Token Cost Radar

1How did Uber grow agent requests 9.4 times without letting spend follow?

Uber's August 27 engineering report says weekly active users across its agentic offerings grew sevenfold from February to August and weekly agent requests grew 9.4 times, while total AI spend has remained relatively stable since April. Holding one model constant, Uber says cost per 1,000 model requests fell almost 34% from its peak and cost per session fell 52% from its June peak. Its cost program combines attribution, model routing, caching, prompt and tool controls, and on-demand MCP tool loading.

2Snowflake CEO: big token numbers are not the objective

Diginomica reports August 28 comments from Snowflake CEO Sridhar Ramaswamy arguing that enterprise AI should optimize value versus cost rather than tokenmaxxing. He says Snowflake supports per-user budgets, including limits such as $30 per user per month, while spreading workloads across proprietary and open models and using precomputation and other optimizations to control spend.

3Google gives agent budgets an actual stop button

Google Cloud has added new FinOps controls for Gemini Enterprise, including pay-as-you-go billing, pooled developer-tool quotas, Flexible Savings Plans that discount eligible token costs by 10% to 20%, and hard monthly project-level spend caps. The caps can halt additional agent API calls once the configured ceiling is reached, moving AI budgeting from reporting toward enforcement.

4Cheaper AI keeps producing bigger enterprise budgets

The Edge Malaysia's current tokenmaxxing analysis frames rising AI bills as a Jevons-paradox problem: lower unit costs encourage wider adoption, longer reasoning, autonomous loops, and heavier agent usage. It argues for formal AI FinOps, dynamic model selection, monthly token budgets, centralized procurement, and cost-per-outcome metrics rather than raw consumption.

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

Submitted August 25, Paritok-4B targets the large file reads and tool outputs that coding agents repeatedly resend to frontier models. Across all 300 SWE-bench Lite instances, it compressed agent context to 25.7% of its original size while retaining 86.5% of uncompressed single-shot solve quality. On line-numbered agent input, it retained 89.3% of solve quality at 27.8% context size.

Why it matters: The paper makes an important economic distinction: compression itself has a price. The authors found that using GPT-5 as the compressor was net-negative at list rates because the compressor cost more than the downstream tokens it saved.

Hybrid Semantic Tool Discovery for Enterprise MCP Gateway

Submitted August 25, SCOUT reframes large MCP catalogs as a retrieval problem. In a production deployment at PayPal spanning more than 2,000 indexed tools across more than 200 MCP servers, the system reduced MCP tool-token consumption from 140,200 tokens, or 70.1% of context, to 1,300 tokens, or 0.8%, a reported 99% reduction.

Why it matters: Prompt caching lowers the price of repeatedly processing a large tool catalog, but it does not free context capacity. Retrieval attacks both problems by keeping capability available while only injecting tools relevant to the current step.

Token Reduction Is Not Cost Reduction

This provider-billed study analyzed 2,848 coding-agent runs across 103 tasks, seven repositories, and three models. Prompt-cache traffic accounted for about 80% of the actual bill, and one compression arm that removed roughly 38% of estimated raw tool-output tokens increased paired billed cost by 6.8%. Aggressive compression also reduced successful patch application in a smaller SWE-bench-derived test.

Why it matters: Tokenminimizing needs a success-adjusted denominator. Removing context can trigger more retrieval, longer trajectories, cache changes, or lost evidence, so fewer tokens do not automatically mean a cheaper successful task.

Fail-Fast, Restart-Smart for Expensive Coding-Agent Trajectories

This arXiv paper trains a lightweight monitor to detect likely failure from partial coding-agent trajectories and restart before the full expensive path completes. On SWE-bench Verified, the monitor transferred across multiple policies and saved 14.6% to 20.4% of execution tokens at a targeted 5% false-positive rate.

Why it matters: A token budget does not have to wait until the budget is exhausted. Early failure detection treats the unproductive tail of a trajectory as something the runtime can identify and cut off.