OpenAI has published a case study describing how V7, an AI platform for enterprise workflows, uses a suite of OpenAI models — including GPT-5.6 Luna, GPT-5.6 Terra, GPT-5.6 Sol, and the newer GPT-6 Astra — to power its V7 Go agentic platform. The post, authored by OpenAI, outlines how V7 Go builds a "Context Graph" that connects entities, relationships, and cited evidence across millions of documents, enabling agents to execute 50–100 step workflows in minutes with claimed 99.9% accuracy.
V7, founded in 2018 by Alberto Rizzoli and Simon Edwardsson, targets finance, insurance, and real estate teams where retrieval accuracy inside workflows is critical. The company says its Context Graph is an order of magnitude cheaper and faster to traverse than long-context approaches, and that on the HERB benchmark for cross-system information retrieval, its retrieval-only system outperformed the official baseline by 69% while reducing hallucinations on unanswerable queries by 38%.
Confirmed
- V7 Go uses GPT-5.6 Luna for high-volume structured extraction across millions of files.
- GPT-5.6 Terra and Sol power reasoning, tool use, chat, and the Go Agent path for complex multi-step workflows.
- GPT-6 Astra is being tested on the most demanding Context Graph queries, including financial analysis across thousands of documents.
- V7 reports that GPT-5.6 Sol reduced tool-call error rate from 2.7% (with GPT-5.5) to 0.2% in its internal benchmark.
- GPT-5.6 Luna delivers a 78% lower cost per document compared to GPT-5.4 mini, according to V7.
- Migration from Chat Completions API to Responses API reduced token use by roughly 5% for PDF-heavy workflows and improved caching reliability.
- On a new "very-hard" graph-query test across four difficulty levels, GPT-5.6 Sol scored 78% and GPT-6 Astra scored 89% accuracy; both scored near 100% on easy, medium, and hard levels.
- V7 Go exposes Context Graph querying and ingestion via MCP server, accessible from ChatGPT and Codex; workflow creation time dropped from ~1 hour to ~20 minutes.
- Customer outcomes cited: asset managers screen deals 21x faster (full day → 15 minutes); a financial services team cut review from 100+ hours to under 10, saving $12,000 in expert costs per task; insurance teams reduced claims processing errors by 13.5% versus manual baseline.
- OpenAI approved and implemented V7's capacity increase requests within hours, versus weeks for other providers.
Unknown
- General availability dates, pricing, or API access tiers for GPT-5.6 Luna, Terra, Sol, or GPT-6 Astra — the post does not state whether these model names reflect generally available products, preview releases, or V7-specific designations.
- Independent replication of V7's benchmark results (HERB, tool-call error rates, graph-query accuracy, cost-per-document figures).
- Rollout scope for the Responses API migration benefits across other workloads or customers.
- Whether the 99.9% workflow accuracy claim is measured against a defined ground-truth set and under what conditions.
- Technical details of the Context Graph architecture (schema, update latency, consistency model) beyond the high-level description.
Our take
The case study reads as a coordinated signal: OpenAI is showcasing a family of post-GPT-5 models through a design partner that sells agentic workflows into regulated industries. The tiered assignments — extraction to Luna, reasoning to Terra and Sol, hardest queries to Astra — look like product-line strategy rather than one monolithic model. What's unproven is whether the benchmark gains hold outside V7's own harness. For buyers, the practical lever is MCP: it turns the Context Graph into callable memory for ChatGPT and Codex.
