OpenAI takes the enterprise spending lead — and OpenRouter flips
Two competitive moves and one rule change: OpenAI's momentum now shows up in spending data, and India just put AI voice agents inside its anti-spam regime.
Astra accounted for about 13 percent of enterprise AI spending tracked by the corporate expense platform Ramp, against roughly 8 percent for Claude Fable — and OpenRouter said its users spent more on OpenAI models than on Anthropic's last week, the first time OpenAI has led that measure in more than two and a half years. Both numbers come from Reuters, which reported them alongside the news that Anthropic is now considering releasing a new model to answer the momentum OpenAI has picked up since GPT-6 Astra launched.
That is the shape of the decision. The market Anthropic has been described as leading for months is being contested on the spending data investors actually watch, and the counter-argument is scale and stickiness rather than a benchmark. Anthropic's annualized revenue run rate topped $65 billion by the end of July, up from about $9 billion at the end of 2025, with a projected roughly $190 billion to $200 billion for 2028; OpenAI's run rate passed $40 billion in July. Switching costs inside large enterprises are measured in quarters, not weeks, and lead changes in this market have historically been temporary.
The deliberation is hard because it is two decisions at once. Anthropic is weighing how to hold its enterprise position and how to strengthen profitability at a moment when rising rates have investors focused on the timing of profit rather than the size of the bet, per the people familiar with the company's thinking. It is also evaluating the safety of its next model as part of that process, one person said, and declined to comment. A lab whose chief executive spent September arguing the industry should slow down has to explain the sequence if it ships early — and a lab that stays quiet while accounts move has to explain that instead.
There is an IPO clock on top of it. Anthropic could push its listing to after the November US midterm elections, two people familiar said, with the elections not expected to matter much for the offering; marketing had previously been expected to begin in mid-October at the earliest. OpenAI removed itself from that race when Sam Altman confirmed it would not go public in 2026. We covered the earlier turn in this contest — Anthropic passes OpenAI on revenue for the first time — and the sequence since is the useful reminder: a revenue lead and a spending lead can belong to two different companies at the same time.
India's telecom regulator amended its anti-spam rules to pull AI-generated and automated voice calls into the commercial-calling regime, requiring every entity that makes them to pre-declare the numbers it will use to its telecom operator. Under the Third Amendment Regulations published Friday, an "application-to-person" call is any voice call initiated by software, an automated platform or autodialing without direct human dialing, including robocalls and prerecorded or artificial voices. Calls made without that declaration are treated as spam, and operators may levy a termination charge of up to 5 paise — about 0.052 cents — per minute on them.
The same amendment reaches the apps that sit on top of the network. Caller-ID and call-management apps can no longer offer users a "report spam" button unless those reports are forwarded to the DLT platform the operators run, which is the change Truecaller objected to, calling it a one-way transfer of commercially valuable data that is anti-competitive. Apps are also barred from blanket-blocking or spam-tagging calls from the 140xx and 1600xx series reserved for promotional, service and government communications; individual users can still block them on their own devices. The regulator's stated reason is the risk of mislabeling, and the practical effect is that the largest independent spam signal in the market now feeds the carriers' own enforcement system.
What is unresolved is what "report" means. Policy researchers told TechCrunch it is unclear whether apps must transmit a single user's report or the reputation data and analytical signals they build from millions of them, and TRAI did not answer questions about scope or about whether phone dialers and operating-system call features are covered. There is a second hole at the edge of the A2P definition: the test is how a call is initiated, not whether a synthetic voice is on the line, which leaves contact-center software and click-to-call services in an ambiguous middle. India is the first large market to write AI voice agents into a spam regime rather than a safety one, and the compliance question there is not capability but bookkeeping.
What to watch: whether TRAI answers the scope questions in guidance, and whether other regulators copy a rule that turns spam enforcement into an AI-calling registry.
If a rival's model starts taking your enterprise accounts, does a safety-first lab ship early or hold the line? Tell us in the comments.
