V7 cuts costs 78% while boosting accuracy with GPT‑5.6 Luna
Read the Portuguese translation →
Author: OpenAI · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: V7 cuts costs 78% while boosting accuracy with GPT‑5.6 Luna · Channel: OpenAI · 01:32 · Watch on YouTube
Topics: AI agents, Task and workflow automation, AI evaluation and benchmarks, Prompt and context engineering, RAG, search and information retrieval
Summary
Section titled “Summary”The speakers say newer GPT models can handle complex enterprise workflows, using long instruction documents to build end-to-end processes with citations. They describe V7 Go’s workflow engine and context graph as supporting persistent company memory, with benchmark errors and hallucinations reduced when used with the newer models. They also report faster research, less need to re-ingest data, and better investment memos. Together, they say these capabilities make previously impractical enterprise use cases possible.
Key points
Section titled “Key points”- The speakers say earlier models struggled with complex processes such as insurance underwriting and screening finance deals.
- Newer GPT models can ingest up to 100 pages of instructions and build workflows to address tasks end to end.
- V7 Go helps create long-horizon workflows with citations that anchor facts.
- The context graph stores information about entities, facts, and relationships according to an ontology matching company data.
- The speakers report that tool-call errors in their context graph benchmark fell from 2.7% to 0.2%.
- They report that hallucinations fell by 38% with GPT Luna on the V7 Go context graph.
- The speakers say document visual understanding is improving.
- Company research that previously took minutes now takes seconds, without re-ingesting all the data from scratch.
- The speakers describe improved depth and accuracy in investment memos.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Building enterprise workflows | The speakers describe newer GPT models handling complex enterprise processes using long instructions and citation-backed workflows. |
| 00:30 | Company memory and context graph | They discuss a company memory system, benchmark results, and how the context graph organizes entities, facts, and relationships. |
| 01:01 | Fewer hallucinations and faster research | The speakers report fewer hallucinations, improved document understanding, faster company research, and better investment memo quality. |
Quotes
Section titled “Quotes”build out a workflow that can solve them end to end. — 00:00
citations that anchor every fact. — 00:00
errors in our context graph benchmark dropped from 2.7% to just 0.2%. — 00:30
we’ve seen hallucinations reduced by 38%. — 01:01
the quality of the output improve. — 01:32
Concepts and entities
Section titled “Concepts and entities”GPT-5.6, GPT Luna, V7 Go, Context Graph, ontology, RAG, OpenAI, tool calling
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 With GPT 5.6 we seen huge improvements in the build out of enterprise workflows with a B7 Go. A few months ago the models couldn’t really understand complex processes like insurance underwriting or screening of finance deal, but the latest GPT models can ingest 100 pages of instructions and build out a workflow that can solve them end to end. B7 Go helps GPT models develop these long horizon workflows with citations that anchor every fact.
00:30 >> And recently a memory system that remembers everything about your company. >> We benchmark every new model and with GPT 5.6 sold we noticed that tool called errors in our context graph benchmark dropped from 2.7% to just 0.2%. >> Yeah, the model’s only half the equation because the other half is the platform where these models run on. And our context graph store all this information about entities, facts, relationships following an ontology that matches your data. Previously models used to make ingestion
01:01 errors especially whether fact was worth remembering or not. >> Older models using rag would also try and guess unanswerable queries. With GPT Luna on V7 Go context graph we’ve seen hallucinations reduced by 38%. Visual understanding of documents is also getting much better. >> And we’re seeing the same benefits in workflows as well. I were previously we took minutes to research companies information and then just give the user an output. Now we take just a few seconds and we don’t need to re-ingest
01:32 all the data from scratch. >> We’re also seeing the quality of the output improve. So previous models used to generate decent investment memos, but they lack the depth and accuracy that we’re seeing with the new models. >> Yeah, so the GPT 5.6 suite is letting us solve use cases that a year ago were impractical or even impossible. And combining that within a workflow engine and a memory system we can help these enterprises that are known for their slowness in retooling towards the AGI. And we’re thrilled to be working with OpenAI towards that.