Strands agent boosts Claude Opus 5 from 30% to 99.95% on ARC-AGI-3
agents anthropic claude
| Source: Dev.to | Original article
A Strands agent boosted Claude Opus 5’s ARC‑AGI‑3 score from 30% to 99.95%, showcasing rapid AI self‑improvement.
A Strands‑built AI agent has pushed Anthropic’s Claude Opus 5 to near‑perfect scores on the ARC‑AGI‑3 benchmark, raising its accuracy from roughly 30 % to 99.95 %. The jump was achieved by wrapping Opus 5 in a lightweight orchestration layer that treats the model’s context as a “living document,” repeatedly checking assumptions against production data, correcting errors in memory and limiting unnecessary sub‑agent calls – practices highlighted in Anthropic’s own rollout notes for Opus 5.
The result matters because it shows how external agents can extract far more value from a given large language model than the model can deliver on its own. Opus 5 already boasts a new tokenizer that yields about 30 % more tokens for the same text, and Anthropic advises careful cost control at the task level. Strands’ approach demonstrates that disciplined agent design—capping sub‑agents, defining precise outputs and reserving the model for high‑impact decisions—can turn a modestly performing LLM into a virtually flawless problem‑solver on a challenging reasoning suite.
Industry observers will watch whether other firms replicate the technique and whether Anthropic integrates similar orchestration capabilities into its own offerings. The next steps include testing the Strands agent on additional benchmarks, monitoring cost‑effectiveness as Opus 5’s pricing model encourages selective use, and seeing if the approach scales to more complex, multi‑modal tasks. If the performance boost holds across domains, it could accelerate the adoption of agent‑centric AI stacks in both research and commercial settings, reshaping how organisations evaluate and deploy large language models.
Sources
Back to AIPULSEN