Token Cost Radar

1What if the token price stays flat but the task gets 30% cheaper?

Reuters reports that Anthropic released Claude Sonnet 5.5 on September 28 at the same $2 per million input tokens and $10 per million output tokens as Sonnet 5. Anthropic says the new model needs fewer tokens to complete the same work. That distinction matters for AI budgeting: a model can lower cost per completed task without changing its nominal token rate.

2Agentic AI puts the per-token meter under pressure

A September 28 SiliconANGLE guest column examines a new Futurum Research report, sponsored by QumulusAI, which estimates that agentic AI can consume 10 to 100 times more tokens per task than simple inference. The report forecasts agent and reasoning inference growing 219% this year and argues that sustained production workloads can make purely variable per-token pricing difficult to predict. The figures are report estimates, not independent measurements, but the underlying FinOps problem is increasingly visible.

3Chinese models process 62.22 trillion tokens in one week

China Daily reports September 29, citing OpenRouter data, that Chinese models processed 62.22 trillion tokens from September 21 through September 27 versus 14.2 trillion for U.S. models. DeepSeek V4.1 Flash led with 19.6 trillion tokens, up 24% week over week, while GLM 5.3 Flash processed 16.3 trillion. The report attributes the shift partly to competitive pricing, open weights, rapid iteration, and price-performance for agent and tool-calling workloads.

4Corporate America routes more work toward cheaper open-weight models

The Financial Times reports September 28 that U.S. companies are increasingly using lower-cost open-weight models as AI infrastructure bills rise. The report says open-weight models have exceeded half of observed usage on some AI platforms, with DeepSeek, Zhipu, Mistral, and Nvidia among the model makers benefiting. The trend is modelmaxxing in practice: reserve premium closed models for workloads where their additional capability justifies the premium.

Total Cost of Agency: Exact Attribution of Memory Injection Cost in Multi-Agent LLM Workflows

This September study decomposes multi-agent cost into base prompt, inference, memory injection, miss penalty, and context accumulation. Across a 200-task enterprise benchmark using real model APIs, memory injection represented about 12% of full billed cost and reached 27.6% of controllable variable cost at workflow depth six. Reducing retrieval capacity from 32 entries to 2 cut injected tokens by 28.7% with accuracy movement within seed-level variation.

Why it matters: Tokenminimizing gets more actionable when the bill is attributed to architectural causes. Memory tokens look like ordinary input to the provider, but the application controls how much memory gets injected.

An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents

This empirical study separates tool-schema filtering, content compression, and history summarization as independent token-saving levers. Tool-schema filtering removed roughly 21,000 to 57,000 tokens from a typical turn and produced an immediate linear saving, while compressed file reads accumulated savings across later turns. The authors caution that single-shot compression quality does not by itself prove lower multi-turn agent cost.

Why it matters: The tool-surface tax compounds because schemas are repeatedly resident. Removing unused definitions can be more predictable than trying to squeeze every file read or conversation turn.

Toollery: Scaling LLM Agents to Thousands of Skills and Tools

Toollery treats large capability catalogs as a retrieval problem rather than placing every skill and tool specification into the model context. Evaluations span a roughly 79,000-capability benchmark, BFCL-V4 with more than 440 tools, and 3,396 proprietary requests across 220 tools. The framework retrieves a compact candidate set before final model selection and reports improved recall over ordinary specification retrieval in its evaluated settings.

Why it matters: On-demand tool loading attacks two costs at once: fewer schema tokens and fewer distractors. The model sees the small set of capabilities relevant to the current request instead of paying attention to the entire warehouse.

Token-Budget-Aware Pool Routing for Cost-Efficient LLM Inference

This serving study estimates total token demand before dispatch and routes short and long workloads into differently configured inference pools. On Azure and LMSYS traces, the authors report 17% to 39% fewer required GPU instances, corresponding to roughly $1.2 million to $2 million in annual savings at 1,000 requests per second. A Qwen3-235B case study projects $15.4 million in annual savings at 10,000 requests per second.

Why it matters: Token budgets can control infrastructure as well as prompts. Expected context footprint determines KV-cache demand and concurrency, making routing by token budget a direct lever on cost-to-serve.