More than half of the Fortune 500 are substantial paying Anthropic customers. Yet Anthropic says most enterprise deployments still cover only a few teams and use cases.
The customer count is a milestone. The gap between buying Claude and using it throughout a company is the more interesting part of the filing.
Buying Claude Is Different From Using It Widely
Among publicly traded customers in that Fortune 500 group, average September-quarter spending was below $1 for every $1,000 in total expenses. Anthropic’s 10 largest Fortune 500 customers spent more than eight times the group’s average share of expenses. Even that top group averaged below 1% of expenses.
Those figures fit Anthropic’s description of deployments limited to a few teams and use cases. A company can be a substantial customer while Claude still touches a small part of its work.
There are signs of broader use. In the September quarter, more than half of customers generating at least $100,000 in trailing annual revenue used both Anthropic’s platform and one of its applications. The comparable share was about 30% in the last quarter of 2025.
Cost per task is moving too, at least by Anthropic’s account. It says Opus 5.5 cut that cost by about 40% compared with Opus 5 through lower pricing and better task efficiency. In Ramp’s sample across labs, token volume rose about 52% between the weeks ending September 13 and October 11, while spending rose 23%.
Lower task cost and use of both products may help explain how usage can grow. They do not tell us how much work it takes to fit Claude into more of a company’s workflows. Anthropic has committed $100 million to a training program with a goal of 10,000 engineers by the end of next year. One commentator says the filing leaves open how much continuing workflow help customers will need.
Google’s TPU Lead Is an Estimated Cost Result
Two official-preview results put Google’s TPU setup ahead on estimated performance per dollar. SemiAnalysis reports 1.42× on an outside customer’s cost basis. Office Chai reports 1.67× for a first submission.
The limits belong beside those numbers. The test replayed traces from real Claude code use, but ran a stand-in model. Cost was estimated rather than taken from a paid bill.
It also compared whole setups: four Google TPUs against Nvidia’s 72 GPU rack. Google’s TPU ran FP8, while Nvidia’s setup ran FP4. That makes the result useful as a cost comparison for these setups, with those choices, rather than a clean comparison of the chips alone.
The preview does not establish that answers arrived faster. Anthropic has announced plans to use up to a million Google TPUs and cited price performance. Whether the benchmark’s price lead survives validation remains open.
The Byte-Model Lead Exists in Fitted Curves
A token model reads common chunks of text. A byte model reads one byte at a time, giving it many more positions to process. In this study, byte students learned from predictions made by an 8 billion parameter Llama 3 teacher and received a marker showing where each teacher token ended.
At small training budgets, the token student was ahead. As the fitted curves extended, the byte students kept improving while the token curve leveled off. The reported lead comes from those curves, not from trained models reaching the projected scores.
At one learning-rate setting, the curves predict average accuracy of 52.4 for the byte student with end-of-token markers, against 48.4 for the token student. On an alternative validation set, the projected gap narrows to 50.5 against 48.7, according to a summary whose figure has not been matched to the paper’s table.
There is a practical limit beyond the uncertainty in those projections. The authors say the end-of-token models remain significantly expensive at inference time. The paper also did not test whether its byte students spell better.
A Browser Agent’s Win Has a Clear Boundary
Eagle Light lets an agent run a JavaScript program inside a browser where the user is already signed in. Each task gets its own space and tabs, so the agent can read a dashboard without taking over the person’s tabs.
In one reported run, it gathered plan, usage, and reset dates from six dashboards in 2 minutes and 27 seconds. A narrower comparison used the same prompt and model on two dashboards, ran each tool three times, and kept the middle run.
In that comparison, Claude in Chrome took 92 seconds, 792,000 tokens, and 16 tool calls. Eagle Light took 48 seconds, 290,000 tokens, and six tool calls. Eagle Light was faster and used fewer tokens and calls in this task. EGO’s separate benchmark also favors it, but uses EGO’s own tasks.
The billing example gives the agent a clearer boundary than a speed result does. Codex opened a Supabase billing page and selected a paid plan, then stopped before the payment button. My rule is to let the agent read and draft while a person handles anything that sends or pays.
That rule is concrete. Anthropic’s larger open question is less settled: how far Claude will spread within its customers, and how much continuing help that work will take.