AI Agent Armies Rise: From Benchmarks to Code, a New Era Dawns

Anthropic's new flagship model has arrived: Claude Opus 4.6 is here, and it's gunning for the throne. Early benchmarks show it outperforming GPT-4 and Gemini 1.0 Ultra on key metrics like MMLU and HumanEval, positioning it at the leading edge of general intelligence, and is available via the Claude API in 159 countries. But it's not just about excelling on standard tests; Opus 4.6 is designed for "agentic work," tackling multi-step tasks requiring planning and revision, which is why Anthropic believes this model moves AI closer to a "vibe working" era according to *CNBC*. Agentic AI is no longer a solo act. Perplexity AI launched its Agent Council, allowing users to create and collaborate with multiple AI agents to tackle complex tasks, and synthesizing their outputs for enhanced problem-solving. This mirrors a broader industry trend, as Anthropic and OpenAI are also shipping products built around managing teams of AI agents that divide up work and run in parallel. However, the efficacy of this supervisory model remains an open question, as current AI agents still require heavy human intervention to catch errors, and no independent evaluation has confirmed that multi-agent tools reliably outperform a single developer working alone. The AI foundry is open for business. OpenAI's Frontier platform launched this week, offering enterprises a dedicated space to build and manage AI agents, treating them with the same operational rigor as human employees. This arrives alongside the quiet rollout of GPT-4o, OpenAI's new flagship model, boasting multimodal capabilities, including accepting any combination of text, audio, image and video as input and generating any combination of text, audio, and image outputs. GPT-4o is not just a spec bump; it's twice as fast and 50% cheaper than GPT-4, fundamentally shifting the price/performance curve. The model's text and image capabilities are now live in ChatGPT for Plus users, with up to 5x higher message limits, while the new voice mode is slated for release in the coming weeks. Even free users get a taste, albeit with usage limits, marking a strategic push towards wider adoption and, perhaps, a new era of AI ubiquity. The key to Opus 4.6 is its enhanced long-context capabilities, sporting a 1M token context window. Anthropic even tasked Opus 4.6 using agent teams to build a C compiler, which later worked on the Linux kernel. The model also uncovered 500 zero-day flaws in open-source code. However, not all benchmarks are created equal: Opus 4.6 placed only 26th on EsoBench, testing its ability to explore and code with novel esolangs, due to a tendency to second-guess itself and hallucinate output tags. It seems even the smartest minds have their off days.