cost

Miles Brundage reflects on how language model tokens were cheap relative to 'human tokens' from the start, which helped drive rapid technology adoption because it was 'born cheap'.

GLM 5.3 FlashX is praised for improved performance but criticized for higher cost, while DeepSeek v4.1 Flash is noted to outperform it.

Astra AI model has announced its pricing with $10 per million input tokens and $50 per million output tokens for API usage.

The author, after trying Qwen's latest app, points out that for Agents to reach mainstream users, three bottlenecks need to be resolved: cost, response speed, and the way humans interact with AI.

Introduces the /swarm multi-agent parallel fan-out mode in the Claude Code ecosystem, which was recently disabled by default with hard limits added, due to runaway recursion and cost explosion.

A Reddit user asks how DeepSeek can make a 300B parameter model cheaper than a 9B parameter model, sparking discussion on efficiency and cost.

DeepSeek-V4-Flash-0731 reportedly outperforms GLM 5.2 while maintaining the same cost as its predecessor, sparking speculation of another market disruption.

Opus 5 is not significantly cheaper to use than Fable.

The article explains how prompt caching works in large language model agents, covering KV cache mechanics, prefill and decode phases, and the impact on latency, cost, and agent design.

AI's growing energy demand from data centers could spark public conflict as electricity bills rise, even among those who previously ignored AI benchmarks.

AI has made building software cheap and easy, but it hasn't reduced the ongoing costs of maintenance, support, and responsibility, leading to many abandoned projects.

A practical benchmark comparing Gemma dense (31B) and MoE (26B) models shows MoE is 25.5% faster and 20% cheaper per query with identical quality, validating theoretical cost savings.

A company shares the practical difficulties of purchasing and maintaining an HGX B300 for AI inference, including high cost (€1.1M), power and space requirements, and limited availability, with throughput estimates for GLM 5.2.

A Fortune 500 'AI First' company scales back AI usage, removing Claude and limiting usage due to high costs, reflecting broader concerns about AI operational expenses.

This paper investigates the computational cost of reasoning in non-English languages, using Japanese as a case study, to highlight inefficiencies in current AI systems.

Grok-4.5 achieves coding performance comparable to GPT-5.5-xhigh while costing half as much.

Chinese AI models are increasingly adopted by US companies as they narrow the performance gap with leading American rivals while being significantly cheaper, with OpenRouter token share rising above 30% weekly since February.

This article explains why AI tokens are expensive: the autoregressive generation process requires predicting token by token and repeatedly computing the entire context, causing computational cost to grow linearly with context length; coding agents, due to multi-turn interactions and file reads, quickly accumulate very long contexts, further increasing token costs.

Jay Alammar highlights how local AI models are becoming critical infrastructure at all scales, referencing a writeup by Sebastian Raschka that covers setting up local models, with a mention of Cohere's North Mini Code (Nomico) released under Apache 2.0.

A tweet asks about the most money people have seen an AI agent waste while trying to fix its own mistake, highlighting the financial risks of autonomous agents.