GPT-5.6 vs. Claude Fable 5 for Physical AI, which performs best?

I'll pull the JuliaHub evaluation article so the paragraphs stick to real names, numbers, and claims rather than invented detail.JuliaHub published a sealed July 2026 evaluation, led by Dr. Chris Rackauckas, Anas Abdelrehim, and Anshul Singhvi, that ranks Anthropic’s claude-fable-5 against OpenAI’s GPT-5.6 family (terra, sol, and luna) inside the Dyad physical-modeling agent. Across 52 graded runs on five modeling-and-simulation problems, claude-fable-5 led with a difficulty-weighted score of 0.889 at $9.60 and 16.1 minutes per trial. gpt-5.6-sol placed second at 0.814 for $1.74 and 13.4 minutes; gpt-5.6-terra scored 0.786 at $1.25 and 12.6 minutes; gpt-5.6-luna trailed at 0.727 for $3.26 and 25.0 minutes. Fable was the only model to clean-sweep all twelve trials on the four core problems, while the full study bill for Fable ran about $125—three to eight times the GPT tier.

The harness stayed fixed so only the model varied: Dyad AI agent, reasoning effort at xhigh, 1M context, and a 128k token budget. Each sealed container packaged a problem, pinned configuration, and hidden ground truth; the agent had to derive physics, write a Dyad model, compile, simulate, and verify before a grader scored simulated trajectories—not source that merely compiles. Problems climbed from constitutive and constrained consistency through relativistic dynamics and steady-state linearization to NASA’s HL-20 lifting body: six-degree-of-freedom rigid-body dynamics, roughly 170 wind-tunnel aero tables, standard atmosphere, and a control-surface mixer from two NASA technical memos, graded on eight physics scenarios. No model fully solved HL-20. Fable led that frontier problem at 0.69 (full-flight tracking within about 0.4 m of the sealed reference), with sol at 0.66, terra at 0.59, and luna at 0.38; every model failed dutch-roll lateral-directional response.

For builders shipping agentic simulation, the failure mode that matters is not broken syntax but physics that looks green and is still wrong. Agents can write their own tests, pass them, and still encode impossible aircraft, column, or particle dynamics. JuliaHub’s work-style fingerprints underline the behavioral split: Fable averaged about 28 tool calls per trial, spent a large share of time deriving and testing, and in one relativistic trial swept solver tolerances across three decades and deliberately sign-flipped a field (residual 2.0 wrong vs 0.066 right) before committing. Sol built strong independent oracles—one constitutive check closed to 3.7e-12—but also misread a pinned four-velocity z-component of -0.62 as coordinate velocity and handed the particle about 27% excess longitudinal momentum while its self-checks stayed consistent. Terra was cheapest and fastest but shortened a required stop from 5.0 to 3.0 and blinded its own nonlinear reference with the same horizon. Luna piled on self-referential validation (68 calls and still ~30% off on one constrained-consistency trial; 544 calls and 138 minutes on HL-20) without external anchors.

On competitive terms, the headline is not a blanket model crown so much as a cost–capability trade inside a specialized engineering stack. Fable buys the highest probability that the committed model is physically right; Sol is the value default at roughly a fifth of Fable’s per-trial price and second on merit; Terra is the budget/speed route; Luna lost on score, cost, and wall-clock. Provider self-benchmarks are treated as untrustworthy by design: JuliaHub ships Dyad with multi-vendor backends and says it wins only when users get the best modeling experience, so the evaluation’s incentives cut against cheerleading a single lab. On the unsolved HL-20, Fable’s edge concentrated on tests with external references (especially the high-weighted full-flight scenario), while Sol extracted both NASA memos early and locked aero against 24 embedded check cases yet still ended a dutch-roll run below sea level in a 48-degree dive.

A companion harness comparison is the larger practical lever. Holding the frontier model fixed and swapping only the agent stack, the same model scored 0.899 inside Dyad AI versus 0.533 in a stock Claude Code-style coding agent at nearly identical cost (about $4.20 vs $4.65), a 0.366-point gap—more than double the 0.162 separating best and worst models under Dyad alone. The stock agent scored zero on every relativistic-dynamics trial it ran. If you are choosing tooling for physical AI, pick the instrument that forces derive–compile–simulate–verify against sealed trajectories first, then buy the best model you can afford inside it: Fable when correctness risk dominates, Sol when unit economics dominate day-to-day modeling. Watch for JuliaHub’s full harness-versus-harness write-up and for whether HL-20-class long-horizon flight problems remain unsolved as new frontier releases land.