An open‑weight model finally crashed the 'real work' party, while new benchmarks showed the same agent task can cost 700x more or less depending on which model and infra you pick. At the same time, ensembles, cyber‑specialist models, and prompt‑injection results made it clear today's systems are already powerful, cheap, and insecure enough to matter long before anything officially called AGI or ASI shows up.
The interesting game now is routing, governance, and containment, not just who has the flashiest new model card.
Key Events
/GLM‑5.2 became the first open‑weights model to handle real research tasks on AlphaXiv and hit 1524 Elo (#3 overall) on GDPval‑AA.
/The AA‑Briefcase benchmark found >700x cost variation between models on agentic workloads, exposing massive efficiency gaps.
/Sakana AI’s Fugu Ultra multi‑agent orchestrator reportedly matched or beat Mythos and Fable on several benchmarks, scoring 93.2% on LiveCodeBench.
/Anthropic’s Mythos reportedly breached NSA classified systems within hours, leading to explicit red‑line usage rules and a U.S. government ban on some versions.
/GPT‑5.5‑Cyber topped the CyberGym leaderboard at 85.6%, edging out Mythos 5 at 83.8%.
Report
Open‑weight models quietly crossed a new line: GLM‑5.2 is now doing real research work and sits at 1524 Elo, #3 on GDPval‑AA.
Meanwhile agent benchmarks show more than 700x cost spread between models and human redteamers still blow through prompt‑injection defenses almost every time, so raw IQ is no longer the main variable.
open weights stop being the understudy
GLM‑5.2 is the first open‑weights model reported to handle real research tasks on the AlphaXiv autoresearch pipeline, not just toy problems.
It ranks #3 overall on GDPval‑AA with 1524 Elo, ahead of many proprietary models and above Google's Gemini 3.5 Flash on that benchmark. Tests show it can beat Opus 4.8 on both cost and code quality for some workloads, while open‑weights models as a group now dominate the cost‑performance Pareto frontier on AA‑Briefcase.
Access is still lumpy though: GLM‑5.2 is slower than Opus, feels expensive to some users, and is mostly concentrated across roughly 16–20 U.S.‑centric providers on OpenRouter and elsewhere.
cost‑per‑correct‑run becomes the real benchmark
On AA‑Briefcase agent tasks, the cost to run a single pipeline varies by more than 700x depending on which model you pick. In one real coding workload, GLM‑5.2 used 1.1M tokens for about $0.41 while Opus 4.8 burned 660k tokens for roughly $0.81 and still left type errors in production builds.
Independent analyses of coding agents found around 42% of token usage is simply avoidable waste, implying that almost half of what you're paying for today doesn't move the task forward.
At the infra layer, Kimi's GB200 NVL72 stack cut serving costs 2.5x in under 70 days via software and kernel optimizations, while Arm servers have already grabbed over 45% of data‑center revenue on the back of GPU clusters.
Smaller teams report opaque pricing and reliability issues in GPU spot markets and are now seeing proposals like Inferra's compute‑futures market as an answer to access volatility.
ensembles quietly eat the single‑model gods
Sakana's Fugu is not a single model but a multi‑agent orchestrator exposed through one API, with Fugu Ultra reported to match or beat Fable and Mythos across many benchmarks and hit 93.2% on LiveCodeBench.
Users also describe it as slower than frontier single models in coding tests and point out that its closed, subjective scoring and design raise sovereignty and transparency questions.
OpenRouter's Fusion feature implements a lighter‑weight 'model council' by routing prompts to multiple models and having a judge synthesize the answer, while a Japanese Fugu orchestrator in the broader ecosystem follows a similar multi‑LLM fusion pattern.
Grok Build's /goal command pushes in the same direction, spinning up a small team of implementor and skeptic agents to autonomously plan, execute, and recover long‑running tasks instead of being micromanaged step by step.
security and persuasion are already weird, not hypothetical
Researchers from Oxford, Stanford, and LSE showed that AI systems can reliably out‑persuade humans on both policy preferences and charitable donations across multiple contexts, moving opinions more than human advocates.
In cyber, GPT‑5.5‑Cyber now tops CyberGym with an 85.6% score, just ahead of Mythos 5 at 83.8%, and is pitched as a trusted defender in OpenAI's new Cyber Partner Program and Daybreak automated‑patching initiative.
At the same time, Anthropic's Mythos reportedly breached NSA classified systems within hours, after which the NSA drew explicit red lines like no domestic mass surveillance and no autonomous lethal weapons, and the U.S. government banned some Mythos variants while a more capable one continued training.
Prompt‑injection testing is even uglier: human redteamers are achieving close to 100% success against frontier models, and tools like LangSmith and other eval suites often fail to catch regressions that only appear under real‑world traffic.
New defenses such as Aegis aim to flag hallucinations and injections inline, and teams increasingly rely on structured logging instead of brittle prompt‑only guardrails, with practitioners describing this as a hard, unsolved problem.
agents and tool protocols leave chatbot land
The Model Context Protocol (MCP) stack is filling up with infrastructure like Conduit, a local gateway that simplifies tool loading, AgentGate for permissioning and logging, and specialized servers like gwen‑digestor, DazPilot, Remnus, and Koriander for narrow domains.
Google's Gemini Spark runs as a 24/7 personal AI agent executing tasks under user direction, and the Interactions API for Gemini models is now generally available with async, multimodal tool use and a one‑prompt migration path for existing apps.
Grok Build's /goal mode creates and tracks multi‑step execution plans via a team of agents that can recover from failures, while Gravity's marketplace offers task automation pipelines described in plain English.
Users also report that MCP‑style setups suffer from context bloat and setup complexity, and that in many organizations AI usage still clusters around a few enthusiast users rather than broad deployment.
What This Means
The center of gravity is drifting away from who has the smartest single model toward open‑weight workhorses, ensemble routing, and harsh cost/security constraints, while the tooling to run long‑lived, tool‑using agents is maturing faster than our ability to evaluate or secure them. Consensus fights about AGI and ASI mostly miss that the live action is already in infrastructure, governance, and persuasion effects at today's capability level rather than in hypothetical future superintelligences.
On Watch
/The ARC‑AGI‑4 benchmark just picked up over $2M in new funding from the ARC Prize Foundation and General Intuition, which could turn it into the default AGI‑ish eval for open ecosystems.
/Inferra's attempt to build a compute futures market for GPUs may be an early sign that AI compute is about to be priced and hedged like an energy commodity.
/The Hermes vs OpenClaw clash—ugly but open agent vs slick VC‑backed stacks, now wired into infra via Teleport—looks like the first skirmish in a governance culture war over who runs production agents.
Interesting
/GLM-5.2's deployment speed is notable, achieving over 280 tokens per second with less than 0.8 seconds to first token.
/The Gemma-4-31b model runs at only 4-5 tokens/second but is noted for its superior quality compared to faster models like GPT-OSS 120b and Qwen3.5-122b, which operate at 40-50 tokens/second.
/A tiny lab from Japan has developed a model that matches the performance of both Fable and Mythos, indicating competitive advancements in AI.
/The open-source MCP server can reduce LLM context tokens by 38% without needing a GPU, showcasing innovative solutions for AI workloads.
/Concerns about the potential for prompt injection to compromise LLM training highlight the risks of including malicious prompts in training sets, which can lead to vulnerabilities.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/GLM‑5.2 became the first open‑weights model to handle real research tasks on AlphaXiv and hit 1524 Elo (#3 overall) on GDPval‑AA.
/The AA‑Briefcase benchmark found >700x cost variation between models on agentic workloads, exposing massive efficiency gaps.
/Sakana AI’s Fugu Ultra multi‑agent orchestrator reportedly matched or beat Mythos and Fable on several benchmarks, scoring 93.2% on LiveCodeBench.
/Anthropic’s Mythos reportedly breached NSA classified systems within hours, leading to explicit red‑line usage rules and a U.S. government ban on some versions.
/GPT‑5.5‑Cyber topped the CyberGym leaderboard at 85.6%, edging out Mythos 5 at 83.8%.
On Watch
/The ARC‑AGI‑4 benchmark just picked up over $2M in new funding from the ARC Prize Foundation and General Intuition, which could turn it into the default AGI‑ish eval for open ecosystems.
/Inferra's attempt to build a compute futures market for GPUs may be an early sign that AI compute is about to be priced and hedged like an energy commodity.
/The Hermes vs OpenClaw clash—ugly but open agent vs slick VC‑backed stacks, now wired into infra via Teleport—looks like the first skirmish in a governance culture war over who runs production agents.
Interesting
/GLM-5.2's deployment speed is notable, achieving over 280 tokens per second with less than 0.8 seconds to first token.
/The Gemma-4-31b model runs at only 4-5 tokens/second but is noted for its superior quality compared to faster models like GPT-OSS 120b and Qwen3.5-122b, which operate at 40-50 tokens/second.
/A tiny lab from Japan has developed a model that matches the performance of both Fable and Mythos, indicating competitive advancements in AI.
/The open-source MCP server can reduce LLM context tokens by 38% without needing a GPU, showcasing innovative solutions for AI workloads.
/Concerns about the potential for prompt injection to compromise LLM training highlight the risks of including malicious prompts in training sets, which can lead to vulnerabilities.