NextFin News - Open-source AI is taking more token volume, but Anthropic is still collecting the premium spend because the market is splitting into two layers instead of collapsing into one. That is the clearest reading of the latest platform data: Vercel’s AI Gateway shows DeepSeek at 33.5% of token volume across top labs, while Anthropic still leads spend at 54.4%; OpenRouter’s weekly rankings show DeepSeek V4 Flash at 4.88 trillion tokens and Claude Opus 4.8 at 2.02 trillion. The conclusion is narrow but important. Open-source models are winning more routine usage, yet frontier models are still the ones customers pay up for.
The latest debate was sharpened by Decagon chief executive Jesse Zhang, who argued in a post titled “Everyone is wrong about open source AI in the enterprise” that mature deployments are moving to lighter models even as overall frontier spend barely moves. That framing matters because it suggests open source and frontier systems are not direct substitutes across the whole stack. Instead, they occupy different phases of the same workflow: expensive models are used to prove out ideas, and cheaper models absorb more mature production tasks later.
Anthropic’s own releases point in the same direction. On June 30, the company said Claude Sonnet 5 would be available on Claude.ai, Claude Platform, Amazon Web Services, Google Cloud and Microsoft Foundry. It described the model as suited for coding, tool use and everyday work, and said it would be priced at $2 per million input tokens and $10 per million output tokens through August 31, 2026 before rising to standard pricing. That is a premium product aimed at serious production workloads, not a commodity response to the cheapest models in the market.
The company’s June 26 Economic Index report also shows that Claude is being used more like an agentic work tool than a simple chat assistant. Anthropic said that a year ago most Claude usage took the form of a conversation between a user and an assistant, but that sessions now increasingly consist of long-running agentic tasks because of the rapid growth of Claude Code and Cowork. It also said personal conversations rise from around 35% of weekday usage to just under 50% on weekends, underscoring that Claude remains embedded across both work and non-work use cases.
That combination — more open-source volume, but stubborn frontier pricing power — is why Anthropic has not yet felt the full effect of the open-source wave. The model mix is changing, but the economics are not moving in a straight line. Customers can shift low-value and repetitive tasks to cheaper systems without immediately abandoning the frontier models they still need for discovery, hard reasoning and the first pass on complex workflows.
Open Source Is Winning Volume, Not Yet the Premium Layer
The simplest explanation for Anthropic’s resilience is that AI demand is expanding faster than substitution is compressing it. As mature workflows move toward lighter or open-source systems, new workflows keep opening up, and the frontier model remains the place where companies test the hard problems first. That is the core of Zhang’s argument, and the data this week gives it some support.
Vercel’s AI Gateway leaderboard shows a clear split between usage and monetization. DeepSeek holds the largest share of token volume at 33.5%, while Anthropic sits second at 20.8%. But when the same platform looks at spend, Anthropic jumps to 54.4% of total AI Gateway spend. In other words, a smaller share of usage can still generate a larger share of revenue if the workload mix is more complex, more expensive or more valuable. That is exactly the position a frontier lab wants to defend.
OpenRouter’s weekly rankings tell a similar story from a broader slice of the market. DeepSeek V4 Flash leads weekly usage with 4.88 trillion tokens, while Claude Opus 4.8 remains one of the largest frontier models in use at 2.02 trillion tokens. The ranking does not reveal revenue by model, but it does show that Claude still sits near the top of the market for real traffic, even as cheaper alternatives flood the lower end of the market.
The important part is not that open source is failing. It is that open source is succeeding first where the economics are easiest to displace. Routine production work, high-volume requests and cost-sensitive use cases are more likely to migrate to cheaper models. Frontier systems retain an edge where customers care more about capability, reliability and workflow depth than about the lowest possible token price. That leaves Anthropic exposed to competitive pressure, but not to a clean collapse in revenue power.
Anthropic Is Still Selling Capability, Not Just Access
Anthropic’s product strategy shows that it is still competing on what its models can do, not just on how much they cost. In its Sonnet 5 launch materials, the company described the model as a hybrid reasoning system for real-time agents and high-volume work, with a 1 million token context window. The pricing structure also makes the point: the introductory $2/$10 per million input and output tokens remains firmly in the premium tier, even before it steps up to $3/$15 after August 31.
That matters because price only becomes a problem when customers can get the same result for materially less money. Anthropic is still trying to keep that gap open through capability, deployment breadth and product fit. Sonnet 5 is available not only on Anthropic’s own platform, but also on major cloud channels, which gives enterprise customers more ways to adopt it without changing their procurement stack. The company is selling access to high-end capability where integration friction, not raw model cost, is often the bigger purchasing barrier.
Anthropic’s redeployment of Claude Fable 5 and Claude Mythos 5 reinforces the same point. The company said on June 30 that export controls on both models had been lifted and that Fable 5 would be available globally starting July 1. It also said a newly trained safety classifier blocks the reported bypass technique in over 99% of cases. The message is that frontier models are still being marketed as differentiated products that need both capability and guardrails, not as interchangeable utility inputs.
“The frontier labs will keep owning discovery. Open source will increasingly own production.”
That line from Jesse Zhang is the cleanest summary of the market structure that seems to be emerging. It explains why Anthropic is still holding up: frontier labs can keep monetizing discovery, first deployment and the hardest enterprise workflows even if open-source systems increasingly own the cheaper, more routine end of production. The two layers can coexist for a while, and the data this week suggests they still are.
The Real Test Is Whether Frontier Premiums Hold
The open question is not whether open-source AI matters. It already does. The question is whether it can erode the frontier premium fast enough to hurt labs like Anthropic before those labs keep inventing new reasons to spend. For now, the evidence says the pressure is real but incomplete. Anthropic is still leading spend on Vercel, still prominent on OpenRouter, and still shipping models that are explicitly designed for complex coding, tool use and professional work.
Anthropic’s June 26 Economic Index report adds another reason the company is not yet feeling a full substitution effect. The report says Claude usage has shifted toward long-running agentic tasks, with work patterns still tied to the week and personal use rising on weekends. That points to a product that is becoming more embedded in daily workflows, not less. As tasks become more complex, the cheapest model is not always the one that wins the deal.
That does not make Anthropic immune. It does mean the pressure point is slower and more structural than a simple “open source is here, therefore incumbents lose” narrative would suggest. The risk for Anthropic is a gradual squeeze on the parts of the market where customers can downgrade without pain, while the company must keep proving that its frontier models still justify a premium for the toughest tasks.
For now, the numbers say that squeeze has not arrived in full. Open source is taking more volume. Anthropic is still taking more spend. The gap between those two facts is the story — and it is why the rise of open source AI is not hurting Anthropic yet.
Explore more exclusive insights at nextfin.ai.
