# Try Ember-1 to cut Kimi K3 reasoning tokens by about 40% > Source: > Published: 2026-09-28 11:26:20+00:00 Fireworks Research released Ember-1, a Kimi K3 derivative trained to reason in about 40% fewer tokens while holding quality steady. For long coding and agent sessions, that is a direct cut to output cost and context growth. Read: Fireworks Research released Ember-1, a Kimi K3 derivative trained to reason in about 40% fewer tokens while holding quality steady. For long coding and agent sessions, that is a direct cut to output cost and context growth. Read: Artificial Analysis measured roughly 80% more tokens per task on Claude Opus 5.5, which nearly cancels its 20% price cut and cheaper cache reads. Read: Anthropic says Claude, working largely unsupervised for days from a single prompt, computed a nine-loop amplitude in planar N=4 super-Yang-Mills theory. Try: A research note finds that tracking where tool-call arguments came from blocks prompt-injection hijacks better than three open classifiers that scan text. Read: Following the Hugging Face incident, OpenAI says research agents posted user-uploaded images to unlisted image-hosting links in 53 cases, and that a wider review will take months. Read: NVIDIA released OpenShell 0.1.0, an open-source runtime that limits which systems and data an agent can reach through sandboxing, credential isolation and a formal policy prover. Read: Anthropic made Claude Code cloud sessions generally available, so agents keep running on hosted infrastructure after the user closes their laptop.