GPT-6.2 Astra is the strongest robot-code author we’ve tested to date.
With no simulator, demos or web access, it wrote a 2.5B parameter neural net that solved 79% of 500 LIBERO trials, and Python code that solved 96.8%, beating Fable 5.2 (89.2%) and GPT-6 Astra (33.2%).
Astra 6.2 also led in our live control tests, solving 60% of trials vs Fable’s 50% at 7x less per success (75 cents vs $5.44).
But for every LLM we tested, writing a policy was 27-77x cheaper per success than live control.
Full results: withplayground.ai/research/h…
3
46
Full leaderboard and methodology: withplayground.ai/blog/robot…
4
118
SLA-Preview is the new #2 model on the Robot Intelligence Index, beating Dream-1 and every frontier LLM except Fable 5.2.
At just 3.35B parameters, it’s the top open-weight model, the only VLA in the top 10, and the only university-led model in the top 5.
With no extra parameters, SLA-Preview nearly doubles π0.5’s score on RoboDojo Real (42.3 vs 22.9) by training its vision to match a 3D vision model. Congrats to the SLA team at HKUST (Guangzhou) on the result!
Full leaderboard: withplayground.ai/blog/robot…
1
4
247
Full SLA (Spatial-Language-Action) paper: arxiv.org/abs/2611.08373
4
159
Frontier LLMs write robot policies that are 30-75x more cost-effective and 17-76x faster than the same LLM controlling a robot live.
Across 500 LIBERO trials, Fable 5.2 solved 89% of episodes with code it wrote and 93% with a hand-written neural net, but only 50% driving live.
Our results suggest existing robotics evals miss some of LLMs’ most useful skills. Most benchmarks score LLMs as live controllers, but in LIBERO that was the slowest, costliest way to use them, and for Fable, the least successful.
Full blog and replication bundle: withplayground.ai/blog/how-s…
We used Fable 5.2 to hand-write a sparse 3.8B parameter model that scored 82.8% (414/500 simulated trials) on 10 randomly selected LIBERO tasks.
Fable built it in 90 minutes with no simulator, demos, web access or gradient descent.
Fable raised the network's score to 92.8% with a single revision, landing within 5 points of fine-tuned π0.5’s published score on the same LIBERO suites.
Fable’s network cost just 7 cents per success over 500 episodes, and was small enough to run on a CPU.
Full blog: withplayground.ai/blog/how-s…
1
9
790
Full leaderboard and methodology: withplayground.ai/blog/robot…
2
441
Playground retweeted
LLMs scoring high on sim benchmarks is great, but they usually get privileged info not available in the real-world. The comparison I am excited to see are frontier LLMs vs. frontier WAMs on real-world benchmarks, run by third parties. The optimal solution is probably a mix of both. Either way, rising zero-shot generalization numbers move the whole field forward.
Yesterday, Fable 5.2 became the first frontier model release to include a robotics benchmark result (64.9% on RoboDojo Sim). The problem is that simulation benchmarks systematically underrate certain model architectures.
Our findings validate sim benchmarks as a tool for relative placements, but only in tandem with real-robot benchmarks. To help close the real-robot gap, we're:
• executing real-world runs of unseen LLM vs. VLA vs. WAM matchups on our hardware
• updating our methodology to weight real-robot results more heavily in our overall robot intelligence ranking
• publishing sim-to-real agreement by model family with every update to our intelligence index
Evaluate your model on our hardware: withplayground.ai/blog/robot…
1
2
217
We’re very encouraged by Anthropic’s decision to report robotics eval results alongside its Fable release. We believe independent, real-robot evals will be vital as frontier labs begin to compete in robotics.
Congrats to the @AnthropicAI team on the release!
2
166








