Claude Opus 4.7 Tops Vals AI Finance Agent Benchmark at 64.37%

The Headline From May 5

At the New York City briefing on May 5, 2026, Anthropic confirmed that Claude Opus 4.7 leads the Vals AI Finance Agent benchmark at 64.37 percent, making it the most capable model on that evaluation as of that date. The number landed alongside ten pre-built finance agent templates and a Moody's data partnership announced earlier in the week.

For buyers comparing models for financial workflows, 64.37 percent is the headline. What it actually measures is more interesting than the score.

What Vals AI Measures

Vals AI runs a finance-specific agent benchmark that tests multi-step reasoning on workflows a junior analyst or associate would encounter: reading 10-Ks and 10-Qs, reconciling figures across documents, building pitchbook sections, running KYC checks, and answering questions where the right answer requires combining tabular data, footnotes, and analyst commentary.

The benchmark grades agents end-to-end, not single-shot. The model has to decide which document to open, which tool to call, when to stop, and how to present the result. That is a much harder bar than a single-turn question.

Why 64.37 Percent Is Hard

A score in the mid-60s on finance agent work is not what general-purpose benchmarks would suggest. Frontier models routinely score in the 80s and 90s on factual or reasoning tests. Vals AI is lower because finance work compounds errors.

• A 10-K has hundreds of pages. The agent has to pick the right paragraph.
• Reconciling balance sheets across years means matching restated line items, not just adding numbers.
• KYC means following a checklist where one missed flag is the whole failure.
• A pitchbook section pulls from three or four sources and must footnote each.

A model that is 90 percent reliable per step still drifts below 50 percent on a six-step workflow. 64.37 percent end-to-end is a real number.

Why Finance Is Harder Than Retail

Retail and customer-service agents have shorter contexts, fewer documents, and more forgiving ground truth. A small phrasing error in a salon booking does not cost anyone money. A small error in a KYC narrative or a leverage ratio does.

Three structural reasons finance is harder for LLM agents:

1. Document density. Filings, term sheets, and credit memos are long and structurally repetitive. The "needle" is rarely a single sentence.
2. Numerical chaining. Most outputs depend on numbers from multiple places. Each lookup is a chance to grab the wrong cell.
3. Compliance ground truth. "Mostly right" is not acceptable for KYC, disclosure, or audit narrative.

These are the same reasons general LLM demos that look magical on consumer queries quietly break on real bank desks.

Reading The Score Like A Buyer

If you are evaluating models for finance work, the useful framing is not "Opus 4.7 is the best." It is "the ceiling for unsupervised single-agent finance work is in the mid-60s today, and Opus 4.7 is at that ceiling."

That has three implications:

• Human review stays. A 64 percent benchmark score means at least one in three end-to-end workflows needs a human in the loop. Most production deployments at JPMorgan Chase, Goldman Sachs, Citi, AIG, and Visa already reflect this.
• Tool design matters more than model swapping. The gap between vendors is narrower than the gap between a workflow with the right tools and one without. Cookbooks for Claude Managed Agents (also announced this week) target exactly this.
• Vertical templates compound the model. Anthropic's ten pre-built finance agent templates for Claude Cowork and Claude Code wrap Opus 4.7 in workflow structure, which is where most of the real lift comes from.

What Changed With Opus 4.7

The Opus 4.7 generation introduced longer context windows, stronger tool-use reliability, and better grounded-citation behavior. The Moody's data partnership announced the same week gives Claude grounded financial reference data, which historically is where general models hallucinate.

In short: better model plus better data plumbing plus task-specific templates. That combination is what moves the Vals AI score from "interesting" to "useful in production."

Where CallSphere Sits Next To This News

CallSphere is an AI voice and chat agent platform for customer-facing communication: voice, chat, SMS, and WhatsApp, in 57 plus languages, with around 14 function tools and a HIPAA-friendly posture. We are adjacent to finance back-office agents, not in the same category.

Where this news matters for CallSphere customers:

• The same engineering disciplines that produce 64.37 percent on Vals AI also produce reliable front-door voice agents. Tool design, grounding, and human escalation patterns transfer.
• Banks, insurers, and wealth managers using Claude in the back office still need a voice and chat front door for customers, applicants, and policy holders. CallSphere fills that slot.
• Our pricing is transparent: Act at $49 per month for single-system task completion, Orchestrate at $99 for coordinating across multiple systems, Custom Build at $149 for bespoke apps and unlimited development. Launch takes 3 to 5 business days, with a free trial.

If you want to see a customer-facing agent live before you decide, book a demo.

FAQ

Q: Is 64.37 percent good or bad? For end-to-end finance agent work, it is the current state of the art. Single-turn benchmarks score higher. Vals AI is harder because it grades the whole workflow.

Q: Does Opus 4.7 replace human analysts? No. A 64 percent score means a human still reviews and finishes most workflows. The lift is in speed and first-draft quality, not full autonomy.

Q: Is CallSphere built on Claude Opus 4.7? CallSphere is model-flexible. Where Claude is the best fit for a customer workflow, we use it. The voice path uses real-time speech models tuned for low latency.