OpenAI has closed the first $11 billion tranche of its new funding round at a $1.25 trillion valuation, with the full round potentially reaching $20 billion — yet the first tranche is less than half of what was originally planned, signaling growing caution among Silicon Valley investors. In parallel, OpenAI released BioBasicsBench with biotech firm Ligo and a cyber-jailbreak rating framework with the Frontier Model Forum, while the anticipated GPT-6.2 Astra did not materialize.
Benchmark
基准测试
Standardized datasets or tasks used to evaluate model performance metrics
Curated Intelligence
Air Street Capital has released its ninth annual State of AI Report, surveying the past year across research, industry, politics and safety. It concludes that the global frontier has consolidated into a three-way race among Anthropic, OpenAI and Google, while flagging benchmark collapse, agent risks spilling into production systems, and AI turning into a geopolitical lever.
SketchAgentBench is a new benchmark aimed at evaluating large language models on 3D spatial reasoning, addressing a gap in current evaluation suites that rarely probe spatial understanding. The available material only states its name and target capability, lacking task design, dataset scale, and experimental results, so its practical value awaits verification via the full paper or code release.
Ai2 has released ToyBlocks, a controlled simulation environment for robot learning and evaluation, with code, datasets, models, and documentation released together. The goal is to let researchers systematically study what a robot has actually learned and why it fails, moving robot learning beyond narrow demonstrations toward transferable capability.
webAI released TwIL-8B, a reasoning model built on IBM Granite-4.2-8B and optimized for math, code and instruction-following workloads, open-sourced under a non-commercial license. The lab claims best-in-group results on MATH-500, AIME 2025 and LiveCodeBench, reaching 88-93% of frontier-model capability on its strengths, while acknowledging limited world knowledge and no general-purpose scope.
OpenAI introduces GPT-6 Cyber, which lifts ExploitGym exploit performance from 32.1% to 48.5% (a 51% relative gain) and achieves 99.93%–100% scope-compliance in alignment tests. The model has already surfaced and validated thousands of vulnerabilities in critical codebases such as Linux and Firefox, including 948 Linux privilege-escalation exploits.
The Information reports that OpenAI could unveil its next-generation Astra and Sol model lines plus a cybersecurity-focused Cyber model as soon as Friday, November 6. The report offers no model identifiers, pricing, context-window size or benchmark results, and it is unclear whether the models would ship first via the API, ChatGPT, Codex or partner channels. The only concrete leads are a "gpt-6.2-astra" string spotted in Codex usage stats and anecdotal gray-test user claims of smoother physics and animation generation, with no independently verifiable data.
Artificial Analysis released version 5, addressing the 'benchmarker's dilemma' in independent LLM evaluation. The piece highlights the difficulty of creating forward-looking, single-number metrics that align with community sentiment and scientific rigor amidst rapid AI progress, warning against optimizing for scores over meaningful capability.
Google updated its Android Bench leaderboard, adding reasoning level annotations and new models. Claude Sonnet 5.5 leads with the highest pass rate, while GPT-6.1 Sol offers high completion rates at significantly lower costs. In contrast, Google's own Gemini 4 Argon performed poorly, ranking fifth and lagging behind competitors.
Artificial Analysis has launched public leaderboards for OpenAI's HealthBench and HealthBench Professional benchmarks. These evaluations assess AI models' performance in realistic health conversations using physician-created rubrics across five axes: accuracy, completeness, communication quality, context awareness, and instruction following. Early data indicates GPT-6 Sol leads in communication quality but lags in completeness, while Claude Sonnet 5.5 tops the completeness metric.