AI Newsletter

OpenAI has officially announced GPT-6 Astra – the model achieves 98.6% on ARC-AGI-3 and demonstrates critical cybersecurity capabilities for the first time (100% on ExploitBench). Internally, the next pretraining model 'Bel' is reportedly already running, which will surpass Astra once more. The launch marks a qualitative threshold that puts competitors like Anthropic and Google under acute imitation pressure.

Thomson Reuters has released its own Large Language Model called 'Thomson', based on Qwen3.5-397B and continuously fine-tuned on 175 years of its own legal, tax, and accounting data. This move illustrates the trend toward vertical, domain-specific frontier models by established data providers. For legal tech and enterprise AI startups, the competitive landscape is fundamentally changing, as proprietary data moats become the decisive defense.

Mistral AI has completed a funding round of $3 billion and reports that its open-weight models achieve GPT-4 level performance at half the API price. The combination of massive capital and aggressive pricing on open weights significantly increases price pressure on commercial providers. For enterprises, a serious alternative emerges to dependence on proprietary cloud models.

Google DeepMind has announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, combining real-time conversation with asynchronous background reasoning. The model directly addresses the market for agentic voice and task automation applications. Despite this new release, the prediction market sees Anthropic as monthly leader with 96% probability – Google remains at 2%.

NVIDIA's Inception program offers AI startups without VC funding, without equity stake, and without deadline free training credits, cloud resources, and introductions to investors – achieving broad resonance in the founder community with 2,529 likes. The program solidifies NVIDIA's position as a central infrastructure gatekeeper in the AI ecosystem beyond pure hardware sales. For early-stage startups, it significantly reduces capital requirements and strengthens binding to NVIDIA's stack.