Which LLM Builds the Best Knowledge Graph? Opus 4.8 vs Fable 5
By Dimitris Goudis | 2026-06-11 | Updated | 8 min
Which AI Model Builds the Best Knowledge Graph? We Tested Two on the Same 134 Videos
You point an AI at your transcripts, your client calls, your documents, and you ask it to build you a knowledge graph — a digital brain. It hands back thousands of nodes and edges. Looks impressive.
Here's the question almost nobody checks: did it actually understand any of it, or did it just shred your content into confetti and call it a graph?
I checked. I took the same 134 YouTube videos, fed them through the exact same pipeline twice — once with Claude Opus 4.8, once with Fable 5 — and built two separate knowledge graphs. Same videos. Same ontology. Same loader. The only variable was the model doing the thinking.
Then I did the part people skip: I verified the output against the original transcripts, line by line, instead of trusting the node count.
This is not an abstract lab exercise. It's exactly what Brain Graph does — turn client calls, documents, and frameworks into an AI knowledge graph you can actually query. So the question "which model should build your brain?" is the whole game. Here's the answer, with receipts.
The setup: one corpus, two models, zero excuses
- Corpus: 134 videos (creators teaching AI agents, Claude Code, automations, second brains).
- Run 1 — Opus 4.8 → graph
social_media. - Run 2 — Fable 5 → graph
social_media_fable_5. - Same extraction schema, same dedup rules, same FalkorDB loader, same audit gate.
Because the input is identical, every difference in the two graphs is the model, not the data. That's what makes this a fair fight.
Then two kinds of measurement:
- Whole-graph metrics — computed deterministically over every node and edge. No sampling.
- Trust metrics — I sampled 578 items (entities, claims, judgment-call edges, classifications) and had independent checker agents verify each one against the actual transcript, defaulting to the skeptical answer when they couldn't find support. 260 grounding checks, 160 inference-edge checks, 146 blind classification calls, 78 gold facts for recall.
Most "AI builds a knowledge graph" demos stop at "look how many nodes." This one starts where that one stops.
The trust number first: is what it extracted actually true?
This is the only number that matters before any other. If the graph is full of confident nonsense, more of it is worse, not better.
Grounding rate = of a verified sample, how much the transcript genuinely supports.
| Metric (verified vs transcript) | Opus 4.8 | Fable 5 |
|---|---|---|
| Grounding rate (higher = better) | 93.8% | 97.7% |
| Hallucination rate (lower = better) | 5.4% | 1.5% |
Fable 5 invented facts about a third as often. Out of 130 sampled extractions, Opus had 7 that the transcript didn't back up; Fable 5 had 2. Neither is reckless — but if this graph is your source of truth in a client call, 1.5% beats 5.4%.
Coverage: did it actually capture what was said?
Grounding is precision. The opposite question is recall — of the real facts sitting in the transcript, how many did each graph actually grab? I pulled 78 concrete gold facts (a named tool, a specific claim, an offer) straight from the videos and checked.
| Opus 4.8 | Fable 5 | |
|---|---|---|
| Gold-fact recall | 85.9% (67/78) | 89.7% (70/78) |
Fable 5 again, by a nose. It both invents less and misses less. That combination is rare — usually a model that grabs more also grabs more junk. Fable 5 didn't.
Is it a graph, or a filing cabinet?
This is the finding that actually changed my mind, and it's the one that matters most for a digital brain.
A knowledge graph is only useful if the same idea from ten different sources becomes one connected node with ten links — not ten disconnected copies. That's the difference between "my brain knows X, and here's everywhere it came up" and a pile of orphaned notes.
| Connectivity | Opus 4.8 | Fable 5 |
|---|---|---|
| Cross-video concept rate (concepts appearing in ≥2 videos) | 6.9% | 13.1% |
| Single-video orphan concepts | 93.1% | 86.9% |
| Total edges | 16,810 | 19,551 |
| Edges per video | 125 | 146 |
Read that top row again. With Opus, 93% of concepts showed up in exactly one video and never connected to anything else. It built a very tidy filing cabinet. Fable 5 recognized the same concept across creators nearly twice as often — it built a graph that actually links.
For a tool whose entire value is connecting your scattered knowledge, that's not a tie-breaker. That's the product thesis.
Where Opus 4.8 wins
It's not a clean sweep. When I forced a head-to-head on the cases where the two models disagreed — same video, different label for the hook, the title formula, the thumbnail type, the concept type — and adjudicated each against the transcript, Opus was right more often.
| Opus 4.8 | Fable 5 | |
|---|---|---|
| Classification accuracy (on 121 decisive disagreements) | 57.9% | 42.1% |
| Inference-edge precision (judgment-call edges) | 90.0% | 88.8% |
So Opus is the sharper classifier. When the job is sorting things into a fixed, controlled vocabulary — and being right about which bucket — it edged Fable 5 out. Its interpretive edges (this video addresses that use-case, targets that persona) were also marginally more defensible.
What was clean on both
Worth saying plainly, because it's the floor you should demand from any model building your brain:
- 0 type collisions on either graph (no concept filed under two contradictory types).
- 100% of entity nodes carried a description.
- 100% provenance — every node traces back to its source video.
Both models cleared the audit gate. The differences above are about quality on top of a clean base, not one model being broken.
The decision: which model should build your knowledge graph?
| What it tells you | Opus 4.8 | Fable 5 | Winner |
|---|---|---|---|
| Trust / precision (grounding) | 93.8% | 97.7% | Fable 5 |
| Risk of invented facts (hallucination) | 5.4% | 1.5% | Fable 5 |
| Coverage (gold-fact recall) | 85.9% | 89.7% | Fable 5 |
| Connects ideas across sources | 6.9% | 13.1% | Fable 5 |
| Raw richness (edges/video) | 125 | 146 | Fable 5 |
| Right when the two disagree (classification) | 57.9% | 42.1% | Opus 4.8 |
| Judgment-edge quality (inference precision) | 90.0% | 88.8% | Opus 4.8 |
Bottom line: For building a knowledge graph you'll actually rely on — more real facts, fewer invented ones, and a graph that connects instead of fragments — Fable 5 is the pick. Choose Opus 4.8 when fine-grained classification into a fixed vocabulary is the core deliverable and you'll post-process the connectivity yourself.
For a digital brain, connectivity and trust win. That's Fable 5.
What this means for your digital brain
The model that builds your brain decides whether it becomes a thinking partner or a search box with extra steps. A graph where nothing connects can't answer "everywhere this idea shows up in my work" — and that question is the entire reason you'd build one.
This is the engine under Brain Graph: your client calls, documents, and SKILL.md files that package your expertise go in; a connected, queryable AI knowledge graph comes out — running on FalkorDB, extracted with Claude. The same memory and guardrails discipline that keeps an AI agent on-task is what keeps a knowledge graph honest: provenance on every node, a description on every concept, an audit gate before anything ships.
Want to see a brain built from your own content, live? Join the next masterclass →
Method and honest caveats
I'd rather under-claim than oversell, so:
- Whole-graph metrics (connectivity, yield, type purity) are computed over every node and edge — no sampling.
- Trust metrics are sampled and verified against the original transcripts by independent checkers (578 adjudications total), defaulting to skeptical.
- Three things I did not measure, and won't guess: repeat-run stability (needs a fresh re-run), audit first-pass rate, and cost/time per run — those live in run logs I didn't capture. Method's defined; numbers are intentionally omitted.
- Both graphs cover the same 134 videos, so every difference is the model.
That's the standard I'd want from anyone telling me their AI "understands" my content. Ask for the grounding rate. If they can't give you one, they didn't check.
FAQ
Which AI model builds the most accurate knowledge graph? In this benchmark, Fable 5 — it had a 97.7% grounding rate versus 93.8% for Claude Opus 4.8, and a 1.5% hallucination rate versus 5.4%. Opus 4.8 was the stronger fine-grained classifier (57.9% accuracy on disagreements).
What is knowledge graph hallucination? It's when a model adds a node or edge to the graph that the source material doesn't actually support — an invented fact stated with confidence. We measured it by sampling extractions and checking each against the transcript. Lower is better; Fable 5 hit 1.5%.
Does a bigger knowledge graph mean a better one? No. More nodes only count if they're true and connected. The Opus 4.8 graph had thousands of concepts, but 93% appeared in just one source and connected to nothing else — a filing cabinet, not a graph. Always read size next to grounding rate and cross-source connectivity.
How do you measure knowledge graph quality? Four numbers: grounding rate (precision), recall (coverage), cross-source connectivity (is it actually a graph), and type purity (are concepts cleanly classified). Raw node and edge counts on their own tell you almost nothing.
How does this relate to Brain Graph? Brain Graph turns your transcripts, documents, and frameworks into exactly this kind of AI knowledge graph — your digital brain. This benchmark is how we decide which model does the extraction so your brain connects ideas and stays grounded in what you actually said.
Should I replace Opus 4.8 with Fable 5? For building a knowledge graph, Fable 5 is the stronger default on this corpus — higher grounding (97.7% vs 93.8%), lower hallucination (1.5% vs 5.4%), better recall, and nearly double the cross-source connectivity. Keep Opus 4.8 where fine-grained classification into a fixed vocabulary is the main job — it won the head-to-head disagreements 57.9% to 42.1%. But don't swap on my numbers alone: I tested 134 videos about AI agents, and your content isn't mine. Run both models on a slice of your own material — same inputs, sample 20–30 outputs, check each against the source, and score grounding, recall, and connectivity. An afternoon on your own corpus beats any benchmark, including this one. Don't trust what other people say works — verify it on your own work.
Built from a transcript-verified benchmark of two knowledge graphs over an identical 134-video corpus.