Today's token-cost story is starting to challenge the meter itself. Futurum Research says agentic AI can consume 10 to 100 times more tokens per task than simple inference and argues that predictable production workloads are pushing enterprises toward reserved infrastructure rather than pure per-token billing. Meanwhile, the Token Price Index fell 8.2% in one week to $2.13 per million blended tokens, an open inference marketplace logged 2.61 trillion fresh input tokens plus 2.10 trillion cache tokens over 28 days, and Stanford and NVIDIA researchers showed that some agent decisions do not need generative output tokens at all. The arc from tokenmaxxing through tokenminning to token yield is becoming architectural: first reduce waste, then route the work, then question whether an LLM should be generating tokens for that step in the first place.
Top Developments (Last 24 Hours)
1What happens when an agent turns one inference call into 100?
A Futurum Research report released September 25 and sponsored by QumulusAI says agentic AI can increase token consumption per task by 10 to 100 times compared with a simple inference call. Based partly on a survey of 824 AI decision-makers, the report says reserved and owned infrastructure already represent 66% of AI compute consumption and argues that sustained, predictable inference workloads can become better candidates for capacity-based economics than variable per-token billing.
Business Wire ↗2The Token Price Index drops 8.2% in a week
The September 25 Token Price Index fell to $2.13 per million blended tokens from $2.32 a week earlier, an 8.2% weekly decline and 9.1% decline over 30 days. The index is the geometric mean blended inference cost across 24 models from 11 providers. Its methodology also carries DeepSeek's off-peak rate because that rate applies for most of the week, another reminder that time-of-day pricing can now affect model-routing economics.
Token Price Index ↗3Cached context approaches the scale of fresh input in an open inference market
Surplus Intelligence's September 25 marketplace snapshot records 2.608 trillion fresh input tokens, 2.096 trillion cache tokens, and 45.06 billion output tokens across 49.6 million requests over 28 days. Its latest seven full days of eligible traffic show an 88% mean realized discount from direct-provider pricing. The marketplace is not a proxy for the entire inference industry, but cache volume equal to roughly 80% of fresh input makes cache economics a first-order part of the bill.
Surplus Intelligence ↗4Stanford and NVIDIA ask whether an agent decision needs output tokens at all
VentureBeat reports September 25 that Stanford and NVIDIA researchers released CLM-8B, a contrastive model for bounded decisions such as tool selection, ranking, and verification. Instead of generating an action name token by token, CLM encodes states and candidate actions and compares their representations. The team reports up to 9 times lower latency than Jev in its tests, although results vary by benchmark and the researchers did not publish an apples-to-apples dollar-cost comparison.
VentureBeat ↗From Tokenminning to Avoided Inference
Tokenmaxxing is the practice of maximizing AI consumption as a proxy for productivity. Tokenminning is the deliberate practice of reducing token consumption while preserving useful output quality, with token minimizing and tokenmining also appearing as variants in current usage. Modelmaxxing means matching work to the least expensive model tier that can perform it reliably. Token yield measures useful work relative to token consumption. Today's emerging step beyond that arc is avoided inference: moving deterministic or bounded operations out of generative token production altogether.
IFX
The IFX Inference Index closed September 25 at 82.80, down 0.18% on the day. Its capability-adjusted board identifies different cheapest models for frontier, capable, and budget intelligence thresholds, currently ranging from a $3.375 blended rate at its frontier gate to $0.4625 at its budget gate. That is modelmaxxing expressed as a market index: buy the capability threshold the task requires rather than the most capable model available.
IFX ↗Digital Applied
A September 25 pricing update documents substantial route-price drift even when vendor list prices do not move. Its observed GLM-5.3 OpenRouter route moved from $0.84 input and $2.64 output per million tokens on September 22 to $1.40 and $4.40 on September 25, while the GLM-5.3-Flash route fell to $0.045 and $0.14 against a $0.15 and $0.50 vendor list. AI gateways can therefore change effective model economics independently of the model maker.
Digital Applied ↗Tokenminning
The tokenminning vocabulary is being formalized into production controls rather than prompt-writing etiquette. Current guidance defines the discipline around metering, spend attribution, model routing, agent caps, and deployment enforcement. That is a useful maturation of the original tokenmaxxing backlash: token efficiency becomes an architectural property instead of asking individual users to remember to be frugal.
Tokenminning ↗The New Stack
Current coverage of skills versus MCP architecture focuses on the resident cost of exposing large tool catalogs to agents. The emerging pattern is a two-layer capability surface in which compact descriptions or skills remain visible while heavier tool definitions and execution interfaces are loaded only when required. The important tokenminimizing principle is broader than any one implementation: capability does not have to mean permanent context residency.
The New Stack ↗LSEG
LSEG's agent-economics guidance identifies two distinct MCP token costs: the instructions the model must read to select a tool and the payload it receives after calling one. Its connector design emphasizes clearer tool descriptions, precise retrieval, and compact responses so agents make fewer misdirected calls and carry less unnecessary data forward. Tool-surface optimization therefore applies on both sides of the call.
LSEG ↗Research Watch
Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
Growing Harness moves recurring agent control logic out of repeatedly reconstructed model context and into reusable executable code learned from task failures. Across BrowseComp-Plus and WebArena-Verified with deployment models from 4B to 120B parameters, the authors report 76.0% to 91.8% fewer LLM calls and 74.4% to 98.6% lower deployed-agent inference cost relative to a tool-calling agent.
Why it matters: This pushes tokenminning past context compression. If recurring reasoning can become tested program logic, the cheapest token is the one the system no longer needs to generate on the next task.
arXiv ↗Reducing Cost of LLM Agents with Trajectory Reduction
AgentDiet removes useless, redundant, and expired information from growing agent trajectories during execution. Evaluated on two models and two software-engineering benchmarks, the authors report 39.9% to 59.7% lower input-token consumption and 21.1% to 35.9% lower computational cost while maintaining comparable agent performance.
Why it matters: Long-running agents repeatedly rebill their own history. Trajectory reduction attacks that compounding input burden without requiring a cheaper model or a provider price cut.
arXiv ↗Total Cost of Agency: Exact Attribution of Memory Injection Cost in Multi-Agent LLM Workflows
This 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 the full billed cost and reached 27.6% of controllable variable cost at workflow depth six. Shrinking retrieval capacity from 32 entries to 2 reduced injected tokens 28.7% with accuracy movement within seed-level variation.
Why it matters: The invoice calls memory ordinary input tokens. Cost attribution reveals which part of that input came from architecture and can therefore be optimized independently of the model.
arXiv ↗AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows
AgentRouter routes individual steps within one agent trajectory among four model tiers instead of assigning the entire workflow to a frontier model. Trained on 50,000 annotated trajectory steps, the authors report 72% lower cost than frontier-only execution while retaining 97.3% of frontier-only quality, with less than 5 milliseconds of routing overhead per step on an A100.
Why it matters: Modelmaxxing becomes more precise when the routing unit is the step. Planning may deserve frontier reasoning while extraction, formatting, or verification may not.
arXiv ↗Phrase of the Day
“Avoided inference”
Avoided inference is the practice of removing an LLM call entirely when caching, deterministic code, retrieval, a specialized decision model, or another cheaper mechanism can perform the required operation.
- AI adoption
- Tokenmaxxing
- Token shock
- Tokenminning
- Modelmaxxing
- Token discipline
- Token yield
- Avoided inference
The likely winners are systems that optimize in the right order: eliminate unnecessary inference, retrieve only necessary context and tools, route the remaining work to the appropriate model, and reserve frontier generation for steps where it changes the outcome.
- deterministic preprocessing
- specialized decision models
- reusable executable agent logic
- retrieval-based tool loading
- cache-aware inference
- cost-aware model routing
- outcome-level AI FinOps
Tokenminning asks how to spend fewer tokens. Avoided inference asks whether this particular job needed a token meter at all.
VentureBeat ↗