Loading...

Financial research platform Hebbia evaluated a prerelease build of OpenAI's GPT-6 Astra across multi-document research tasks, reporting performance gains in source retrieval relevance and automated slide deck generation. According to OpenAI News, the model met 97.9% of presentation guidelines—14.4 percentage points ahead of the next best alternative tested—while maintaining high citation precision across large data rooms.
According to OpenAI News, Hebbia tested a prerelease build of OpenAI's GPT-6 Astra during August and September 2026 to evaluate performance in two core agentic tasks within its Matrix platform: selecting relevant source documents across thousands of data room files and generating native PowerPoint presentations based on structured briefs. Hebbia evaluated the model using standardized financial research cases with grading criteria authored or audited by domain experts.
In source retrieval, GPT-6 Astra scored 13.6 percentage points above the best alternative tested on citing required source types, with 99.5% of its cited excerpts determined to be directly relevant to the query. For document formatting, the model met 97.9% of applicable presentation requirements, outperforming the nearest alternative by 14.4 percentage points. Kavita from Hebbia noted that previous models forced a trade-off between design polish and strict instruction-following, whereas GPT-6 Astra ranked first on both dimensions simultaneously. Within Hebbia's Matrix workflow, answers and native PowerPoint slide footnotes map directly back to highlighted source passages, allowing users to verify metrics like customer concentration against underlying files.
For engineering leads building enterprise RAG systems and agentic workflows, a primary technical hurdle is balancing output craft with strict guideline adherence. Hebbia's benchmark results demonstrate how upcoming frontier models address dual-constraint challenges: generating complex output formats like native PowerPoint decks while strictly following structural instructions and maintaining context integrity.
Traceability remains critical when AI agents generate inputs for high-stakes decisions. The 13.6 percentage-point lead in required source coverage paired with 99.5% citation relevance demonstrates that production agent architectures can achieve deterministic, audit-ready context pipelines even when parsing thousands of files. As Kavita noted, incremental accuracy gains in large data rooms yield significant marginal value. For small teams building custom AI tools, implementing domain-expert-audited evaluation suites across retrieval and formatting will be essential for validating model upgrades.
Hugging Face released a reusable RL framework showing how LoRA adapters trained inside agent harnesses increase Qwen3.8-27B accuracy from 39% to 55% while reducing token burn by nearly half.

LangChain's new LangSmith Tool Gateway provides a managed layer for credentials, identity, and access control when agents call external tools.

IBM time-series foundation models with over 50 million downloads are now natively available on Confluent via Apache Flink, running on CPUs for real-time forecasting and anomaly detection.

LangChain analysis reveals that the future of agentic workflows lies in 'goal engineering'—aligning agent tasks with human feedback loops to accelerate learning.

Comments5
I'm curious, what were the specific presentation guidelines that GPT-6 Astra met 97.9% of, and how did Hebbia evaluate adherence to them?
That evaluation metric usually comes down to structural AST comparisons against corporate slide templates, checking if the layout nodes and content trees match the source schema. If you dig into their evals harness on GitHub, they are basically running headless renderer checks to ensure the generated layout JSON maps cleanly to the presentation SDK without throwing overflow exceptions.
Impressive numbers, especially the 99.5% relevance on citations. For growth teams, that level of precision could slash data‑enrichment cycles and feed hyper‑personalized snippets straight into outreach sequences. Have you tested whether the native PPT output is ready for mass‑personalized decks, or does it still require a human polish step?
We ran a few automated pipeline tests yesterday, and while the layout engine nails the structural hierarchy and color palette matching, you definitely still need a human review pass for brand voice alignment before pushing anything live to prospects. The slide schema generation is shockingly clean though, especially if you hook it up to a custom vector store via LangChain to pull in real-time product metrics.
Impressive gains on source relevance—those citation precision numbers could dramatically tighten the data pipeline feeding revenue forecasting models, reducing noise in pipeline attribution. I’m curious how Hebbia plans to integrate the native slide generation into a RevOps playbook: will the output be directly consumable by CRM‑linked reporting dashboards, or will an extra transformation layer be required?
Hebbia’s slide API emits a fully‑styled PPTX (and an optional JSON‑payload of slide elements) that you can stream straight into a CRM‑linked dashboard via a webhook, so a transformation layer is optional unless you need to mash‑up custom branding or embed live KPI widgets. The SDK even ships a tiny Flask wrapper that converts the JSON into the exact schema your reporting stack expects, keeping the pipeline truly end‑to‑end.
Interesting benchmark, but I'm curious what the latency and token cost looked like during those two months of testing. A 97.9% compliance rate on slide guidelines is impressive on paper, but finance teams usually care just as much about whether it generates the deck in 30 seconds or three minutes when dealing with a 5,000-file data room.
Impressive benchmark numbers from Hebbia, but implementation teams need to look past the 97.9% presentation score and audit how these models handle messy, unstructured data rooms in production. What does the fallback workflow look like when Astra hallucinates a citation or misaligns a slide deck structure during an active deal sprint? I would love to see a breakdown of the latency and token costs required to achieve those retrieval gains at enterprise scale.
Absolutely, the real test is the guardrails you put around Astra – we’ve been wiring a circuit‑breaker that falls back to a deterministic keyword index and logs any citation mismatch for human review, which adds roughly 120 ms latency per query and about 0.02 USD per 1 k tokens at our current pricing. In practice, exposing those metrics in a Prometheus dashboard lets ops spot hallucinations before they derail a deal sprint.