Claude Opus 5 Gets Tested in Tornadoes, Collapsing Buildings, and Sand Simulations

Claude Opus 5 was just released, and Anthropic positions it as a cutting‑edge model comparable to Fable 5 but at roughly half the price, charging $5 per million input tokens, $25 per million output tokens, and offering a Fast mode about 2.5× faster.

Community members quickly began stress‑testing the model by asking Opus 5, Fable 5, Kimi K3 and GPT 5.6 to each generate three self‑contained HTML physics scenarios: a tornado that sweeps an entire house away, a demolition ball that punches through an apartment, and an overloaded truck that collapses a truss bridge.

Cost and outcome details show Opus 5 consumed 55.9 K tokens, costing $1.40, and successfully made the house fly from the funnel top, shattered walls at impact points, and caused the truck to fall into a river. Fable 5 cost $2.82 (about double) but the building collapsed before the ball hit it. GPT 5.6 cost only $0.31, yet the ball never reached the building and the bridge broke like a chopstick. Kimi K3 was cheaper still but also failed to avoid the physics teacher’s scrutiny.

Another visual demo featuring cherry blossoms in the wind highlighted that Fable 5 remains the strongest, Opus 5 is decent, and Kimi K3 is the cheapest. A sensational post titled “OPUS 5 MURDERED GPT 5.6 AND FABLE 5” showcased the results in a video with a so‑called “GOD‑MODE” guide.

A more advanced group created a high‑level sand simulation that also handled water, fire, smoke, lava, and acid with density‑based layering and reactions, processing tens of thousands of particles at 60 FPS. Their ranking list placed models from Kimi K3 Max and Qwen 3.8 Xhigh up to GPT 5.6 Sol Ultra and Opus 5 Max.

The author notes that what used to be a weekend project for front‑end developers has become a zero‑shot, single‑file AI arena. Early testers say Opus 5 writes more clearly than Fable 5, excels at browser automation and complex data review, but can be verbose and behaves like a well‑trained student. Its cost advantage over GPT 5.6 Sol is not guaranteed, and some tests show Opus 5 costing $4.20 versus Fable 5’s $9.60 for comparable tasks.

The real excitement, the author argues, is not a leaderboard shuffle but that on the first day after release users are no longer asking whether the model can write code; they are asking whether it can understand how a world collapses, flows, burns, and then recreate that world themselves.

Key prompt guidelines:

State the full task up front: give complete specifications; Opus 5 is better at autonomously completing complex long tasks.

Be concise and explicit about requirements; use effort to control depth of thinking without limiting answer length.

Set reporting cadence: a single opening line, then only update when there is a key finding or direction change.

Avoid redundant verification: the model self‑checks; repeated validation only adds cost.

Lock task boundaries: narrow tasks must have clearly defined scope to prevent the model from expanding the work.

Use sub‑agents cautiously: delegate only large, independent work to sub‑agents; handle small tasks directly.

Official prompt guide and announcement links:

https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5
https://www.anthropic.com/news/claude-opus-5