It directly boosts developer productivity — companies can ship features faster and maintain legacy systems with less effort.

• It fits naturally into existing tools (GitHub, IDEs, CI pipelines), so it can be sold as a recurring subscription per developer seat.

• It touches the core of how all future digital products will be built.

Some concrete markers:

• GitHub Copilot (Microsoft) has around 1.3 million paying subscribers and roughly 20 million users overall; external estimates put its annual recurring revenue at about USD 400 million in early 2025.

• Analysts size the AI coding assistant market at roughly USD 4.7–6.8 billion in 2025, with potential to reach USD ~14.6 billion or more by the early 2030s.

• For big model providers like Anthropic, projections show total revenue leaping into the double‑digit billions by 2026, with a significant share tied to developer and coding‑related workloads.

Because coding is demanding — it tests logic, long‑term consistency and integration with tools — companies like Anthropic have been able to justify higher prices for their best models, arguing that the extra quality pays for itself in developer time saved.

The “vibe coding” startup wave and valuations

On top of the big platforms, a whole ecosystem of “AI coding” or “vibe coding” startups has emerged, and investors are paying up.

A few examples, with links:

• Cognition – maker of the “Devin” coding agent

  • Site: https://cognition.ai

  • Blog on funding and vision: https://cognition.ai/blog/funding-growth-and-the-next-frontier-of-ai-coding-agents

  • Reported raise: nearly USD 500 million at a roughly USD 9.8–10.2 billion valuation in 2025, with ARR ramping from low single‑digit millions to tens of millions within a year.

• Devin product information and commentary

  • Overview: often linked via Cognition’s main site above.

  • Coverage of the “AI coding agent” model and funding: for example, Remio’s write‑up: https://www.remio.ai/post/cognition-ai-backs-devin-ai-coding-agent-with-400m-to-transform-developer-tools

• Sourcegraph – “Cody” AI coding assistant

  • Company site: https://sourcegraph.com

  • Cody product: https://sourcegraph.com/cody

  • Funding news:

    • Series D announcement summary: https://salestools.io/en/report/sourcegraph-150m-series-d-2024

  • Raised USD 150 million at a USD 2.625 billion valuation, with total funding around USD 225 million.

• Broader “AI coding / agent” startup lists and valuation context

  • “85 Hottest AI Startups to Watch in 2026” (includes several coding‑adjacent tools): https://wellows.com/blog/ai-startups

  • “Top AI Agent Startups 2026 (Funding & Valuation)”: https://aifundingtracker.com/top-ai-agent-startups/

  • “AI startups revolutionize coding industry, leading to sky‑high valuations” (Reuters‑style coverage): you can link out to a representative news article if you like this angle.

The pattern is clear: investors are comfortable valuing pure‑play coding‑AI companies in the USD 2–10 billion range on revenues that are often still in the tens of millions. That tells you how central they think coding will be to AI monetisation.

Where DeepSeek V4 fits in: cheaper, and possibly just as good

This brings us back to DeepSeek and its upcoming V4 model.

DeepSeek is already known for offering very low‑cost, high‑performance models, with earlier versions priced at a fraction of legacy US models. Chinese competitors like Moonshot AI / Kimi, MiniMax and Zhipu AI have followed similar strategies: push quality up, push token prices down.

DeepSeek V4 is expected to:

• Focus specifically on coding and long, complex codebases.

• Aim to be comparable to, or better than, best‑in‑class coding models from US providers, based on early benchmarks and commentary.

• Be priced aggressively, extending DeepSeek’s playbook of undercutting incumbents and forcing price adjustments across the market.

If DeepSeek V4 can deliver coding quality in the same ballpark as premium offerings from OpenAI, Anthropic or Google, at a fraction of the cost per token, two things follow:

1. End users get a better deal.Enterprises and developers will see lower AI bills and more options — including Chinese models — for critical workloads like coding, documentation and code review.

2. Pricing power at the top gets challenged.Companies that have been able to charge a premium for coding‑grade quality may find it harder to defend those prices if a credible, cheaper alternative exists at scale.

Why these “minor” news items are worth watching

For a general reader, it is easy to gloss over headlines about token usage and model launches. But the combination of:

• 61% usage share for Chinese models on OpenRouter,

• Rapidly falling token prices from both US and Chinese providers, and

• A serious attempt like DeepSeek V4 to enter the high‑value coding market,

points toward a world where:

• AI becomes cheaper and more competitive.

• The benefit shifts gradually toward end users through lower prices and more capable tools.

• The map of AI power may be more evenly split between US and Chinese providers than today’s market valuations suggest.

AI is getting more exciting every day it seems. Stay tuned.