Token Cost Radar

1How did Uber grow agent requests 9.4 times without growing the AI bill with them?

Uber Engineering reports that weekly agent requests grew 9.4 times from February to August while total AI spending remained relatively stable from April. Holding one model constant, cost per 1,000 model requests fell almost 34% from its peak and cost per session fell 52% from its June peak. Uber attributes the improvement to model routing, caching, prompt and tool optimization, attribution, and on-demand MCP tool loading.

2The average agent run is cheap. The expensive tail is not.

Metapress reports on an internal Revenium audit of 14,680 agent runs in which the top 1% generated 46% of total spending and the top 5% generated 77%. One unattended coding-agent session ran for four days, made 4,819 calls, and cost $3,762. Revenium's findings argue that average cost per run can hide the portion of agent spending most likely to break a budget.

3Token volume rises 47% while spending rises 7%

KuCoin reports that a JPMorgan research update found OpenRouter token volume increased 47% month over month while token spending increased 7%. The reported volume-weighted average price fell 28%, with lower-priced models such as GPT-5.6 Luna accounting for most of the incremental volume. If the trend holds, cheaper tokens are expanding consumption considerably faster than the bill.

4DeepSeek makes the weekend an inference discount

DeepSeek's current V4 pricing charges half its peak rate during off-peak periods, with the entire UTC weekend classified as off-peak. V4 Flash is listed at $0.22 per million input tokens and $0.66 per million output tokens off-peak, versus $0.44 and $1.32 at peak. For delay-tolerant evaluation, batch, and background workloads, scheduling has become a literal cost-routing decision.

CacheRouter: Tool Discovery Without Sacrificing Prompt Caching

CacheRouter identifies a conflict between progressive tool disclosure and prompt caching: changing the visible tool list can shrink the prompt while invalidating reusable prefixes. Its dual-path architecture keeps a small fixed tool surface in the main prompt and handles long-tail discovery separately. Prototype experiments reported token-level cache-hit rates of 90.99% and 95.2%, reducing modeled input cost to roughly 12% and 8% of a no-cache baseline under the DeepSeek pricing used by the authors.

Why it matters: Tokenminimizing and caching can work against each other. The cheapest prompt is not necessarily the cheapest execution if changing it destroys discounted reusable context.

SchemaRouter: Field-Aware Tool Routing for Efficient Agentic RAG

SchemaRouter chooses both which tools to call and which response fields to retrieve. On its 110-query benchmark, it used 227 retrieved-context tokens versus 2,066 for a fetch-everything baseline while maintaining comparable answer accuracy. A more aggressive minimum-field configuration lost accuracy.

Why it matters: The target is relevant context, not minimum context. Tool payloads can be narrowed dramatically, but removing evidence beyond the useful threshold turns token savings into lower token yield.

Paritok-4B: Intent-Conditioned Context Compression for Coding Agents

Paritok-4B compresses large file reads and tool outputs before they are passed to coding models. Across all 300 SWE-bench Lite instances, it reduced context to 25.7% of its original size while retaining 86.5% of uncompressed single-shot solve quality. The authors also found that using GPT-5 itself as the compressor was economically negative at list prices because compression cost more than the downstream tokens it saved.

Why it matters: Compression has its own invoice. A token optimization is only a cost optimization when the machinery performing it costs less than the inference it eliminates.

Resource-Aware Consensus for Multi-Agent AI

This recent arXiv paper models multi-agent communication topologies according to both consensus speed and token cost, then adds the remaining computational budget directly to the system state. Its framework derives conditions for reaching agreement without exhausting the token budget and evaluates adaptive switching among differently priced coordination patterns.

Why it matters: Multi-agent systems can consume tokens simply by talking among themselves. Budget-aware coordination makes inter-agent communication an explicit resource allocation problem rather than assuming more discussion is always better.