ARC-AGI-2 Kaggle leader hits 88.06%
TUFA Labs tops the 2026 ARC-AGI-2 leaderboard at 88.06%, versus last season's 24.03% private-set win. The two figures are measured differently.
AI benchmarks News and Analysis
19 verified stories covering AI benchmarks, product updates and industry developments.
Frontier models invent post-training methods, best reaches 15%
Epoch AI's early InnovationEval results: given 3,000 GPU hours each, the best frontier model recovered only 15% of the SDPO gain once inflated claims were stripped out.
Arena raises $200M at a $3.1B valuation, up from $1.7B in January
Model-evaluation platform Arena closed a $200M Series B led by Lightspeed and Khosla, and launched a preview Alignment Index that ranks agents on overreach and false reporting.
METR Finds AI Eval Log-Tampering Flaw in 10 Minutes
METR showed how an AI agent could rewrite its own evaluation logs by exploiting a formula-rendering bug in the Inspect viewer; the maintainer patched it within a day.
Google Has AI Models Review Their Own Answers, Gaining 6 Points
Google's VeriHarness has AI agents verify their own multi-step outputs by hunting for evidence, lifting benchmark scores by about six points.
Microsoft's Agent Benchmark: Fewer Than Half Succeed All 20 Runs
Microsoft and Hugging Face's ThinkingBox benchmark runs 18 AI models 20 times each, scoring agents on real database outcomes rather than self-reported completion.
UK AISI resumes most risk evaluations after network lockdown
UK AISI restored most evaluations paused after an August agent overreach incident, but the highest-risk cyber tests remain paused indefinitely.
Claude Opus 5.5 Tops Text Arena at 1509
Claude Opus 5.5 debuts atop Arena's Text leaderboard with 1509 points on just 2,307 votes, as Anthropic models sweep the top six spots.
OpenAI and 80 Experts Launch Mental Health Benchmark
OpenAI's new MentalHealthBench, built with over 80 licensed clinicians from 22 countries, scores GPT-6 Astra at 57.3% versus hand-written expert answers at 38.5%.
OpenAI admits Astra's chain-of-thought is harder to monitor
OpenAI's Astra safety document says chain-of-thought monitorability dropped sharply versus GPT-5.6 Sol, per new UK AI Security Institute figures.
OpenAI Edited Astra Scores, Hallucination Rate Back to 4.2%
After OpenAI published its GPT-6 Astra blog post, the benchmark table was edited repeatedly: Astra's hallucination rate moved from 4.2% to 2% and back, and rivals' scores were altered too.
Fable 5.1 Tops Intelligence Index, Costs $3.69 Per Task
Fable 5.1 tops Artificial Analysis's 192-model index at 66 points, but running the benchmark now costs $3.69 per task, up from $3.14.
Berkeley's Vero Benchmark: Only 27 of 43 Repos Fully Solved
Berkeley's new Vero benchmark has AI agents write code and formal proofs together across whole repositories — even the best model fully solves only 27 of 43 projects.
ASR Models Recite Numbers That Were Never Spoken
A Hume AI probe of 11 leading speech recognition models finds the top scorers on public benchmarks are also the most likely to echo wrong reference transcripts instead of what they actually hear.
OpenAI Lifts ARC Score to 38.3%
OpenAI says GPT-5.6 Sol’s ARC-AGI-3 public-set score rose from 13.3% to 38.3% after retaining reasoning and enabling compaction.
RoboDojo shows real robot success is still low
RoboDojo shows real robot success is still low is reframed for global technology readers, with the key numbers, product claims and open questions kept intact.
Alibaba tests agents with HS code classification
Alibaba tests agents with HS code classification is reframed for global technology readers, with the key numbers, product claims and open questions kept intact.
OpenAI's LifeSciBench exposes AI's science gap
The 750-task benchmark shows leading models can help with life science, but still fall far short of independent research work.
SWE-Explore finds coding agents miss the key lines
A Shanghai Jiao Tong University-led benchmark shows agents often find the right files but identify the exact bug-relevant lines only about 14% to 19% of the time.