token-efficiency

Google's blog post introduces Agent Skills support in Genkit for Go, demonstrating how progressive disclosure lets agents load specialized knowledge on demand to reduce token usage and improve reliability.

This paper introduces SIEVE, a search-inspect-fetch strategy that uses Boolean Query Language to make deep-research agents retrieve only relevant document sections, achieving higher accuracy with 20.7–50.6% fewer tokens across multiple benchmark datasets and agent backbones.

Hermes Agent has become dramatically more efficient, especially for smaller/weaker local models, thanks to Nvidia's Nemo Relay and optimizations like reducing turns, context load, and token waste across 250k conversations.

The article argues that model routing isn't the real problem to solve—token efficiency is. It advocates for a holistic closed-loop approach combining cheaper defaults, preference-aware routing, and better caching (e.g., Coinbase's 5%→60% cache hit improvement) to maximize useful intelligence per dollar.

Omar argues that token efficiency in AI models is underestimated and cites Artificial Analysis reporting DeepSeek completing benchmark tasks at 105x lower cost than Fable.

Benchmark results for MindControl, a sampler-level guided reasoning budget for llama.cpp, show multi-stage signaling reduces token consumption without accuracy loss, and sometimes improves it on easier tasks.

ReToken introduces a single learnable retrieval token that selects sparse query-relevant visual tokens from a pre-filled visual KV cache, improving long-context vision-language model performance on visual retrieval tasks. It yields consistent gains across image and video benchmarks while enabling efficient long-video inference on a single H100.

LangChain releases Deep Agents v0.7, a leaner agent harness that cuts base input tokens by 65% through removing the default system prompt, trimming tool descriptions, and making TodoListMiddleware opt-in, with comparable performance validated across multiple models.

GPT-5.6 Sol, a model that solved open math problems, initially struggled with the ARC-AGI-3 benchmark due to a harness memory limitation. Enabling two API settings tripled scores with 6x fewer output tokens.

A blog post reflects on Rails' conventions as ideal for AI agents and compilers, citing the Spinel ahead-of-time Ruby compiler as a project that could eliminate the need for application rewrites after scaling.

JetBrains benchmarked the Caveman token-compression skill on Claude Code across 86 tasks, finding real output-token savings of about 8.5% (not the advertised 65%) with no detectable degradation in task quality.

The article reflects on the rapid adoption of AI code assistants among developers and the emerging focus on token efficiency and cost metrics as AI usage becomes more corporate and metered.

Introduces Latent Briefing, a method for multi-agent systems to communicate via KV cache compaction, achieving 31% fewer tokens with same accuracy and up to 20x speedup.

A developer benchmarks Rote, a memory manager for browser agents that sends page diffs instead of full re-renders, showing a 37% reduction in token growth compared to Browser Use, but with trade-offs on short tasks.

Anthropic released Opus 5, focusing on token efficiency and cost reduction rather than a major capability leap, offering performance close to Fable at half the cost.

The author tested the chain of draft paper on their own setup, achieving 74% fewer tokens for the exact same answers.

This paper proposes Agentic Context Management (ACM), treating agent memory as a lifecycle problem with five primitives, and presents Maximem Synap, a reference implementation achieving strong benchmark results.

Grok Build adds features for token usage tracking, model configuration, prompt batching, and diagnostics to help developers improve token efficiency.

Google announced Gemini 3.6 Flash with improved coding efficiency and lower token costs, alongside Gemini 3.5 Flash Lite and a cybersecurity-focused model, while still developing Gemini 3.5 Pro and hinting at Gemini 4.

A user reports that using three local AI models (GLM 5.2, DeepSeek v4 Flash, Qwen 3.6 35B A3B) over 7 days with 500 million tokens can cover most business automation needs.