{
  "schema_version": "1.1",
  "id": "s2:https://openai.com/business/customer-stories/hebbia",
  "slug": "hebbia-0og5kbd",
  "url": "https://feed7.dev/p/hebbia-0og5kbd",
  "title": "Hebbia sees improvements in financial research with GPT-6 Astra",
  "why_included": "Hebbia reports GPT-6 Astra improved source coverage, citation relevance, and presentation adherence in prerelease financial-research tests. The results suggest where to evaluate models in document-heavy workflows.",
  "summary": "In Hebbia’s August–September 2026 prerelease tests, GPT-6 Astra led the best alternative by **13.6 percentage points** on required-source coverage. **99.5%** of its cited excerpts were relevant, and it met **97.9%** of applicable presentation requirements, a **14.4-point lead** on that measure.",
  "practical_implication": "For agent workflows that turn large document sets into client-facing artifacts, evaluate source coverage and citation relevance separately from instruction following and slide quality. Keep links from claims and generated slides back to the supporting passages so reviewers can check key figures.",
  "agent_context": "In Hebbia’s August–September 2026 prerelease tests, GPT-6 Astra led the best alternative by **13.6 percentage points** on required-source coverage. **99.5%** of its cited excerpts were relevant, and it met **97.9%** of applicable presentation requirements, a **14.4-point lead** on that measure.\n\nFor agent workflows that turn large document sets into client-facing artifacts, evaluate source coverage and citation relevance separately from instruction following and slide quality. Keep links from claims and generated slides back to the supporting passages so reviewers can check key figures.\n\nThese are Hebbia’s results on its own financial-research cases, using domain-expert criteria; they do not establish performance on other workflows or datasets. The acquisition data-room walkthrough is hypothetical.",
  "source": {
    "name": "OpenAI",
    "url": "https://openai.com/business/customer-stories/hebbia",
    "published_at": "2026-11-20T00:00:00.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Official Release",
  "layer": "model",
  "domains": [
    "research",
    "data"
  ],
  "topics": [
    "model-selection",
    "agent-evals"
  ],
  "verification": {
    "status": "official_source",
    "label": "Official Source",
    "method": "source_feed",
    "verified_at": null
  },
  "uncertainty": [
    "These are Hebbia’s results on its own financial-research cases, using domain-expert criteria; they do not establish performance on other workflows or datasets. The acquisition data-room walkthrough is hypothetical."
  ],
  "connected_context": {
    "meaning": "Hebbia adds task-specific evidence for evaluating document agents: source coverage, citation relevance, and presentation compliance are distinct measures. Its results complement Legora’s multi-document review account, whose reported workflow improvement lacks a specified measure, while remaining limited to Hebbia’s own financial-research cases.",
    "corpus_size": 1135,
    "generated_at": "2026-11-21T11:05:58.543Z",
    "connections": [
      {
        "title": "Legora reviewed 41 documents in minutes with GPT-6 Astra",
        "source_name": "OpenAI",
        "source_url": "https://openai.com/index/legora-financial-statement-review-with-astra",
        "feed7_url": "https://feed7.dev/p/legora-financial-statement-review-with-astra-1muuhfh",
        "reason": "Both report GPT-6 Astra on multi-document financial work; Hebbia specifies source coverage, citation relevance, and presentation criteria, while Legora’s workflow improvement measure is unspecified."
      },
      {
        "title": "Introducing Virtual Models on AI Gateway",
        "source_name": "Vercel",
        "source_url": "https://vercel.com/blog/ai-gateway-virtual-models",
        "feed7_url": "https://feed7.dev/p/ai-gateway-virtual-models-0563pz8",
        "reason": "Hebbia’s separate quality measures provide task-specific criteria that teams could evaluate before switching a production task to a candidate model."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-11-20T00:00:00.000Z",
  "modified_at": "2026-11-20T00:00:00.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/hebbia-0og5kbd",
    "json": "https://feed7.dev/p/hebbia-0og5kbd.json",
    "markdown": "https://feed7.dev/p/hebbia-0og5kbd.md"
  }
}