Tag
Revolut moved ~2,000 engineers onto self-hosted GLM models for coding, retiring Claude Code and Codex to an exception list — running on 56 Nvidia B200s at a blended 30 cents per million tokens (about a tenth of prior Anthropic/OpenAI spend), with engineering lead time improving 70%.
Framer's Harness 2 upgrades agent code execution by running JavaScript as it streams, cutting costs by up to 33% and speeding up execution by 20% while reducing tool steps and cached token usage significantly.
The founder of EvalShift shares a model swap case where cost dropped 59% and latency improved 75%, yet agent behavior regressions like fake tool completions and 25% tool-selection divergence would still block the ship, arguing cost/latency alone can't certify a migration.
MIT Technology Review argues teams should measure AI development by cost per merged pull request instead of per-token spend, recommending model routing and owned open-weight capacity to curb the rising economics of agentic coding.
新分析指出，最便宜的 LLM 并不总是完成任务的最便宜方式：在会话中途切换模型可能因丢失上下文缓存而抵消成本节省，智能路由应优化整个工作流而非仅看模型价格。
Propasafe-Hybrid 是一个句子级的混合式宣传检测系统，将微调的 Transformer 分类器与 LLM 技术分类相结合，并通过成本感知的预过滤阶段减少 token 消耗，同时提供可解释的标注和理由。
作者分享如何用一台12年前的办公PC运行本地脚本（IMAP邮件轮询、Tesseract OCR、Piper TTS），把Telegram机器人例行任务从昂贵的LLM API调用转为零token的本地处理，显著降低成本。
MIT Technology Review examines the overlooked data storage costs of AI inference, noting that a single 8-second AI-generated video can leave behind 500MB of back-end data (130x the user-visible file) and citing IDC/McKinsey forecasts of 35% CAGR inference demand growth through 2030.
作者探讨构建研究型 AI 智能体时的两种架构选择：直接调用一体化 deep research API（Gemini Deep Research、OpenAI、Tavily Research 等），或基于搜索/抓取原语（Firecrawl、Exa、Tavily）自建规划与合成循环，并就质量、可观测性、扩展性与成本估算向生产环境实践者提问。
作者导出10小时Grok Bot使用数据发现89%消耗来自后台子任务、96%输入token用于重读旧上下文，并总结出5条规则，预计可节省25-30%的用量。
BotsCrew reports benchmark results comparing GPT-5.6 Luna to GPT-5.2 ahead of OpenAI's January 14 shutdown of GPT-4o and GPT-4.1, finding the cheaper Luna models win on cost but perform worse on long tool-heavy conversations and guardrails.
The author shares a Towards Data Science article on combining cost, quality, and latency into a single convex score for evaluating AI agents, turning business constraints into math that guides agent optimization instead of ad-hoc prompt tweaking.
A practitioner ran a cost benchmark on three parallel research agents sharing a memory of page digests and prefix caching, finding essentially no savings ($1.11 vs $1.09) because concurrent starts caused empty recalls, prefix cache misses, and digest overhead in output tokens.
MiniMax 发布 M3.1 Flash 模型，速度提升 3 倍、成本降低 70%，专为智能体（agents）场景优化。
研究发现安全漏洞挖掘AI代理60.4%的token用于理解代码与推理触发方式，并提出AVRI工具复用代码关系知识，在Codex和OpenCode上分别降低18.0%和23.7%成本且不损失成功率。
The paper introduces TokenEconomizer, an experience-driven external controller that reduces token costs of coding agents by injecting micro-skills, pruning irrelevant tools, and interrupting failure loops, achieving 18.4–30.9% savings across Claude Code, Codex, and GitHub Copilot CLI with comparable task success.
该论文提出一种基于约束上下文老虎机的个性化 LLM 路由框架，通过 primal-dual GLM-UCB 算法在线学习用户偏好，在最小化推理成本的同时以高概率保证响应质量满意度。
一篇博客文章系统分析了决策模型（如 Jev）的经济学，探讨其 choice/score/noul 输出类型、几乎免费的输出 token 成本，以及由此衍生的"决策层级"内存层级类比。
A free browser-based LLM cost calculator lets you compare local GPU inference (with RTX 4070 Ti Super / 5090 presets) against cloud API pricing, highlighting that cached-input token rates — often guessed at by other calculators — most determine where the local-vs-cloud crossover sits.
推文讨论用开源的 MiniCPM-V 4.7 在自有硬件上做图像理解，避免每次上传图片都调用付费视觉 API 产生持续云成本，并介绍如何将其集成到 agent 工作流中。