Token Cost Radar

1How much are you paying before the agent even starts working?

The Stack reports September 16 that Uber found MCP tool-schema overhead reaching a 50,000 to 70,000-token baseline per session in parts of its AI coding environment. Uber moved toward CLI resolution, on-demand tool search, and code-mode batching, reportedly bringing per-session schema overhead close to zero while weekly coding-tool users rose sevenfold and agent requests rose 9.4-fold between February and August. Uber says total spend has remained roughly flat since April, with cost per 1,000 requests down about 34% and cost per session down about 52% from its peak.

2JPMorgan puts a $2,000 monthly ceiling on some engineers' AI spend

Business Insider reports September 17 that some JPMorgan engineers using Claude have a $2,000 monthly spending limit, with users able to request increases. JPMorgan declined to detail its token-limit strategy but said it is tying AI costs to measurable business value across outcomes, productivity, quality, speed, capacity creation, risk reduction, and business impact. The cap arrives alongside a sandboxed Devspace architecture intended to constrain agent access to internal systems.

3Huawei expects agents to consume more than 90% of AI tokens

Reuters reports that Huawei expects autonomous agents to account for more than 90% of global AI token traffic by 2035. Huawei says annual token consumption will surge as agents continuously perceive, reason, make decisions, and call external tools, putting substantially greater pressure on inference infrastructure than conventional request-response AI.

4Huawei turns KV cache into a storage tier

Huawei introduced OceanStor M900 Context Memory Storage on September 17, explicitly targeting the growing cost of KV cache in long-context and multi-turn inference. The system pools KV cache across on-chip memory, DRAM, and SSD storage at petabyte scale. Huawei says the architecture is designed to reduce redundant inference work and lower token costs as million-token contexts and agentic workloads push conventional memory capacity and economics.

SCOUT: Hybrid Semantic Tool Discovery for Enterprise MCP Gateway

SCOUT treats large MCP catalogs as a retrieval problem rather than placing every tool schema into every model context. In a production deployment at PayPal, the authors report reducing tool-schema consumption from 140.2 thousand tokens, or 70.1% of context, to 1.3 thousand tokens, or 0.8%, while indexing more than 2,000 tools across more than 200 MCP servers.

Why it matters: This is tokenminimizing at the tool surface. A capability can remain available without remaining resident in every prompt, allowing tool catalogs to scale without context cost growing linearly with them.

ContextPipe: Database-Inspired Context Assembly for Long-Horizon Agents

ContextPipe treats agent context construction as query planning under a hard context budget and byte-sensitive prompt cache. In its preliminary SWE-bench Pro evaluation, it reduced total token volume by 31%, LLM calls by 23%, and response time by 9% compared with append-only context construction, although KV-cache hit ratio declined.

Why it matters: The cache result matters more this week as KV reuse becomes an infrastructure layer of its own. Removing context can save fresh tokens while destroying reusable prefixes, so the economical context is not automatically the shortest one.

Overthinking Loops in Agents: A Structural Risk via MCP Tools

This study examines MCP tools that induce cyclic agent trajectories, causing models to repeatedly call tools, refine answers, and consume resources without an obvious individual failure. Across tested tool-capable models, the authors observed token amplification as high as 142.4 times.

Why it matters: Token budgets need a trajectory dimension. A perfectly reasonable per-call price can still produce an absurd bill when the agent architecture allows repeated calls to multiply unnoticed.

Can Your AI Agent Be Cheaper? Task Specifications Change Token Spend

Across 2,700 agentic coding runs, this study found that reducing a full task specification to a bare user story increased token spend by 29.7%. Sensitivity varied substantially by task, from 13% to 115%, showing that smaller prompts can create larger downstream bills when agents must explore or reconstruct missing requirements.

Why it matters: Token yield depends on useful context, not simply less context. Removing information that prevents wandering is tokenminimizing's false economy.