The latest AI news roundup from Latent Space highlights a wave of frontier model releases. Claude Opus 5.5 tops SimpleBench at 88.4% and is considered Anthropic's best vision model yet, while the GPT-6 family—Astra, Sol, and Luna—shows strong performance in gaming and coding benchmarks. Google's Gemini 3.8 Flash scores 41 on the AA Intelligence Index at 291 tok/s with 1M context and is free in Cline, and Xiaomi's MiMo-V2.6-Pro is an open-weight, omni-modal model with 1M context at a fraction of the cost of rivals.
A notable theme is the rise of decision models, led by TypeSafe's Jev. Trained with reinforcement learning for calibrated decisions, Jev returns typed decisions with probabilities instead of reasoning text. According to the Jev-as-a-Judge paper, it costs $0.044 per 1K judgments at 152ms median latency—about 277x cheaper than GPT-6—and a cascade that escalates low-confidence calls to GPT-6 Astra keeps 99% accuracy at 57% of the cost. Alternatives like CLM are faster but less accurate, and Fastino's GLiNER2.5-Decide adds structured decisions at CPU-friendly speeds.
Ecosystem signals show growing adoption of these efficient models: Ramp matched GPT-5.6 Luna reranking accuracy with 10x lower tail latency and 3x lower cost, and turbopuffer now natively includes Jev for reranking. The source is a single newsletter, so there are no conflicting accounts—just a broad snapshot of a busy week in AI development.