Video: "Creating a New Fable? Testing Mixture of Agents in Hermes Agent" by Julian Goldie on YouTube.
The gated model problem
Claude Fable 5 and Mythos 5 launched in early June 2026 and were pulled within days by a US export control order that blocked access for non-US users. GPT-5.6 is similarly in a limited access programme with no confirmed timeline for broader availability. For UK businesses that want to work with the best available models, this creates a practical gap: the most capable models are either unavailable or inaccessible at scale.
Mixture of Agents is one of the most coherent responses to this situation. Instead of waiting for access to a single better model, you combine the best models you do have access to — Opus 4.8, GPT-5.5, and others available via API — and aggregate their outputs. The question Julian Goldie's test addresses is whether that combination meaningfully closes the quality gap, or whether it just costs more to produce results that are still materially worse than a frontier model.
What the Fable test covered
The test Julian Goldie runs in this video compares MoA panel output against benchmark scores associated with Fable 5-class performance on structured content and SEO tasks. The comparison is not a direct head-to-head — Fable 5 is not accessible to run — but uses published benchmark scores and Hermes Bench results to triangulate where the MoA panel lands relative to what Fable 5 is reported to achieve.
The tasks covered are practical ones: multi-section content outlines, entity-aware meta descriptions, internal linking structures, and search intent analysis. These are the categories where frontier models show the largest advantages over older models, because they require sustained coherence across a long output and accurate recall of domain relationships. They are also the tasks UK businesses are most likely to want AI assistance with.
What the benchmark results show
The MoA panel — Opus 4.8 and GPT-5.5 in the primary configuration — produces output that Hermes Bench scores in the range of Fable 5's reported numbers on structured content tasks. It does not match Fable 5's reasoning performance on complex multi-step problems, and Julian Goldie is clear about that distinction in the video. The claim is not that MoA equals a frontier model — it is that for the specific category of structured content work that most businesses actually need AI for, the gap is narrow enough that it may not matter in practice.
The cost picture matters too. A Fable 5-level output via MoA costs more per task than a single Opus 4.8 call but substantially less than what Fable 5 access programmes charge where they are available at all. For volume content work — dozens of briefs, hundreds of meta descriptions, ongoing keyword research — the economics of MoA compare favourably even before factoring in the access problem.
What this means for UK businesses
If you are currently using Opus 4.8 or GPT-5.5 for structured content tasks, Mixture of Agents is worth evaluating. The setup cost is moderate — a Hermes configuration with the right panel — and the quality uplift on the tasks it is designed for is measurable. If you were planning to move to Fable 5 once access opens more broadly, it is worth knowing that MoA may already give you most of what you were waiting for.
If you are doing simpler, lower-volume tasks where a single model performs adequately, MoA adds cost and complexity without a commensurate benefit. The right use is in production pipelines where consistent, high-quality output on structured tasks is the priority.
Where this connects to NordSys
If tracking this kind of AI-agent news makes you wonder whether one could actually run inside your business, that's exactly what our AI Agents do — named, briefed and managed for you, no setup fee, from £6 a day.
See our AI Agents →