SAN FRANCISCO — First-day API telemetry on Claude Helios released by Anthropic on Wednesday afternoon showed pharmaceutical and legal-services workloads accounting for the largest share of inference volume during the model’s initial 27 hours of enterprise availability, an early indication of where buyers see the new model’s stronger reasoning capabilities translating into tangible workflows.

The summary, published in the form of a four-page transparency note on the company’s developer site, reported that the Helios endpoint had processed roughly 1.7 billion input tokens and 690 million output tokens since its 9 a.m. Pacific launch Tuesday, with median latency at 4.6 seconds for prompts under 4,000 tokens and 14.1 seconds for longer-context calls. Pharmaceutical-vertical customers accounted for 23 percent of inference volume by token count, legal services for 19 percent, and financial services for 14 percent.

The disclosure was the first of what Anthropic said would be a weekly cadence of operational summaries during the launch period, a transparency commitment the company had described in Friday’s framework signing but had not previously specified in detail. The summaries are intended to feed into the new Frontier Model Assurance Council’s monitoring brief and, the company said, to provide a public record of usage patterns that academic researchers and regulators could use to ground policy work.

“What you are seeing in the first 27 hours is not a final answer about the shape of Helios usage,” said Dario Amodei, Anthropic’s chief executive, in a brief video address posted alongside the disclosure. “But it is consistent with what we have been hearing from beta customers for the past three months. The model is differentially good at multistep, evidence-bound reasoning, and that is where the early enterprise demand is concentrating.”

The pharmaceutical demand was concentrated, the disclosure showed, in three sub-segments: literature-synthesis workflows, in which Helios reads dozens or hundreds of trial reports and produces structured summaries; protocol-drafting workflows for new clinical investigations; and a smaller but rapidly growing category of safety-monitoring workflows in which the model is asked to surface anomalies in adverse-event reporting. Recursion Pharmaceuticals, one of the early-adopter customers Anthropic had named in Tuesday’s launch announcement, was identified in the disclosure as a representative pharmaceutical user; Recursion, in a separate statement Wednesday, described the model’s performance on its internal evaluation set as “the first qualitative step-change we have seen since the original Claude 4 family.”

Legal-services demand traced a similar pattern. Latham & Watkins, also named Tuesday, was the single largest legal-vertical user by inference volume. The firm’s chief information officer, in a Wednesday LinkedIn post that was circulated by Anthropic, said Helios was being used in three areas — diligence review on cross-border transactions, regulatory-mapping workflows for clients facing the new wave of AI legislation in the United States and the European Union, and a pilot for what the firm called “principled second-pair-of-eyes review” on contractual drafting. The post emphasized that all use was governed by the firm’s existing AI policy, which requires attorney sign-off on any output going to a client.

Financial-services usage, while smaller by volume than pharma or legal, was the most concentrated by spend, the disclosure indicated. The category was dominated by what Anthropic described as “retrieval-augmented research workflows” — calls that combine the model with the customer’s own document store and surface analytical syntheses against internal questions. The disclosure did not name financial-vertical customers but indicated that “more than a dozen” of the company’s existing enterprise accounts had moved meaningful workload onto the new endpoint within 24 hours.

A senior research engineer at Anthropic, briefing reporters on background Wednesday afternoon, said the most surprising signal from the first day’s usage was the meaningful pickup in long-context calls. Roughly 28 percent of inference by token count came from prompts above 16,000 tokens, well above the equivalent share for the outgoing Claude Opus 4 endpoint. The engineer attributed the shift to the model’s improved performance on long-document workflows, citing internal evaluations on the company’s needle-in-haystack benchmarks.

Inference pricing was set at $5 per million input tokens and $25 per million output tokens, modestly above the outgoing Opus 4 tier. Several sell-side analysts noted that the pricing had not appeared to constrain initial uptake; one Morgan Stanley analyst, in a Wednesday note, estimated that Anthropic was on track to deliver roughly $9.2 million in incremental run-rate revenue from Helios in the first week of availability if the current usage pace held.

The disclosure also included a brief operational incident report. A six-minute service interruption occurred at 4:14 a.m. Pacific on Wednesday morning, the company said, when a routing layer in the company’s primary serving cluster experienced a memory-pressure event. The incident was contained, no requests were lost beyond the immediate retry window, and the post-mortem was being conducted with the new assurance council’s interim staff as observers — a procedural innovation the framework had introduced.

The interim director of the assurance council, former federal chief information officer Pavithra Ramaswamy, in a brief statement issued Wednesday evening, said the council had reviewed the operational disclosure and the post-mortem and “considers the level of detail provided to be consistent with the framework’s intent.” She added that the council expected to publish its own monthly aggregated summary, separate from the per-company weekly disclosures, beginning in June.

The Wednesday disclosure also touched on what Anthropic described as “early signals on misuse and refusal patterns.” The company said its automated moderation pipeline had handled 41 distinct refusal events in the first 27 hours that the safety team had flagged for additional review, including 14 in categories the company’s safety policy lists as Tier-2 (potentially dual-use) and 3 in Tier-3 (high-severity). None of the Tier-3 events had resulted in capability disclosure, the company said. The reviews are being shared with the council under the framework’s evaluation-sharing rules.

Consumer access through the Claude.ai product is scheduled to begin May 26, after the staged-exposure window. A senior product official at the company said the consumer rollout would include “additional friction” on workflows that the safety team had flagged in the enterprise window, and that the consumer release was being designed to allow a quick rollback if the council or internal safety processes identified new concerns.

The broader frontier-model release calendar has been compressed. OpenAI, which has indicated it expects to release its next-generation model in the third quarter, has not yet specified whether its release will move through the same 60-day notification window or proceed under the more flexible terms the framework permits for pre-existing in-flight development. Google DeepMind’s next Gemini release is expected in the same window. Meta, the most reluctant of the framework signatories, has continued to indicate that its open-weight release model would receive a separate evaluation track.