AI benchmarks

Google is back at the AI frontier: Gemini 4 Argon ties OpenAI's GPT-6 Astra at 53 on the Artificial Analysis Intelligence Index, at roughly 60% of the cost per task. The weekly AI roundup also highlights Europe closing the gap with Mistral Large 4, drawing attention to intensifying global competition among leading AI labs.

Chinese AI firm Zhipu is reportedly launching GLM-5.3-Highspeed, a new GLM model reaching speeds of about 490 tokens per second, more than twice as fast as GLM-5.3-FlashX, which peaks around 200 tokens per second. The model is said to be rolling out gradually in ZCode 3.15.1, with access being made available to randomly selected users.

Opper has launched Jevman, a benchmark that has AI decision models play Pac-Man to evaluate their reasoning and planning abilities. The project was shared on Hacker News, where it drew early attention from developers and AI researchers. Using classic arcade games as a testing ground is becoming a popular way to measure how well AI agents handle sequential decisions and long-term strategy.

Timescale has released pg_vectorsearch, a new vector search index for PostgreSQL that it says outperforms the IVFFlat and HNSW index types available in the widely used pgvector extension. The project is open source on GitHub. Developers are discussing the benchmarks and how the new index compares in practice for similarity search and AI workloads running on Postgres.

A developer known as Lasu has released Parsimony, an open-source benchmark that measures how much code AI models add when generating software. The tool is now live online, and its launch on Hacker News is drawing engagement from developers interested in comparing how verbose or efficient different AI coding assistants are.

A new benchmark tests AI systems on theoretical physics problems, and leading models are performing strongly on it. Researchers say the results show AI moving beyond routine calculation into genuine reasoning about abstract scientific concepts. Commenters are debating whether this signals AI's arrival as a real research tool or whether the benchmark overstates what the models truly understand.

An Arizona court ruling on the use of artificial intelligence could influence how courts across the United States handle similar cases. The decision is drawing attention from legal observers who see it as a potential benchmark for other states weighing how AI evidence and tools fit within judicial proceedings.

A new robotics benchmark from Vals AI finds that today's frontier language models score under 2 out of 100 when tasked with controlling a simulated humanoid robot. The result highlights how far current AI remains from general-purpose physical control, despite strong performance on text and reasoning tasks, and is drawing attention in the AI and robotics communities.

Pi 1.0, a new open-source coding agent, has debuted with claims of reducing token consumption by around 40% compared to existing tools. The release is being positioned as a notable efficiency gain for developers using AI coding assistants, where token costs are a major running expense. Details beyond the headline figures remain limited, and independent benchmarks of the agent have not yet been reported.

A discussion on why large language models produce oddly consistent, distorted pelican drawings is gaining attention online. Asking AI systems to generate a picture of a pelican has become a popular informal test of their image-generation abilities, and the results tend to share the same strange quirks. Commenters are debating what these shared errors reveal about how the models work.

Massachusetts is weighing measures that could make it a national benchmark on artificial intelligence safety regulation. If the state moves ahead, its approach could influence how other US states craft rules for AI systems, filling a gap left by the absence of comprehensive federal legislation and drawing attention from policymakers, tech companies and consumer advocates.

Google's new Gemini 4 Argon model has undergone a full round of real-world testing, with results pointing to a major leap in its ability to handle physics problems. Reports describe the model's 'physics prowess' as awakened, suggesting capabilities well beyond previous versions of the Gemini line.

New research from AI startup Samaya reports that frontier AI models outperform human financial analysts on earnings prediction tasks. The company's blog post compares model forecasts against expert benchmarks, suggesting large language models can now match or exceed professional judgment in a core financial analysis function. The findings are drawing attention among finance and AI communities, with debate likely over methodology and how much weight to give results published by an AI vendor.

Developers are discussing pg_vectorsearch, an index for PostgreSQL that its authors say is faster than the IVFFlat and HNSW index types offered by the popular pgvector extension. The claim has drawn attention on Hacker News, where commenters are debating the benchmarks, accuracy trade-offs, and how it compares with existing vector search options for AI and retrieval workloads.

Meta is reported to have trained AI models with roughly 1GW of computing power over the past year, and these closed-source models are said to top benchmark leaderboards across the board. The puzzle being discussed is why Mark Zuckerberg is not selling API access to them, a departure from the industry trend of monetising frontier models and from Meta's own history of open-source releases.

Google has released Gemini 4 Argon worldwide for developers and enterprise customers. The new AI model features a 1-million-token context window and reportedly ranks at the top of the Vals Index benchmark. Tech and developer communities are discussing the rollout and what its large context capacity and strong benchmark results mean for building AI applications.

Meta has launched Muse Wonder, a new AI product that reports say narrows the gap with rivals OpenAI and Anthropic. Coverage highlights how to try the model and early benchmark results comparing its performance to competing systems. The release signals Meta's continued push into the frontier AI race alongside the industry's leading labs.