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

I'll pull the source article for concrete names, numbers, and product details so the paragraphs stay factual.JuliaHub researchers Dr. Chris Rackauckas, Anas Abdelrehim, and Anshul Singhvi ran a sealed July 2026 evaluation of frontier models on physical modeling and simulation inside the Dyad agent. They compared Anthropic’s **claude-fable-5** with OpenAI’s GPT-5.6 family (**sol**, **terra**, **luna**) across five problems of rising difficulty, for 52 graded runs total. On difficulty-weighted score, **claude-fable-5** led at **0.889** ($9.60 per trial, 16.1 minutes), followed by **gpt-5.6-sol** at **0.814** ($1.74, 13.4 min), **gpt-5.6-terra** at **0.786** ($1.25, 12.6 min), and **gpt-5.6-luna** at **0.727** ($3.26, 25.0 min). Fable was the only model to sweep all twelve trials on the four core problems; no model fully solved the hardest task, NASA’s HL-20 flight vehicle.

The protocol held the harness fixed and varied only the model: same Dyad agent stack, reasoning effort (**xhigh**), 1M context, 128k token budget, and sealed problems. Each run put the agent through a container that combines model, harness, pinned config, and a sealed problem; the agent must derive, write Dyad code, compile, simulate, and verify before a grader scores the full simulated trajectory against ground truth rather than whether the code merely compiles. Problems range from constitutive and constrained consistency through relativistic dynamics and steady-state linearization to a long-horizon HL-20 case with six-degree-of-freedom dynamics, roughly 170 wind-tunnel tables, atmosphere, and control-surface mixing from two NASA memos. On HL-20, Fable scored 0.69, Sol 0.66, Terra 0.59, and Luna 0.38; everyone passed gravity, nobody reproduced dutch-roll lateral-directional response, and Fable’s full-flight track stayed within about a third of a meter of the sealed reference while Sol ended about nineteen meters low.

For engineers building agentic simulation or CAD-to-physics workflows, the failure mode that matters is not syntax. An aircraft, separation column, or charged-particle model can compile and run green while encoding impossible physics, and agents that write their own tests can bless those mistakes. Transcript fingerprints show distinct temperaments: Fable averaged about 28 tool calls per trial and verifies by trying to break its own work; Sol builds strong independent oracles but can over-specify the task and misread pinned state (for example renormalizing a four-velocity so a fixed z-component became coordinate velocity instead); Luna churns (68 calls on one constrained-consistency trial, still about 30% off) and often validates only against its own equations; Terra is cheapest and fastest but trims margins—on steady-state linearization it shortened `stop = 5.0` to `3.0` and blinded its own nonlinear reference with the same truncated horizon. Capability for physical AI is less “does the agent ship code” and more “does the committed trajectory match reality.”

On pure accuracy, Fable owns the top of the table and the cleanest core-problem record, at three to eight times GPT trial cost (about $125 for the full Fable study run). OpenAI’s GPT-5.6 stack owns the value frontier: Sol is second on merit at roughly a fifth of Fable’s per-trial price, Terra is the cheapest and fastest credible path, and Luna trails on score, cost, and wall-clock. That ranking is inside one instrument. In a companion harness comparison with the model pinned, the same frontier model scored **0.899** in the Dyad AI harness versus **0.533** in a stock general-purpose coding agent (Claude Code with the same docs and a Dyad compile tool) at nearly identical spend—a **0.366**-point gap, more than double the **0.162** separating best and worst models in the model-only study. JuliaHub’s incentive is multi-vendor (Dyad ships several backends); the write-up is not a vendor self-score on a public leaderboard.

If “best” means highest probability the committed model is physically right, buy **claude-fable-5** inside a domain harness that forces derivation, compile, simulate, and external grading. If “best” means physics per dollar for day-to-day modeling, **gpt-5.6-sol** is the rational default, with **terra** when cost and latency dominate. Watch next for JuliaHub’s full harness-versus-harness write-up, Dyad 3.0’s modular skills path for custom engineering workflows, and whether other labs publish sealed trajectory-match evals—not chat benchmarks—for modeling and simulation. The model ranking is real; the larger lever is still the agent stack that refuses to grade “runs green” as “physics correct.”