Hermes Fable 5 + Which AI Models To Pair With It: Real GoldieBench Scores For Fable, Opus And Sonnet, And What They Mean For A Hermes Setup

Hermes Fable 5 setups work best as a pair: Fable on the main Hermes chat, and a lighter Claude model on subagents and cron jobs.

GoldieBench gives you numbers to choose that partner with.

Claude Fable 5 scores 8.10 out of 10 on the bench, Claude Opus 5 scores 8.27 and Claude Sonnet 5 scores 7.01.

Below are the scores I read on the live pages on 4 November 2026, what they mean for a Hermes agent, and the pairing I'd run.

If you want my team to sanity-check your model choice, book a free SEO strategy session.

The Two Ways Hermes Reaches Fable

Before the scores, a quick word on wiring.

The official experimental plugin claude-subscription-directsdk runs Claude Code as a Hermes model provider on your Claude plan.

Set default: fable and Hermes runs the pinned route claude-fable-5-1[1m], with a 1,000,000-token window.

Plan-wise, commit fdbf4225d2 treats Fable as part of Max and Ultra, while Pro accounts and standard Team or Enterprise seats pay for it with usage credits.

The other way in is Hermes's built-in Anthropic provider, with --model claude-fable-5-1 and an API key, documented in the Hermes providers guide.

The GoldieBench Numbers

Every score is an average of judged one-shot builds from an identical prompt.

The Fable page puts Fable 5 at rank 6 of 26 ranked frontier models, with 4 golds, 2 silvers and 1 bronze, and data refreshed on 27 October 2026.

How To Read A GoldieBench Score

The methodology matters as much as the number.

Every GoldieBench score is a one-shot, single-file build from the same prompt for every model.

There are no retries and no hand-fixing.

Each build is rendered for real, screenshotted and scored from 0 to 10 by one judge model on one rubric.

Failures score as failures.

So a high average means a model gets complex builds right first time, which is a useful signal for a planner that has to lay out a job in one pass.

What 8.10 Means For A Hermes Main Chat

Fable 5 is strongest where the build is complex.

On the Fable page it averages 8.2 on games against a field average of 7.1, and 8.1 on sims against 6.9.

Its best single builds score 9.0, on the Skyrim and Voxelcraft tasks.

That's a model that holds a large, multi-part brief together, which is what a Hermes planner has to do.

Pages are its weak spot: 7.6 against a field average of 7.7, across just 3 tasks.

So for a main chat that writes plans, specs and long documents, Fable earns its seat.

For a quick landing-page draft, the bench gives it no edge.

What The Bench Doesn't Tell You

Read this before you lean on any score above.

The Fable page says GoldieBench does not measure multi-turn agent work, long-context recall or API latency.

A Hermes agent does all three.

So the bench tells you which model builds best in one shot, not which one runs the best four-hour session.

For that, Anthropic's own guidance is the better source.

Its models overview says to start with Opus 5.5 for most workloads and to use Fable 5.1 for demanding reasoning and long-horizon agentic work.

The Fable 5 Score vs The Fable 5.1 Model

Every GoldieBench Fable score is for Fable 5.

There's no Fable 5.1 page on the site yet.

Anthropic released Fable 5.1 on 1 September 2026 and now lists Fable 5 as legacy.

Its Fable 5.1 page shows a 1M-token window, 128K max output and thinking that's always on.

The Fable 5 bench page lists the context as 200,000 tokens, or 1M with extended thinking, which differs from Anthropic's figure, so I go with Anthropic's 1M.

The plugin's fable short name points at Fable 5.1, so the model you run is newer than the one that was scored.

Pairing 1: Fable Main Chat, Sonnet Subagents

This is the default pairing I'd run.

Fable plans the work and checks the output.

Sonnet 5, at 7.01, does the batch tasks: summaries, rewrites, checklists.

The Hermes delegation docs say children inherit the parent model unless delegation.model is set, so set it.

model:
  provider: claude-subscription-directsdk-experimental
  default: fable
delegation:
  provider: claude-subscription-directsdk-experimental
  model: sonnet

A 1.09-point gap between Fable 5 and Sonnet 5 matters on a hard build, much less on a summary that Fable reviews anyway.

Pairing 2: Fable Main Chat, Lighter Cron Jobs

Scheduled jobs are the easiest place to waste Fable.

The Hermes cron docs say an unpinned job runs on cron.model, and on your main model if that isn't set.

hermes config set cron.model_provider claude-subscription-directsdk-experimental
hermes config set cron.model haiku
hermes cron edit <job_id> --model sonnet

Haiku has no GoldieBench score, so I can't rank it here.

For a daily inbox sort or reminder, a score doesn't decide it anyway: the job is simple and repeats.

Pin the report-writing jobs to Sonnet and leave the rest on the fleet default.

Where Opus 5 Fits

Opus 5 is the one number above Fable 5 in this table: 8.27 against 8.10.

The Fable page shows Fable 5 trailing Opus 5 head-to-head, 12 wins to 29 with 6 ties.

It also shows Fable 5 leading Claude Opus 5.5 by 27 wins to 18.

So if your Hermes work is mostly one-shot builds, Opus 5 is a fair main-chat alternative on this bench.

If it's long agent sessions, Anthropic's guidance still points at Fable 5.1.

Two Small-Business Pairings

A web design studio building client prototypes.

1. The designer opens a Fable chat with the client brief and brand notes.
2. Fable writes the build plan and splits it into sections.
3. Sonnet subagents draft each section to the plan.
4. Fable reviews the assembled prototype and lists fixes.

A dental practice's patient reminders.

1. A Haiku cron job reads tomorrow's appointment list each evening.
2. It drafts reminder messages in the practice's tone.
3. It queues them for the receptionist to approve.
4. Fable is only opened when the practice manager needs a long policy document rewritten.

The Pick

Run Fable on the Hermes main chat.

Put Sonnet 5 on subagents, and Haiku or Sonnet on cron, pinned per job.

Treat the Fable 5 score as a strong one-shot signal, not proof of long-session skill, and remember the plugin is experimental.

For the wider shortlist, see the best Hermes Agent model guide, and the full scoring rules sit on the methodology page.

Run a test before you commit: give the Fable main chat one real planning job and watch which pieces it delegates.

If the subagent output needs heavy fixing, move that route from Haiku to Sonnet.

Want the line-up ready to install? Join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe exact commands and config for both Hermes-on-Fable setup routes
• AI Income Deskwhich Claude plan covers Fable in Hermes, and four small-firm use cases
• AI Tool Verdicta scorecard verdict on both ways to run Fable inside Hermes
• agentos.guidefive prompts that make your agent switch Hermes onto Claude Fable
• aiprofitboardroom.comthe complete Hermes Fable 5 guide on the AI Profit Boardroom blog

Set Fable as the main model, set the delegation and cron models today, and if you want a second opinion book a free strategy session before you lock in your Hermes Fable 5 pairing.