Which AI writes the better take? You decide — blind.

Gemini 4 Carbon is now available, adding a new Gemini model option for production evaluation. Teams shipping LLM features should run prompt, tool-calling, latency, cost, and safety regressions before routing traffic, because this can change model selection and fallback strategy for Gemini-backed systems.

An 800M-parameter agentic language model, BotPsy, achieves state-of-the-art tool-use performance for its size, enabling on-device deployment of personal assistants for tasks like web search and device control, reducing reliance on costly cloud LLMs and improving data privacy for users. This shift allows for cheaper and more private processing of routine workloads. It breaks the need for frontier-level intelligence for simple tasks.

Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens for shorter prompts, while sitting on the Artificial Analysis efficiency frontier: no model scores higher for less per task. For production agents, it is a serious default candidate for high-volume tool calls, routing, extraction, coding subtasks, and long-document parsing, but you should benchmark at your actual reasoning level because token price alone will not predict total task cost.

EarthRecord tests 12 multimodal agents on 238 land-change questions over 56 sites using four decades of raw Landsat, MODIS, VIIRS, and Sentinel-2 data, requiring agents to handle cloud masking and cross-sensor harmonization themselves. The key failure mode is not tool use but grounding: agents can complete many workflows, yet when one satellite stream is altered, about half of their correct answers follow the corrupted data rather than real ground change. For production Earth-observation agents, this means end-to-end completion metrics are unsafe; you need sensor-perturbation and cross-instrument consistency tests before trusting outputs as physical change.

SAE-feature topic monitors trained only on prefill activations generalized to per-token monitoring during decoding on Gemma-3 1B and 4B across seven topics, with labels able to switch mid-response. Practically, this means you can train cheaper offline classifiers on prompt/prefill traces and deploy compact runtime monitors during generation; without SAE features, matching performance required classifiers several times larger, and common perturbations cost at most 5 macro-F1 points.