Meta AI's EvoHarness-RL framework, which trains agents to manage their execution environments through a unified Belief, Progress, Experience interface, achieved a score of 96.9 per cent on the ALFWorld agentic benchmark using a Qwen3 8B parameter model — edging past Claude Opus 4.5's 96.4 per cent and representing a forty-nine-point improvement over the base model's unassisted performance. A "harness annealing" mechanism reduces latency and compute costs as agents internalise routine execution patterns over time. The framework demonstrates that frontier-level agentic performance on structured task environments is achievable with open-weight models at substantially lower inference costs than frontier model APIs — a finding with direct implications for the enterprise AI pricing dynamics that have allowed the major laboratories to command significant premiums on agentic workloads.
LLMs
Google Releases Gemini 4 Argon — Tops Frontier Benchmarks and Targets Autonomous Cybersecurity
30 September: Google's Gemini 4 Argon leads the Vals AI benchmark index over OpenAI GPT-6 Astra and …