Back to Feed
What Is Fable 5?
The Signal
Access to Fable 5, a guarded version of the Mythos 5 artificial intelligence model, reveals that public benchmark scores are significantly depressed by automated refusals. The central tension lies in whether the guarded Fable 5 reflects the true potential of the underlying model, with the narrator arguing that once unrestricted, the model demonstrates performance competitive with or superior to OpenAI and Anthropic’s Opus.*
The Case
- Fable 5 benchmarks are suppressed by safety refusals, causing a 20-point drop on Terminal Bench compared to Mythos 5 runs.
- The narrator claims Fable 5 is highly cost-efficient, reportedly outperforming Anthropic’s Opus models while using fewer tokens and dollars.
- Performance on the coding benchmark SWEBench Pro is cited at 80%, though the speaker critiques the metric itself as 'garbage' because models may merely reconstruct known pull requests.
- Frontier Codebench scores — 30% for Fable 5 versus 13% for Opus 48 — are presented as more reliable evidence of coding superiority.
- The model is described as having achieved a 'meaningful upgrade' in vision and spatial reasoning, potentially surpassing OpenAI’s capabilities, though this remains an unsupported assertion without visual data.
The 1 Minute Signal Take
The evidence suggests that raw model capability is currently obscured by safety constraints and high-variance benchmark quality, making the 'leader' narrative difficult to verify. Skip this video unless you need a granular breakdown of the speaker’s personal, non-audited testing methodology for code-heavy tasks.
Time saved:
Tags
Back to Feed