Tag
A developer shares how they cut DeepSeek API costs by roughly 90% through prefix caching optimizations, reducing browser agent task costs below $0.005 and enabling a free, ads-supported browser agent.
Unused tools in an AI agent's toolset still consume tokens and add noise to tool selection, so agents should load only the tools required for the current task.
An experiment testing whether a coding CLI actually reads AGENTS.md found the file was silently ignored, and even when read, bloated instruction files increased token costs and failed to improve performance. The author recommends writing only what the model cannot infer from code.
A developer benchmarks 13 search API configurations inside an AI agent, revealing that hidden token costs from reading payloads can dominate the total bill and vary by up to 67x across providers.
A tweet points out that most Claude Code token costs come from re-reading the entire codebase on each edit, and highlights a new category of tools that use codebase mapping, context pruning, and project memory to drastically cut token usage, with a list of 10 tools in the linked article.
A paper finds that self-reflection loops like Self-Refine and Reflexion do not beat repeated sampling at equal token cost across open models from 1.5B to 7B, with all 18 self-inspection comparisons negative.
Someone has open-sourced the Codex infinite ammo method, using Sol for planning and Luna Max for execution, saving about 59% of token costs on simple tasks and 65% on complex tasks.
The author discusses the problem of AI agents getting stuck in tool loops and proposes building 'Moven AI', an open-source SDK that acts as a circuit breaker to abort runaway agents before they incur excessive costs.
An updated benchmark shows self-hosting Kimi K3 on 8×B300 nodes achieves 86.4% task resolution at roughly 20% higher hardware cost compared to GLM-5.2 on B200 nodes, though with lower throughput.
After a Codex run wasted over 100,000 tokens without completing a task, the author built a tool that reliably finishes code generation jobs.
NVIDIA details how its full-stack inference software, co-developed with open-source ecosystems like PyTorch, reduces token costs by up to 5x on Blackwell GPUs, with real-world deployments from Baseten, Cognition, Deep Infra, and others demonstrating performance gains.
NVIDIA announces Vera Rubin platform, delivering 10x more throughput per megawatt than Blackwell and claiming lowest token cost for AI factories through extreme co-design across seven chips and five rack trays.
Chamath revealed that his company found AI token costs double every 45 days while downstream productivity improves by at most 5%. He believes AI development is approaching a bottleneck and suggests companies reconsider their strategies or even consider exiting.
Chamath reveals the harsh reality of AI costs vs. returns: token costs double every 45 days, but downstream productivity gains are at most 5%. Large model capability improvement has hit an asymptote, and within the next 3-4 years, every company will face an ultimate reckoning between cost and benefit.
Meta's aggressive pricing for its Muse Spark 1.1 model undercuts competitors like OpenAI and Anthropic, potentially sparking a token price war that could lower costs, prevent a duopoly, and spur innovation in the application layer.
Fable 5 model only used 4 prompts and $173 worth of tokens to create a game called 'Super Smart Racing', demonstrating its extremely strong generative capabilities.
NVIDIA's full-stack inference software, codesigned with hardware, has reduced token costs by up to 5x on the Blackwell platform in just one month, enabling lower cost per token for AI factories. Companies like Baseten, Cognition, Deep Infra, and Together AI are using the stack to optimize inference performance.
A report indicates that operating an AI datacenter in orbit costs 8 to 12 times more per token than a terrestrial datacenter, highlighting significant cost barriers for space-based AI computation.
Bellwethr is developing an open methodology for tracking the real USD cost of a single inference token from capable models, with a draft benchmark suite and community contributions underway.
Announcing version 1 of mattpocock/skills, a collection of AI skill definitions that reduces token costs by 63% and introduces new skills for codebase design, domain modeling, and more.