Open Source AI Is Booming, Yet Anthropic Is Not Losing Ground
Cheap open models now handle a third of the tokens on Vercel, but frontier labs still capture the spend. The split reveals where the real margins hide.
The News
The surge in open source artificial intelligence is not, so far, coming at the expense of the frontier laboratories that helped create the category. Fresh usage data reported by TechCrunch on 7 July 2026 shows that low-cost open weights models and premium proprietary systems are each capturing a different stage of the same life cycle, rather than fighting over the same customers.
The numbers are striking. On Vercel's platform, DeepSeek now processes just over a third of all tokens, yet Anthropic still accounts for more than half of the total spending on AI. On OpenRouter, DeepSeek's V4 Flash pushes 5.3 trillion tokens each week, while Anthropic's Opus 4.8 handles just over 2 trillion. Volume and revenue, in other words, have decoupled.
The reason sits in the price gap. Opus 4.8 costs $1.37 per million tokens, roughly 23 times the 6 cents per million that DeepSeek charges. Cheaper challengers such as Z.ai's GLM-5.2 and Nvidia's Nemotron are pushing in the same direction, expanding raw throughput without collapsing demand at the top of the market.
Why It Matters
The framing that matters here is a division of labour. As Jesse Zhang, chief executive of Decagon, put it: "The frontier labs will keep owning discovery. Open source will increasingly own production." Companies reach for the most capable and expensive model when they are proving out a new use case, then migrate the mature, high-volume workload to a cheaper open model once the problem is understood.
That pattern echoes the early cloud computing era, when businesses paid a premium for managed services to figure out what worked before optimising costs at scale. The fear that open weights would hollow out the frontier labs assumed a single market. In practice there appear to be two, and premium spending keeps rising because each solved problem seeds the next expensive experiment.
The risk for the frontier labs is that the discovery phase keeps shrinking as open models improve. If capability narrows faster than price, workloads graduate to production sooner, and the premium tier needs fresh frontier tasks to justify its economics.
Indian Angle
For India's developer economy, the 23-fold price gap is not an abstraction, it is a survival calculation. Startups building in rupees have always been unusually cost-sensitive, and a model that runs at 6 cents per million tokens changes what is viable for consumer apps serving hundreds of millions of low-ARPU users. Expect Indian SaaS and agentic-workflow firms to adopt exactly the split Zhang describes, prototyping on frontier models and shipping production on open weights.
It also validates the strategic bet made by India's own model builders. Sarvam and Ola's Krutrim have leaned towards open or openly licensed approaches, and MeitY's IndiaAI Mission has channelled public money towards indigenous foundation models on the assumption that open ecosystems lower the cost of national adoption. The Vercel data suggests that owning the cheap, high-volume production layer is a defensible position, not a consolation prize.
The caution for Indian founders is dependence. If discovery value concentrates in a few foreign frontier labs, Indian firms risk building on a research pipeline they do not control. That is the case for keeping some frontier ambition at home, even as cheap tokens do the daily work.
FAQ
Does open source AI threaten Anthropic's revenue?
Not yet. Anthropic still commands more than half of AI spending on Vercel despite handling a minority of tokens, because its Opus 4.8 model is priced around 23 times higher than DeepSeek and is used for higher-value discovery work rather than bulk production.
How much cheaper are open models?
DeepSeek charges about 6 cents per million tokens against $1.37 per million for Opus 4.8. That gap is why open weights models now carry far more raw volume, such as DeepSeek's 5.3 trillion tokens a week on OpenRouter, without capturing the majority of revenue.
What does this mean for Indian startups?
Indian teams can prototype on frontier models, then move mature, high-volume features to open models to protect unit economics. For rupee-denominated products serving large low-ARPU user bases, the price gap can decide whether a feature is commercially viable at all.
Where can I read the original report?
The underlying usage figures and analysis were published by TechCrunch on 7 July 2026. The article draws on data from Vercel and OpenRouter and a comment from Decagon chief executive Jesse Zhang.
This story was reported by TechCrunch. Read the full original coverage at TechCrunch.