Source: ai-research/claude-blog-hebbia-financial-diligence.md — Anthropic first-party “Working at the Frontier” customer case study, claude.com/blog/working-at-the-frontier-how-hebbia-builds-ai-for-financial-diligence-that-cant-miss-a-detail (fetched 2026-07-16; exact publish date not captured in extraction). Part of the same July 2026 case-study rollout as the sibling posts on Cognition and Thomson Reuters, both fetched and ingested the same day.
Hebbia builds Matrix, a research and diligence platform for institutional finance serving more than a third of the top 50 asset managers plus tier-1 investment banks and law firms. It runs every new Claude model through its own finance-specific benchmark before adopting it, and Claude Fable 5 posted the largest accuracy gain Hebbia’s applied AI research team has measured — on both halves of that benchmark. This article covers the two-test methodology, the specific gains, the named researchers behind the eval, and how Hebbia is restructuring its product around the Claude Agent SDK to push further into multi-step diligence work.
Key Takeaways
- ~20% relative accuracy gain on Hebbia’s question-answering/citation-finding test over financial documents — “the best he had seen from any new model,” per researcher Joe Renner. Citation match held roughly steady, which the team attributes to the model better understanding the evidence it finds, not just retrieving more of it.
- On the agent-system test (open-ended, multi-source analysis mirroring real customer workflows), Fable 5 held every part of a multi-part request in view simultaneously and cited every answer back to its source — a query-dropping failure mode prior models exhibited on complex, multi-part requests.
- Hebbia is adopting the Claude Agent SDK to decompose diligence jobs into smaller, repeatable, checked steps rather than a single large model call — “no matter how brilliant the model is,” firms still want control over which documents feed each step of an analysis.
- Concrete before/after economics, per founding PM Divya Mehta: a pitch-deck research job took a junior banker 2-3 days pre-AI; pre-Opus AI models compressed that by 12-24 hours; earlier Opus models dropped it further to about a day end-to-end; Hebbia’s Matrix plus Fable 5 now produces the deck, financial model, and internal research in a couple of minutes.
- Customer conversations have shifted from defensive to ambitious over two to three years — from “is the math right?” to “how do I sequence more steps and generate ten to twenty slide decks in one click with high fidelity?”
How Hebbia holds the line on accuracy
Hebbia’s customers make investment decisions based on analyses spanning thousands of dense documents — public filings, credit agreements, internal documents, structured CRM data — where a wrong number can change the outcome of an entire deal. Hebbia’s meta-prompting layer turns a plain-language request into a set of prompts, and Claude runs each step of the analysis across hundreds of documents at once. Each answer lands in its own cell on a grid in Hebbia’s Matrix product, giving full transparency, traceability, and steerability into how a conclusion was reached.
Keeping those answers accurate at scale is the job of Hebbia’s applied AI research team, led by Adithya Ramanathan. His framing: the point of the work is finding signals — getting a model to draw on the right data, in the right context, and surface what a customer actually wants to know. “When you’re connecting it to the right data and putting it in the right ecosystem, that’s when you get the alpha that finance professionals actually chase.”
The team runs every new model through Hebbia’s own finance-specific benchmark, head-to-head against the model it would replace, expanding what the benchmark measures with every release. Founding product manager Divya Mehta: “The bar is extremely high, and our customers hold us to that extremely high bar — and rightfully so. At the end of the day, they’re making investment decisions at a very large scale based on the analysis and final work product built in Hebbia.”
Clearing Hebbia’s evals by the widest margin yet
Researcher Joe Renner runs each new Claude model against two tests replicating real finance knowledge-worker use cases:
- Question-answering and citation-finding over financial documents.
- An agent-system run through the tools Hebbia’s chat product actually uses, on open-ended, multi-source analysis.
Claude Fable 5 cleared both by the widest margin Renner had measured. On the first test it posted roughly a 20% relative accuracy gain — the best he’d seen from any new model release — while citation match held steady, which the team reads as better comprehension of the evidence rather than just better retrieval. Mehta: “It comes down to two seemingly fundamental qualities: the ability to find the right information from a dense data set, and then synthesize it correctly. These seem like fundamental model capabilities, but they have massive impact when we think about finance and research workflows.”
On the agent run, Fable 5 held every part of a multi-part request at once, answering all of them and citing each back to its source. It also showed more reach on open-ended analysis, reasoning from a wider cross-section of the data and surfacing conclusions the team thought were worth a closer look. Renner traces this to how the model holds a long task together: it keeps every part of a request in view, prompts its own sub-agents and tools so the right facts come back, and grounds each claim in the source rather than inferring it.
Setting a new standard for deal diligence
The information that gives customers an edge usually sits in unstructured, proprietary documents — historically harder to analyze at scale than the structured, quantitative data finance already models well. That might be a data room with thousands of documents (find the signal, cite it, draft each section of an investment memo) or a credit deal’s full document set (the credit agreement, amendments, side letters, each running hundreds of dense technical pages), where the job is extracting the full covenant package and operating restrictions from an unstructured mass. “These are actually the types of documents that Anthropic models have always done really well at,” Mehta says.
Earlier Sonnet and Opus models could already pull out and synthesize a credit agreement’s covenants inside Matrix. With Fable 5, Hebbia is reaching for the rest of the job: the multi-step analysis on top of those covenants, comparing them against live monitoring data, flagging risks, all the way to a first draft of the covenant review and internal memo — work credit firms used to pay outside teams a great deal to produce by hand.
The economics of compression
| Era | Time for a first-draft pitch deck |
|---|---|
| Before AI | 2-3 days (junior banker learns the company, pulls financials, builds slides) |
| Pre-Opus AI models | Compressed by 12-24 hours off the baseline |
| Earlier Opus models on Hebbia | About a day, end-to-end |
| Matrix + Claude Fable 5 (today) | A couple of minutes for the deck, financial model, and internal research |
Hebbia has codified the whole job into a Matrix that gathers data across sources in deterministic agentic steps, runs the analysis, and builds the final deliverables — freeing the banker to spend time on which buyers to pursue and how to position them, rather than on document assembly. Mehta: “Compressing the deal lifecycle has a massive impact on a firm’s ability to compete for those investments.”
Try It
- If evaluating a new model for a high-stakes domain, build a domain-specific benchmark before adopting — Hebbia’s approach (two tests: direct QA/citation-finding, plus an agent-system run through real product tooling) is a concrete, replicable template for any team weighing a model swap in a regulated or high-precision workflow.
- Decompose agentic work into checked steps even when the model is capable enough to one-shot it. Hebbia’s move to the Claude Agent SDK is explicitly about retaining control over which documents feed which step — a discipline worth adopting regardless of how strong the underlying model is.
- Watch for the “citation match holds steady while comprehension improves” pattern as a specific signal of a genuine capability gain, distinct from a model just getting better at surface-level retrieval.
Open Questions
- The exact publish date of this post was not captured in the Tavily extraction — it sits in the same July 2026 “Working at the Frontier” case-study series as the Cognition and Thomson Reuters posts, but no explicit dateline was visible on the fetched page (unlike the Cognition post, which carried a “July 10, 2026” dateline).
- Hebbia’s finance-specific benchmark is described but not named or published — unclear whether it is proprietary-only or will see any public disclosure.
- No pricing, seat count, or contract-value detail is given for Hebbia’s Anthropic relationship.
Related
- Claude Fable 5 and Claude Mythos 5 — the main model-release article this case study substantiates with a named, concrete customer example (the topic index’s brief “Hebbia finance” benchmark mention this article expands into full detail).
- Claude Agent SDK — the SDK Hebbia is adopting to decompose diligence workflows into smaller, checked steps.
- Working at the Frontier: How Cognition Trusts Claude Fable 5 to Work Through the Night — sibling case study in the same series: another domain expert running its own benchmark against a new Claude release, with concrete before/after economics.
- Working at the Frontier: How Thomson Reuters Builds AI for High-Stakes Professional Work — sibling case study, a second high-stakes professional-work domain (legal vs. finance) with a strikingly similar citation-verification-first framing.
- The Verification Frontier — Hebbia’s “grounds each claim in the source rather than inferring it” is a direct, named instance of this thesis in a live commercial product.
- Picking the Right Model — Hebbia’s own-benchmark-first adoption methodology is a concrete real-world example of this article’s general framework.