Today's token-cost story is shifting from price to throughput economics. Public inference traffic is now running at roughly 17 trillion tokens per day on one major routing dataset, while a separate market index estimates $14.9 million in daily notional inference spend and says that figure is up 12.2% week over week. StepFun is adding another million-token Chinese model at $1 input and $2.70 output per million tokens, and open-model serving prices are reaching pennies per million. The vocabulary arc from tokenmaxxing through tokenminimizing to token yield now has a scale problem attached: when intelligence becomes cheap enough to consume by the trillions, efficiency has to be measured in useful work rather than tokens avoided.
Top Developments (Last 24 Hours)
1What does token discipline look like at 17 trillion tokens a day?
Agent Economy's September 20 snapshot, using public OpenRouter ranking data, estimates 509.9 trillion tokens of inference over the trailing 30 days, or about 17 trillion tokens per day. The latest measured day, September 19, reached 17.82 trillion tokens. The dataset represents routed public-model traffic rather than the entire AI market, but it illustrates why small improvements in routing, caching, and context efficiency can become economically significant at current inference volumes.
Agent Economy ↗2Public inference notional reaches $14.9 million a day
The Grid Intelligence Indices September 20 print estimates $14.9 million in daily notional spend across its OpenRouter top-50 methodology, up 12.2% week over week. Its basket unit cost rose 31.8% to $7.72 per million tokens while routed token volume increased only 1.9% to 16.18 trillion tokens per day. The divergence is a useful reminder that model mix can move the AI bill even when aggregate token volume barely changes.
Grid Intelligence Indices ↗3StepFun prices a million-token agentic model at $1 input
Current September 20 reporting on StepFun's Step 5 Preview lists API pricing at $1 per million input tokens and $2.70 per million output tokens, with a 1 million-token context window and a 95% cached-input discount. The sparse model activates 27 billion parameters from 600 billion total. Open weights are scheduled for October 15, adding another non-US model to the increasingly crowded price-versus-capability routing surface.
Unrot ↗4The inference price index holds at 83.46
IFX's September 20 close was unchanged at 83.46. Its 29-model basket spans blended prices from $0.06 to $11.25 per million tokens, with an average of $2.54. Its capability-adjusted board names different cheapest models at frontier, capable, and budget thresholds, while DeepSeek V4 Pro on open hosts is up 33.75% over 30 days. The spread reinforces that modelmaxxing is increasingly about buying the required capability rather than simply choosing the lowest nominal rate.
IFX ↗From Tokenmaxxing to Token Yield at Scale
Tokenmaxxing is the practice of maximizing AI consumption as a proxy for productivity. Tokenminimizing is the practice of removing avoidable consumption while preserving the required result. Modelmaxxing means selecting the best-value model capable of completing each task. Token yield measures the useful output or completed work produced relative to token consumption. Today's inference-volume numbers give that arc a new dimension: optimization that looks trivial per request can become material when multiplied across trillions of daily tokens.
FlexAI pricing tracker
A September 20 pricing snapshot for FlexAI's Token Factory lists 28 open models, with the cheapest input rate at $0.010 per million tokens and cheapest output at $0.030. Cached-input discounts are typically about 85%, and supported contexts reach 1 million tokens. Open-model inference is pushing the raw price floor downward while making workload placement, cache behavior, and capability requirements increasingly important to meaningful cost comparisons.
Price Per Token ↗Silicon Analysts
The September 20 weekly pricing ledger tracks current public list rates across major model providers and flags changes rather than assuming rate cards are static. Its current table spans from $0.20 per million input tokens for GPT-5.6 Luna to $10 for Claude Fable 5, with output rates ranging from $1.20 to $50. The practical FinOps lesson is that model pricing itself is now a moving input to routing policy.
Silicon Analysts ↗DeepSeek pricing
DeepSeek V4.1-Flash remains one of the clearest examples of cache-aware pricing. Current rates list off-peak cache-hit input at $0.003 per million tokens, fresh input at $0.15, and output at $0.60. Peak rates double those figures. A cache hit is therefore 50 times cheaper than fresh input off-peak, meaning stable prefixes can have a larger economic effect than many prompt-trimming exercises.
DeepSeek pricing analysis ↗Tokscale
A public coding-agent ledger updated September 20 recorded 135.2 million tokens for the day at an estimated cost of $99.14, with 130.8 million tokens classified as cache reads and only 3.8 million as fresh input. Across the account's longer history, cache reads represent 96.5% of 273.1 billion tokens. One workload is not a benchmark, but it demonstrates why raw token totals can badly misrepresent cost when cached and fresh tokens are priced differently.
Tokscale ↗ToolRouter
Recent MCP measurements found 47 tool definitions occupying 9,286 context tokens, about 198 tokens per tool with the tokenizer used in the test. Representing a larger catalog as 1,345 individual tool definitions would consume about 257,000 tokens. The tool-surface lesson remains unusually clean: a capability can be available to an agent without its complete schema being resident in every request.
ToolRouter ↗Research Watch
ContextPipe: Database-Inspired Context Assembly for Long-Horizon Agents
ContextPipe treats agent context assembly as query planning under a hard context budget and byte-sensitive prompt cache. In a preliminary SWE-bench Pro evaluation, it reduced total token volume by 31%, LLM calls by 23%, and response time by 9% versus append-only context construction, although KV-cache hit ratio declined.
Why it matters: Tokenminimizing is not simply deleting context. A smaller prompt can save fresh tokens while damaging cache reuse, so the economical context is the one that minimizes total execution cost rather than token count alone.
arXiv ↗Token Budgets: An Empirical Catalog of 63 LLM-Agent Budget-Overrun Incidents
This study catalogs 63 confirmed production budget-overrun incidents across 21 agent orchestration frameworks and groups them into eight failure classes. Its mitigation case study uses affine ownership in Rust to make budget cloning, double spending, and post-delegation reuse compile-time errors. Across its reported multi-agent tests, the typed approach prevented cap violations that occurred in an unconstrained delegation pattern.
Why it matters: Agent token budgets are becoming a reliability primitive rather than a dashboard metric. A spending cap that can be accidentally duplicated during delegation is not really a cap.
arXiv ↗Harness Engineering for Predictable Agentic Systems
This empirical study tests deterministic execution constraints including forced tool selection, bounded retries, output validation, and structured planning. After adding structured planning, three of four model-task cells reached perfect reported reproducibility and task success reached 100% in three cells. Token cost fell in every tested cell, although latency effects differed substantially by model.
Why it matters: Bounded execution can improve token yield by reducing wandering and retries, but the latency split is a warning against assuming that every cost control improves every operational metric.
arXiv ↗Dual-Pool Token-Budget Routing for Cost-Efficient and Reliable LLM Serving
This serving study routes short-context and long-context requests into separately configured inference pools according to estimated token budget. On Azure and LMSYS traces serving Llama-3-70B on A100 GPUs, the authors report 31% to 42% fewer GPU-hours, 5.4 times fewer preemptions, and projected annual fleet savings of $2.86 million.
Why it matters: Token budgeting reaches below the API meter. Expected context size changes KV-cache allocation and concurrency, making token footprint a useful infrastructure-routing signal as well as a billing quantity.
arXiv ↗Phrase of the Day
“Token yield”
Token yield is the amount of useful output, successful work, or business value produced relative to the tokens consumed to produce it.
- AI adoption
- Tokenmaxxing
- Token shock
- Tokenminimizing
- Modelmaxxing
- Token discipline
- Return on Token
- Token yield
The likely winners are organizations that can increase useful work faster than token consumption grows, using model routing, cache reuse, selective context, bounded agent execution, smaller tool surfaces, and explicit workload budgets.
- AI FinOps teams
- cost-aware model routers
- AI gateways
- agent token budgets
- cache-aware inference
- retrieval-based tool loading
- cost-per-completed-task observability
At trillions of tokens a day, saving tokens is arithmetic. Getting more useful work from them is economics.
Agent Economy ↗