article — General

Claude Mythos 6? The Biggest AI News This Week

About this Resource

A roundup of the biggest AI developments this week, covering Anthropic’s mysterious Model 2, Codex 2.0, Gemini 3.7 Flash, DeepSeek V4 Pro, and NVIDIA’s latest AI model.<h2> </h2><h2>Claude Mythos 6? Major AI Leaks, Gemini 3.7 Flash, Codex 2.0 & More</h2>The AI world has had another busy week, with major updates from Anthropic, Google, OpenAI, DeepSeek, and NVIDIA. From a mysterious new Claude model to faster coding tools and cheaper AI models, several developments could shape what comes next. <h2>Anthropic’s Mysterious “Model 2” Could Be Mythos 6</h2>The biggest story is Anthropic’s mysterious Model 2, an unreleased AI system that the company describes as a Mythos-class model that is slightly more capable than Mythos 5. This has led to speculation that it could eventually become Claude Mythos 6, although Anthropic has not confirmed that. What makes Model 2 particularly interesting is its performance on Cobbench 2, an internal benchmark designed around real AI research and development tasks. Mythos 5 scored 50.3%, while Model 2 reached 62.8%, a significant 12.5 percentage-point improvement. Anthropic estimates that a model reaching around 85% on this benchmark could potentially automate a large portion of the research work represented by the test. That does not mean AI researchers are about to be replaced. The benchmark covers a specific set of historical research tasks, and Anthropic has not completed its full evaluation of Model 2. Still, the results suggest that Anthropic may already have a significantly more capable generation of AI internally. <h2>Claude’s Invisible Watermark Raises Questions</h2>Anthropic has also introduced an invisible watermarking system for text generated by newer Claude models. The system is connected to transparency requirements under the EU AI Act. Importantly, the watermark does not appear to add hidden characters, metadata, or special tokens to the text. Instead, Claude subtly changes its word choices to create a statistical pattern that detection systems can recognize. Anthropic says the system cannot identify a specific user or conversation and is being applied globally. However, concerns have been raised about whether heavily edited human writing could potentially be identified as AI-generated. <h2>Codex 2.0 Could Make Long Coding Sessions Much Better</h2>OpenAI is also preparing a major update to Codex, its AI coding platform. In testing involving an extremely large 741-turn conversation, the updated Codex system reduced average loading time from 27 seconds to just 1 second, while using 41% less memory, making 98% fewer requests, and loading 99.6% fewer transcript items. The improvement isn't necessarily that the underlying AI model suddenly became dramatically smarter or faster. Instead, the Codex application is becoming much better at managing huge, long-running coding conversations. For developers who work on large projects, this could make AI coding sessions considerably smoother. <h2>Gemini 3.7 Flash Brings More Speed for Less Money</h2>Google's Gemini 3.7 Flash is another major release. It is designed as a fast, lower-cost model for coding, AI agents, web development, and general knowledge work. The model showed significant gains over Gemini 3.6 Flash across several benchmarks, including coding and web development. Its web-development ranking also jumped substantially, while it demonstrated stronger performance on complex documents and business workflows. One of its biggest advantages is speed and price. Gemini 3.7 Flash launched at roughly half the previous Flash model's pricing, while testing showed it could complete tasks significantly faster. For everyday users and developers, that combination of good performance, high speed, and lower cost could make Flash models increasingly attractive. <h2>DeepSeek V4 Pro Arrives, But Prices Are Rising</h2>DeepSeek has officially released DeepSeek V4 Pro, adding several features aimed at AI agents and production workloads. One useful addition is adjustable reasoning effort. Users can choose lower reasoning for simpler tasks, higher reasoning for more demanding workflows, and maximum reasoning for difficult problems. The model also adds native OpenAI-compatible API support, making it easier to integrate with existing tools. However, the release also comes with a major change, DeepSeek is becoming more expensive. V4 Pro introduces peak and off-peak API pricing, with peak pricing reaching $1.32 per million input tokens and $3.96 per million output tokens. Off-peak pricing is 50% lower. DeepSeek is still relatively inexpensive compared with many leading AI providers, but the price increase highlights a bigger issue: as AI models become more powerful, the computing power required to operate them at scale is becoming increasingly expensive. <h2>NVIDIA’s Tiny-but-Powerful Nemotron 3.5 Lightning</h2>NVIDIA has also released Nemotron 3.5 Lightning, an open 30-billion-parameter mixture-of-experts model designed for fast, always-on AI agents. Despite having 30 billion total parameters, only around 3 billion parameters are active for each task. NVIDIA claims this allows the model to deliver very high output speeds while maintaining strong capabilities. The model also supports a 1-million-token context window, making it particularly interesting for developers building local or continuous AI agents. <h2>Final Thoughts And What This Means for AI</h2>The biggest takeaway from this week's AI news is that progress is happening on several fronts at once. Anthropic appears to be developing increasingly capable AI systems that can perform sophisticated research tasks. Google is pushing faster and cheaper models, OpenAI is improving the infrastructure around AI coding, and DeepSeek continues to compete with efficient open models while adjusting its pricing. At the same time, NVIDIA's Nemotron 3.5 Lightning shows that smaller, efficient models are becoming increasingly capable. The important thing to watch isn't just which company has the biggest model. The AI race is increasingly about reasoning ability, speed, cost, coding, agent capabilities, and how efficiently models can run. And if Anthropic's Model 2 is any indication, the next generation of AI systems could be significantly more capable than today's models.