A new Menlo Ventures report reveals that enterprise AI deployment spend surged from $33B to $177B in 2026, a 5.4x increase. Agent traffic surpassed human traffic on OpenRouter, accounting for 80% of token consumption and driving the growth. Anthropic, OpenAI, and Google captured 69% of the market, with Anthropic's share of enterprise API spend rising significantly to 55%.
Token Consumption
Token 消耗
The number of minimal semantic units used by LLMs for input/output, directly impacting cost
Curated Intelligence
Ant Digital released LingDT-Decision-1.0 (4B/27B), a model optimized for structured decision-making tasks, claiming over 90% reduction in token consumption and 10x faster response times compared to traditional LLMs in scenarios like customer service and risk control. Featuring native multi-choice and sequential decision capabilities, it outperforms competitors like Microsoft-Decision-1 on 23 benchmarks and is now available on the DTMaaS platform.
NetEase LobsterAI released its second 'China Office Agent User Behavior Report', revealing that average token consumption per task reached 202k in September, a 3.2x increase from March, with daily active user growth hitting record highs. The report highlights that while cache hit rates are generally high across models, varying caching discount strategies result in up to sevenfold differences in actual costs, making models like DeepSeek V4.1 Flash highly cost-effective.
In response to claims that over half of China's AI compute is spent on short dramas, ByteDance's Doubao PR head Liu Xing clarified that only about 10% of the company's total compute is used for video generation model training and inference. This follows data from Chinese Online suggesting 55% of AI token consumption goes to short videos and drama generation.
Akka empirically tested a spec-driven AI porting workflow across 65 open-source projects, finding that structured specifications significantly improved first-pass implementation quality. While 57 projects saw gains in code size or performance, the study revealed that using Opus saved ~40% tokens compared to Sonnet but doubled the time, indicating diminishing returns on increased computational effort.
Anthropic reported $16.8B in Q3 revenue, a 45% quarter-over-quarter increase that exceeds its total 2025 earnings but marks a sharp deceleration from Q2's 140% surge. The slowdown is attributed to seasonal demand dips and competitive pressure from OpenAI's new Astra model, which Anthropic countered with Opus 5.5. As the company prepares for a November IPO, it highlights profitable inference metrics to reassure investors amidst intensifying price wars and closing performance gaps from open-source models.
Anthropic has introduced dynamic workflows in Claude Managed Agents, enabling the parallel orchestration of up to 1,000 sub-agents per execution. A lead agent plans and distributes tasks, merging results to achieve significantly higher bug detection rates in large codebases compared to single-agent runs, though token efficiency remains a point of debate.
To address the high token consumption of deep iterations in Claude Design, a 'cloud-finalize, local-maintain' workflow is proposed. The core method involves exporting the finalized design as HTML from the web interface and subsequently using Claude Code to maintain it as a local static resource, thereby decoupling from the cloud tool and significantly reducing costs.
Users discovered that Claude's 'auto-compact' feature does not reduce context load as expected but instead defaults to loading the last 1 million tokens into the context. This mechanism causes rapid depletion of API usage limits (5-hour caps), sparking debates over billing logic and performance optimization.
Huawei-backed openJiuwen has released X-Router, a self-evolving model routing technology optimized for Ascend hardware. By dynamically assessing task complexity and assigning requests to the most suitable models, it reduces token consumption by over 50%. The key innovation lies in its ability to optimize routing strategies through feedback loops without manual rule maintenance.