Ember-1 hits Kimi K3 quality with 40% fewer reasoning tokens
The reasoning-cost squeeze is on: Fireworks undercuts K3 on Terminal Bench by 52%, while Z.ai and Qwen ship 'Prime' speed tiers and AGENTS.md hides a catch.

Copy markdown
Kimi K3 quality, ~40% fewer tokens, on the API
Fireworks Research's Ember-1 is built on Kimi K3 but trims its reasoning: ~40% fewer tokens overall for comparable answers. It scores 82.0% on Terminal Bench 2.1 (K3-max: 80.9%) at 51.9% lower cost, and 92.2% on SWE-bench Verified. It's live now as a Research Preview on Fireworks Serverless and OpenRouter — $3/$15 per M ($0.30 cached), 1M context — a near-drop-in swap for workloads where per-task cost and latency bite.
GLM 5.3 Prime: same weights, 1.5-2x faster
Z.ai shipped a high-speed tier of GLM-5.3 — identical weights, 1.5-2x the output throughput via inference acceleration — at $2.80/$8.80 per M, 1M context, 131K max output. If GLM already runs your agent loops, Prime buys throughput without changing models or re-running your evals.
Qwen3.8 Max Prime joins the speed-tier wave
Alibaba's Qwen3.8 Max Prime is a higher-throughput cut of Max — multimodal in (text, image, video), reasoning and tool-calling on by default — at $4/$12 per M ($0.50 cached), 1M context, API-only via Alibaba Cloud. Two 'Prime' launches in a day signal the frontier's next axis is speed and cost, not just benchmark scores.
Update: AGENTS.md reportedly loads only with telemetry on
A developer reports that Claude Code 2.1.277+ gates AGENTS.md loading behind a remote feature flag (tengu_agents_md_mod) that only resolves when telemetry is enabled. Set DISABLE_TELEMETRY=1 or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 and your AGENTS.md is reportedly skipped with no warning. Workaround: point to it via @AGENTS.md from a one-line CLAUDE.md. No Anthropic response yet.