NVIDIA AVO Scores 100 on ARC AGI 3 After Starting as a CUDA Optimization Agent
NVIDIA has revealed new results from AVO, or Agentic Variation Operators, an autonomous agent architecture originally developed to optimize GPU kernels and sustain complex software engineering work without requiring humans to direct every individual step. The system has now achieved a 100.00 RHAE score across the complete ARC AGI 3 public set, solving all 183 levels across 25 environments while using Claude Opus 5 as its underlying model. The result is particularly notable because ARC AGI 3 does not provide traditional natural language instructions, rules or clearly stated objectives. Instead, agents must interact with unfamiliar environments, observe what their actions change, infer the objective and determine how to complete each level.
According to NVIDIA, AVO completed the 183 levels in 6,624 environment actions. NVIDIA compared this with VISTA, which reported 7,542 actions using Claude Opus 5 to complete the same public levels, giving AVO an approximately 12% reduction in actions. NVIDIA is careful to note that this is not a controlled comparison because the 2 systems use different memory structures, agent backends, observation methods and context management strategies. The company also stresses that the 100 score applies only to the ARC AGI 3 public set and does not represent a result on the semi private or fully private competition sets.
Our general-purpose coding agent just scored 100% on the ARC-AGI-3 interactive reasoning benchmark.
— NVIDIA AI (@NVIDIAAI) August 21, 2026
NVIDIA AVO completed all 183 levels across all 25 public environments, figuring out what to do with no instructions, explicit rules, or stated goals. pic.twitter.com/UgROuDrMtn
What makes AVO especially relevant to NVIDIA is that it did not begin as an ARC benchmark project. The architecture was initially developed for autonomous software engineering and GPU kernel optimization, where an agent needs to inspect existing code, develop optimization hypotheses, modify implementations, compile them, run hardware tests, evaluate performance and continue iterating when an approach fails. NVIDIA says AVO operated continuously for 7 days during an attention kernel optimization experiment, exploring more than 500 optimization directions and committing 40 kernel versions. On NVIDIA DGX B200 hardware, the resulting multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention 4 by up to 10.5% across the configurations NVIDIA tested. AVO then adapted the evolved kernel to grouped query attention in approximately 30 minutes of additional autonomous work.
The technical significance is not simply that an AI system can generate CUDA code. High performance GPU kernels are extremely sensitive to memory access, scheduling, synchronization, register pressure, tensor operations and the characteristics of the specific architecture running them. A change that appears efficient from source code alone can easily perform worse once executed on real hardware. AVO addresses this through an iterative feedback loop where generated changes are repeatedly tested against actual performance rather than accepted because the code merely compiles. That approach aligns closely with NVIDIA's wider focus on software optimization around CUDA, which continues to be one of the largest competitive advantages surrounding its GPU ecosystem, NVIDIA Blackwell running DeepSeek V4 at nearly 3,500 tokens per second similarly showed how optimized CUDA kernels and software tuning can produce major performance gains even after the underlying GPU hardware is already deployed.
AVO also highlights an increasingly important distinction between an AI model and an AI agent. NVIDIA notes that ARC Prize separately reports an approximately 30% result for Claude Opus 5 under a different evaluation configuration, while the complete AVO system reached 100 on NVIDIA's public set run. NVIDIA explicitly warns that the 2 numbers should not be interpreted as a controlled measurement of how much performance AVO alone added. The broader point is that persistent memory, supervision, tools, context management and recovery mechanisms can substantially change what the same underlying model family is capable of completing over extended workloads.
That shift is becoming increasingly important as coding agents move beyond short code generation tasks toward workflows that may involve hundreds of tool calls, experiments and revisions. NVIDIA is already optimizing its hardware around these workloads, with the GB300 NVL72 delivering up to 20 times higher concurrent agent throughput per megawatt than H200 in NVIDIA's published AA AgentPerf results. AVO represents the software side of the same strategy, where increasingly capable agent systems can potentially optimize the code that eventually runs on NVIDIA hardware.
AVO is more interesting as a GPU engineering tool than as another 100% AI benchmark headline. NVIDIA already has one of the industry's strongest advantages through CUDA, but autonomous optimization could multiply that advantage by allowing agents to explore hundreds of kernel variations, test them directly on hardware and preserve useful discoveries without engineers manually driving every experiment. The ARC AGI 3 result suggests that the architecture behind that optimization loop can also transfer into very different tasks where rules must be discovered through interaction. The important limitation is equally clear. 100% was achieved on the public set, not the private competition benchmark, so it should not be interpreted as evidence that ARC AGI 3 has been fully solved. What NVIDIA has demonstrated is something more immediately practical: agent architecture can dramatically influence how much useful work a frontier model can sustain.
Could autonomous agents eventually optimize CUDA kernels better than human engineers for most workloads, or will expert developers remain essential for finding the biggest performance breakthroughs?
