Tag
Garry Tan mentions efforts to slim down GStack to reduce token usage and task latency, with a reply noting CI pipeline times dropping from 7 to 3 minutes while asking about latency spikes from 20 running subagents.
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。
A user suggests that Grok products should proactively alert users when their token burn rate is extremely high on a task, to prevent accidentally exhausting weekly usage limits.
llm-echo 0.5 发布更新，改进了模拟 token 用量统计，现在会考虑工具调用、系统提示和推理 token，用于调试 LLM CLI。
该推文分析认为Anthropic的"增长放缓"可能被过度解读，指出8月假期季节性因素及OpenAI新模型Astra的发布分流了部分用量，但长期看Anthropic在高难度任务上仍将占据重要份额，开源模型则主导标准化任务。
Anthropic's Claude Opus 5.5 model family has been hugely popular, with OpenRouter data showing that within roughly 6 days of its launch, Anthropic reclaimed a significant share of token usage and spending from OpenAI.
Vercel的AI Gateway十月生产指数显示，开源权重模型已处理71.5%的token流量并占近23%的支出，自6月以来支出份额翻了四倍多；整体AI支出较12月增长13倍，而每token均价下降47%。
一名用户详细拆解了自己一周内使用 Grok Bot 的费用：升级到每月 $200 套餐，其中 52% 的成本来自 16 步以上的长任务，90% 的 token 是对已有上下文的重复读取。文章给出了测量用量、脚本化重复任务、将长任务交给独立代理等省钱建议。
A hands-on comparison of Grok Bot and Muse as personal AI agents over 20+ hours each, concluding Muse wins on price with generous token allowances while Grok Bot wins on smarts but burns through expensive usage tiers fast due to frequent scheduled jobs and repeated context re-reads.
A practitioner shares tips on conserving AI usage limits by choosing between single-agent and multi-agent setups in Codex and Claude, noting that subagents and repeated cross-checking burn tokens without improving results. The author tests Opus 5.5 as an orchestrator that dynamically picks Haiku 5.5 Short or Long with tuned cache settings, cutting usage to roughly 5% of the weekly allowance over two hours.
Aaron Levie 借 Prime Intellect 的最新实验指出，AI 智能体群将是未来 token 消耗的主要场景，可在软件开发、网络安全、生命科学等企业领域并行协作，实现人类无法企及的执行方式，这也是未来需要 1000 倍算力的原因。Prime Intellect 的 Prime Agent 此前指挥超过 2000 个智能体在两周内用 Rust 重写自身，消耗了 2000 亿以上 GLM-5.3 token。
文章预测一两年内与AI聊天将显得过时，AI代理将在后台协作、自我复制并以群体方式运行，这可能使token用量激增1000倍，届时算力与网络带宽都将成为重大瓶颈。
A technical comparison shows that running a cheaper, lighter model like Gemini 3 Flash can cost far more than a pricier reasoning model like Gemini 3 Pro — Flash burned ~1,000 steps and $14 in tokens before failing, while Pro succeeded in 85 steps for just $1, highlighting how token usage varies across models and even runs.
作者对比了 $200 的 GPT（Codex，使用 GPT-6 Astra）与 Claude（主要使用 Opus 5.5）订阅额度，估算 Claude 每周额度约等值 $3,400 API 成本，而 Codex 约为 $1,700，考虑缓存命中率后 Claude 的耐用度约为 Codex 的 10 倍。
A user reports that Android charging notifications repeatedly woke their OpenClaw Astra agent, resulting in 127.8M recorded tokens in a single day — though ~98% were cached input tokens, not billed at the full rate. The post seeks advice on debouncing system-notification events or isolating background events to smaller contexts.
Jamon Holmgren shares that a `retro` skill (from Matt Pocock) helped him find ~15 real issues in his AI agent harness that were wasting tokens, such as linter error hints that misled agents into skipping checks and incomplete build symbols.
Vercel AI Gateway的CEO @rauchg 表示，大多数token聚合器充斥着噪音，包括那些要求你把数据用于训练的token促销，以及声称支持ZDR（零数据保留）但信誉存疑的提供商；而Vercel AI Gateway凭借真实客户用量、40万+付费客户规模和零加价，成为理解全球AI token流向的最可信数据源。
The tweet discusses predictions that AI agents will use 100x more tokens than humans by 2030 and introduces Fo, a personal AI product that combines AI with human assistance for real-world tasks.
The tweet compares Opus 5.5's performance to Astra on a task, noting it performs better but uses 10x more tokens and draws faster.
According to Bloomberg, Harvey's costs for renting OpenAI and Anthropic models have risen sharply due to a 20-fold surge in token usage, causing gross margins to drop from about 50% to -50% by June.