On Wednesday (7th October), Elon Musk posted a short note outlining that Grok Bot, would use the best back-end model for any given task including Anthropic’s Claude Opus 5.5, whatever would give the user the best outcome would be used. My feeling is that with hindsight, this moment will be more significant than it now appears. It is the first time a company with its own frontier model has publicly committed to routing its users to a competitor’s. Most people I have spoken to over the past year said that would never happen, a company with its own frontier model would never route to a competitor.
Where We Are
At the end of July I wrote, a piece on The Commoditisation of Intelligence, arguing that the model layer would go the way of the browser. Netscape charged for access to the web and was ultimately usurped by those who gave it away for free. Even Google, the ultimate category winner, never really made money on Chrome, it made money above and below.
Two months on, that argument looks close to settled. On Vercel’s AI Gateway, which routes tens of trillions of production tokens a month, open-weight models ran 56% of all tokens in August. The average price per token has more than halved in five months. Routing work between cheap models and the frontier has gone from a developer habit to standard practice. This trend will only accelerate as the infrastructure needed to enable open source continues to become faster, better and cheaper.
Ben Thompson made a version of the same point in his piece Frontier Overhangs. His argument is that frontier capability is now ‘good enough’ that customers will no longer do whatever it takes to reach the cutting edge.
I think the question then becomes, not whether intelligence commoditises but rather who captures the value when it does.
For the past few weeks I have been testing an idea. Namely, that OpenAI or Anthropic should just build their own version of OpenRouter. Thus, allowing them to host open-weight models next to their own, route each task to whatever model fits it, and take a cut either way.
As this thesis was tested, it became clear that the players best placed to win are frontier companies that can also host other models, because their product was never ‘our model only’. That view was correct in terms of who would move first, and I believe my view is correct that this will prove to be the winning view.
The Router is Critical
In August, Stripe agreed to buy OpenRouter for more than $7 billion. Reacting to the deal, Bankless argued that it lets Stripe offer hybrid products the frontier labs (OpenAI/Anthropic) cannot easily replicate, because they would never dare mix their models with a competitor’s.
The purchase price is a very clear nod in the direction that market consensus believes value is flowing, specifically, to the layer that decides which model handles which task. The data supports this to a certain degree, showing a of polarisation in the market:
- ‘The Premium’- A small[er] premium tier for the hardest work: In June, Anthropic took 61% of Vercel gateway spend on 32% of tokens, and 72% or more of spend where mistakes are costly: coding agents, back-office agents and app generation. Even here the premium has limits. When Opus 5 arrived at half of Fable 5’s price, nine in ten teams running Fable cut their usage.
- ‘The Good Enough’- A large and growing market section that wants cheap and ‘good enough.’ Open-weight models went from under 4% of gateway spend in June to 14% in August, so the dollars are starting to follow the tokens. AT&T now runs 40% of its AI workloads on open models and says routing cut its coding costs by as much as 56%.
The crucial point for me, is that whoever sits between these two segments of the market, namely ‘the premium’ and the ‘good enough’ owns the customer by being able to match make between task and model, taking a cut in the process.
On paper, that should be OpenAI or Anthropic. They have the best harnesses, and the harness is what keeps a customer when the model underneath changes. Between them they own the touchpoints, ChatGPT had 900 million weekly users at OpenAI’s last official count, and Anthropic took 40% of enterprise LLM spend in 2025 according to Menlo Ventures, against OpenAI’s 27%.
But their customers are not waiting around for this. In May, Marc Benioff said Salesforce would spend around $300 million on Anthropic tokens this year (roughly $20,000 for each of its 15,000 engineers) and called for an intermediary layer that sends complex work to a frontier model and simpler work to cheaper ones. It has since been fairly widely reported that Salesforce, is building that layer itself.
The router layer is being built, by several companies who believe it to be in their best interests. To date it is being built by companies whose product was never just their own model anyway:
1) The Hyperscalers: AWS Bedrock sells Claude, Llama, Mistral and others alongside Amazon’s own Nova models. Additionally, Google’s Vertex AI offers Gemini next to third-party models, Claude included. Most interestingly of all, Microsoft’s Azure catalogue does the same, and Satya Nadella outlined to Ben Thompson that Microsoft’s agent harness will rotate through OpenAI, Anthropic and any open-weight model.
2) The platforms: As mentioned, Stripe now owns OpenRouter. Further, Meta is reportedly building a competitor to OpenRouter called Switchboard. When Reflection unveiled Beam this week (an American open-weight model it claims matches China’s GLM-5.2 at a fraction of the compute) it said it would distribute through hyperscalers and neoclouds. (not OpenAI/Anthropic)
3) SpaceXAI: Among the model makers it was SpaceX that made the first move. SpaceX absorbed xAI in an all-stock deal in February, agreed to buy Cursor for $60 billion in June and leases compute to Anthropic and Google. All of this is to say that it is a frontier lab sitting inside an infrastructure and distribution business that already sells capacity to its rivals. The perfect fit to move into the routing layer first.
A lot happened in the last week in this regard. SpaceXAI put its latest model, Grok 4.7, on Amazon Bedrock on 28 September, Google’s enterprise agent platform on 1 October and Microsoft Foundry on 7 October, the same day Musk outline Grok Bot would utilise Claude. And because Anthropic runs Claude on SpaceX’s Colossus 1 data centre, making Claude the default in Grok Bot turns SpaceXAI into both Anthropic’s supplier and its customer.
There is arguably a less flattering reading that exists. Musk launched Grok 4.5 in July as “Opus-class”; on Artificial Analysis’s index it ranked fourth, behind Claude Fable 5, GPT-5.5 and Claude Opus 4.8. Making Claude the default is partly an admission that Grok isn’t always the best answer. However, this reading is also my central point. The player behind the models has recognised the importance of opening up to keep the customer.
Interestingly, OpenAI, briefly had some of the plumbing. Its evals platform let customers call third-party models, billed at OpenRouter list prices. This evals platform will shut down next month, disappointingly so (for me) as I would have liked to see how this would have played out as OpenSource develops.
I previously outlined that the current dynamic felt analogous to the browser. To take that metaphor a step further, and a better reference for my central argument is that, the winner does not look like Google, but rather it looks like Amazon. Let me elaborate:
In the fourth quarter of 2025, third-party sellers accounted for 61% of the units sold on Amazon, many of them competing directly with Amazon’s own products. Amazon books its cut (commissions and fulfilment fees), a cut which came to $172 billion last year. Amounting to about a quarter of Amazon’s $717 billion in revenue in 2025. The marketplace also feeds an advertising business that grew 22% to $69 billion, much of it bought by those same sellers. Essentially, selling a competitor’s product keeps the customer on Amazon and pays Amazon either way.
But owning the marketplace does something far more strategically important. It tells Amazon what customers buy, in what volume and at what price. Presumably Amazon uses that to decide where to launch vertical products of its own.
I believe that is a pretty solid playbook for intelligence. Host other models alongside your own, route each task to whatever model fits it, and take margin on every token including the ones you didn’t train. Then use what the router layer tells you about which tasks run at scale and which customers will pay a premium for. Therefore, the router layer isn’t a business case on its own anymore, in the same way that I also believe that to be true for intelligence. The value is in the vertical, but the ability route to other models helps own the customer and also informs on how you choose your verticals.
Can OpenAI/Anthropic do this?
The labs own the customer at this point in time, that much is clear. Nobody opens AWS to ask a question, hundreds of millions of people open ChatGPT. Developers choose Claude for coding agents and then find it on Bedrock, not the other way round. And the line between lab and hyperscaler is blurring, for example, OpenAI is building its own data centres through Stargate, and both labs are signing compute deals that would have been hyperscaler territory three years ago.
So OpenAI and Anthropic can make this move. There are a few reasons why they haven’t:
1) Admitting defeat? Does routing a customer to a competitor’s model looks like an admission that yours isn’t always the best? Well, SpaceXAI has just shown that a frontier company can make that admission and present it as a feature, the flexibility to choose whatever gives you the best outcome.
2) Channel conflict is genuinely problematic. Anthropic sells through AWS, Google Cloud and Microsoft; OpenAI through Microsoft. Becoming a router means competing with the partners that distribute your product.
3) It is geopolitically challenging. The open models developers route to most today are Chinese, and hosting them would be untenable for a US lab such as OpenAI/Anthropic. I am watching Thinking Machines and Reflection very closely here, a growing set of credible western open models would open this up.
Ben Thompson’s view is that open models only look cheap because the labs are compute-constrained, and that once supply catches up they can win the volume back on a superior cost structure. My contention would be that that approach doesn’t solve the problem. Routers have made model choice a runtime decision, so a cheaper lab is still a supplier to whoever owns the routing layer albeit just earning less.
The prevailing narrative has been that the market for frontier intelligence is big enough for both OpenAI and Anthropic. The frontier-only market may not be big enough for two closed labs. The router-plus-vertical market is far bigger but it doesn’t belong to them by default.
By default, it belongs to the companies that already host everyone: Google, Amazon, Microsoft, and now SpaceXAI. The labs’ only way off that path is to become the router themselves, using the touchpoints they already own, and to build vertical products on what the router tells them.
The race has started, and the labs are not in it yet. If one of them joins, it has the best hand at the table. If neither does, they become the best-paid suppliers in someone else’s store..