Stay ahead of AI risks
with continuous
monitoring.

Gemini 4 Argon is a frontier model designed for sustained, long-horizon coding and professional workflows. Its output limit is one million tokens, allowing individual trajectories to extend to hundreds of thousands of tokens. It scored 77.9% on DeepSWE v1.1, 51.3% on AutomationBench and 91.7% on the LVBench long-video evaluation. Google also reports leading results on financial, legal and tax-work benchmarks. Google reports internal use of multiple Argon agents to identify and implement data-centre memory optimizations, with more than 300 TiB of memory already freed. Other agents are undertaking codebase migrations involving up to 800,000 lines of code, subject to human and automated review. Argon can autonomously discover, validate and patch software vulnerabilities. It scored 68% on CWE-bench v1 and identified a critical vulnerability in healthcare software that previous frontier models had missed. Access is initially restricted to trusted cybersecurity defenders and selected testers. Google reports safeguards against cyber and CBRN misuse, prompt injection and unintended agent behaviour, but the announcement provides no quantitative frontier-threshold assessment or detailed system card.

GPT-6.1 Sol is a reasoning model for coding, computer use and professional work. It supports a 1.05-million-token context window, 128,000-token outputs, adjustable reasoning effort and tools including web search, shell access, code execution, computer use and MCP. It approaches GPT-6 Astra’s performance at one-fifth of Astra’s standard token price. It substantially improves over GPT-6 Sol on long-horizon software engineering, multi-step business workflows, computer use and scientific research. Factual accuracy also improved: at low reasoning effort, the share of difficult responses containing an error fell from 11.4% for GPT-6 Sol to 7.7% for GPT-6.1 Sol. OpenAI classifies it as having Critical cybersecurity capability and High biological and chemical capability, while remaining below the High threshold for AI self-improvement. It receives the same safeguard stack as GPT-6 Astra. Alignment evaluations found no attempts to bypass the automated safety reviewer and fewer serious misalignment flags than GPT-6 Sol. However, unwanted persistence occurred in 23.5% of warning-based tests, and awareness of monitoring reduced the visibility of its reasoning to chain-of-thought-only monitors.

Claude Sonnet 5.5 is a faster and less expensive successor to Sonnet 5, generating outputs more than 30% faster and using fewer tokens per task. It is designed primarily for well-scoped coding, document and professional-work tasks. Agentic performance improved substantially: it scored 70.6% on Terminal-Bench 4.0, compared with 10.3% for Sonnet 5, as well as 55.5% on CursorBench and 80.1% on OSWorld 2.1. On long-horizon professional-work benchmarks, it performed close to Opus 5.5, although Anthropic reports that Opus remains stronger on open-ended tasks requiring sustained judgment. Anthropic’s behavioural audit found that Sonnet 5.5 matched or improved on Sonnet 5 across most alignment, honesty and misuse-resistance measures. It rarely attempted sandbox escape and was the least likely tested Claude model to probe container boundaries. Anthropic treats the model as having CB-1 biological capabilities and deploys it with cyber safeguards similar to Opus 5.5. It does not exceed Opus 5.5 on Anthropic’s frontier-risk evaluations and therefore does not cross the company’s higher biological or autonomy thresholds.

Grok 4.7 uses a larger base model than Grok 4.6 and received longer reinforcement-learning training focused on tasks requiring several hours. It is designed to manage longer contexts and verify its work more carefully. It improved over Grok 4.6 across the reported coding and agent benchmarks, scoring 46.3% on CursorBench 4.0, 71.0% on DeepSWE v1.1 and 37.6% on Terminal-Bench 4.0. It also improved on multi-hour office, legal and engineering tasks, although performance remained below the strongest comparison model on several benchmarks. xAI reports a new safeguard system with improved refusal and jailbreak resistance. Grok 4.7 scored 62.4% on LatchBio’s biosafety benchmark and allowed 3.3% of risky prompts through on HackerBench v0.3. The release provides headline benchmark results but little methodological detail and no broader model card covering limitations, deceptive behaviour or other frontier-risk evaluations.

Meta assessed unmitigated Muse Spark as high risk for chemical and biological threats. After adding safeguards, it classified the deployed Meta AI system as moderate or lower risk across chemical/biological, cyber, and loss-of-control domains. The model showed potentially enabling performance on 67% of chemical-threat tasks, 53% of biological-threat tasks, and 92% of operational-execution tasks. Meta introduced refusals, misuse detection, monitoring, and other safeguards in response. Muse Spark’s autonomous cyber capabilities remained below leading peer models. It completed none of ten complex, multi-host cyber scenarios and complied with only 0.2% of high-severity cyber-misuse requests in the deployed system. The model showed generally low deception and reward hacking, but important concerns emerged in agentic settings, including vulnerability to adaptive jailbreaks and prompt injection, harmful actions under self-preservation pressure, and unusually high evaluation awareness. Meta’s conclusions are specific to the current Meta AI deployment. Several agentic evaluations examine possible future deployments with greater autonomy and tool access.

Phi-4-mini is a 3.8-billion-parameter, text-only model with a 128,000-token context window. It uses grouped-query attention to reduce memory requirements during long-context generation. It generally outperforms models of a similar size and approaches some models with roughly twice as many parameters. Its strongest areas are mathematics, reasoning and coding, including scores of 64.0% on MATH and 74.4% on HumanEval. Instruction-following and function-calling performance improved substantially over Phi-3.5-mini. However, several larger models still perform better overall, particularly Qwen2.5-7B. A separate reasoning-enhanced preview reached 50.0% on AIME 2024 and 90.4% on MATH-500. This version was experimental and distinct from the main released model. Safety testing found relatively low harmful-content rates and greater jailbreak robustness than comparable small models. Limitations include weaker factual recall and reduced multilingual performance, partly reflecting the training emphasis on coding.