Estimating Fable and Sol model sizes

I'll pull the linked source so the paragraphs stay grounded in what was actually claimed about Fable and Sol sizes.Researcher @anpaure published Frontier model size estimates, an interactive chart of implied total parameters for current frontier models, and shared it on X. On Hacker News the link sat at 7 points with 0 comments. The two headline figures are Claude Fable 5 at a model estimate of 4.5T parameters and GPT-5.6 Sol at 3.1T.

The page does not treat those numbers as single opaque guesses. It combines audited capability signals from the Artificial Analysis Intelligence Index and the Epoch Capabilities Index, uses No-CoT Time Horizon as the primary parameter-scaling branch, and treats API price and compute as lower-weight checks. For Fable and Sol specifically, the public estimate is the geometric mean of parameter-count forecasts from a 20-person poll of researchers and engineers. The forest plot shows a wider weighted factor range for Fable (3.2–6.6T, public estimate 4.4T) than for Sol (2.9–3.4T, public estimate 3.2T), so Sol’s band is tighter around its point estimate.

For engineers building on Claude and Codex, parameter scale is a proxy for latency, cost, and when a smaller tier is enough. If Sol is ~3.1T and Fable ~4.5T, routing, caching, and agent budgets should not treat the two flagships as interchangeable black boxes. The poll-plus-evidence method is reproducible: you can reweight channels and see which inputs move the estimate.

On the same scale, Claude Opus 5 sits at 3.0T, disclosed Kimi K3 at 2.8T, Claude Opus 4.8 at 2.7T, GPT-5.5 at 2.3T, GPT-5.6 Terra at 2.1T, Claude Sonnet 5 at 1.6T, and both Grok 4.5 (disclosed) and GPT-5.6 Luna at 1.5T. That puts Fable above the rest of the chart and Sol roughly in the Opus 5 / Kimi K3 band, with OpenAI’s family clearly tiered (Sol > Terra > Luna).

Treat the numbers as estimates, not vendor specs. Watch the repo and chart for poll updates, factor reweights, and any official disclosures that pin Fable or Sol the way Kimi K3 and Grok 4.5 already are.