Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon top Coding Agent Index

A Reddit user reports a new Google image generation and editing model appearing on AI Arena under the name "resplendent_flash," speculated to be Nano Banana 2.5. No official Google announcement accompanies the sighting.

Ivan Leo, a developer experience engineer at Google DeepMind, presents the Gemini Interactions API and Managed Agents, tracing the shift from single completions to function calling to agents. The talk argues existing APIs didn't keep up as models became agents.

Pi pod executes sessions of the pi coding agent in isolated "pods" on a server you run, with composable environments and minimal curated tooling. The self-hosted edition is available now; a hosted service is coming soon.

Anthropic's guide says Opus 5.5 always thinks before replying and decides how much, so 'think step by step' lines should be deleted from prompts. It keeps going on long multi-part work better than Opus 5 did, with early testers running coding tasks for hours.

A blog post benchmarks Strata against llama.cpp on a single RTX 4070 12GB: 53.2 tok/s decode at 60k context versus 27.06 tok/s for llama.cpp with MTP and NVMe prefetching, and 2,013 tok/s prefill. Strata, ninfer, DwarfStar, Splash, llamAmpere and gufo each target a short curated model list and one hardware family.