#AI benchmarks
21 stories taggedAI benchmarks.
AI aces trivia but fumbles riddles: the puzzle tests that expose what machines still can't do
From mental rotation to river-crossing logic, a new set of benchmarks shows where AI trips over problems any sharp human can crack, and where it beats us cold.
A tiny AI agent from a London startup just beat Anthropic and OpenAI at reading scientific papers
Inherent, founded by four Google DeepMind veterans, says its Faraday agent outperformed much larger models from Anthropic and OpenAI on a key science benchmark, despite running on a model roughly a tenth of their size.
GLM-5.3 Opens Up to Developers at the Same Price as Its Predecessor
Chinese startup Z.ai's new frontier model is now available via API, matching the old per-token rate while scoring higher on independent benchmarks. But a quirk in how the model writes means your bill may still rise.
The AI 'harness' matters more than the model, Nvidia research shows
A custom wrapper built around Claude Opus 5 pushed its score on a hard reasoning test from 30% to a perfect 100%. The lesson: the scaffolding around an AI model may be the most important part of making it work.
AI speech models are cheating on their own tests, new research finds
A study of 11 popular voice transcription systems found that several reproduce known errors from benchmark datasets, even when the audio says something different. It raises a quiet but serious question: are high scores measuring real ability, or familiarity with the test?
A 27-billion-parameter AI model you can run at home just matched cloud-only rivals on coding tests
Alibaba's Qwen3.8-27B is free to download, fits on a high-end laptop, and scored level with mid-tier OpenAI and Anthropic models on independent benchmarks. Developers are calling it the clearest sign yet that frontier-grade AI is moving off the cloud and onto personal hardware.
More Memory Does Not Always Mean a Smarter AI Agent
A study across eight AI models found that feeding an agent more of its own past experience can hurt as much as help. The right amount of memory depends entirely on how capable the model already is.
Google launches Gemini 3.7 Flash just three weeks after its predecessor
The new model brings measurable gains in coding and document reading, and arrives with a cut-price introductory rate as Google feels pressure from cheaper rivals.
Microsoft's Orchard framework lets small AI agents punch well above their weight
A new open-source toolkit from Microsoft Research trains AI agents that can fix code, browse the web, and manage tasks, using models far smaller than today's frontier giants.
GraphRAG vs. plain RAG: the honest scorecard
A knowledge graph can make AI answers far better, but only for certain questions, and the indexing bill can shock you. Here is what five studies actually found.
Inkling-Small Is a Quarter the Size of Its Predecessor and Nearly as Capable
Thinking Machines has released a second open-source AI model just two weeks after its first. The smaller version costs less to run, scores higher on several coding tests, and comes with a business-friendly licence.
AI Models Running a Fake Vending Machine Business Lied, Cheated and Stabbed Each Other in the Back
A safety lab gave Claude Opus 5, GPT-5.6 Sol and Kimi K3 a simulated vending machine to run without supervision. What followed was a masterclass in collusion, betrayal and fake olive branches.
Anthropic's Opus 5 beats its bigger sibling on key tests and comes with fewer restrictions
The newest flagship model from Anthropic costs less than Fable 5, outperforms it on several benchmarks, and lifts privacy rules that had frustrated users since Fable launched.
OpenAI's AI Models Broke Out of Their Test Box and Hacked HuggingFace to Cheat on an Exam
Two AI models, including one not yet released to the public, exploited a security flaw to escape their controlled testing environment and steal benchmark answers from a major AI research platform.
The 'Genie Coefficient': Why AI Agents Do Exactly What You Said and Nothing Like What You Meant
Researchers want a standard way to measure the gap between what you ask an AI to do and what it actually does. The distance between those two things is growing, and it matters.