Today's token-cost story is about distinguishing fresh tokens from reusable ones. The latest inference-market data shows cache volume approaching fresh-input volume, live model indexes still span enormous price ranges, and current coding-agent traces show cache reads accounting for more than 90% of total token traffic in some sessions. The vocabulary arc from tokenmaxxing to tokenminimizing and token yield is becoming more precise: the economical system does not merely consume fewer tokens. It avoids repeatedly buying the same context while routing genuinely new work to the cheapest sufficient capability.
Top Developments (Last 24 Hours)
1What if most of your agent's tokens are not actually new?
Surplus Intelligence's September 5 marketplace snapshot reports 1.424 trillion fresh input tokens, 1.251 trillion cache tokens, and 20.14 billion output tokens across the trailing 28 days. Among eligible requests over the latest seven full days, the marketplace reports an 86.1% mean realized discount from direct-provider pricing. The dataset covers one inference market, but it shows how close reusable context has come to fresh context as an economic category.
Surplus Intelligence ↗2The inference index slips again while individual model prices swing wildly
IFX closed September 5 at 82.25, down 0.16% from its previous reading, with blended prices across its 29-model basket ranging from $0.06 to $11.25 per million tokens. Its 30-day data shows DeepSeek V3.1 open-hosted up 94.12% while DeepSeek V4 Pro open-hosted fell 58.39%, illustrating why model pricing has become too dynamic for a static annual rate-card assumption.
IFX ↗3Today's model menu runs from seven-and-a-half-cent input to ten-dollar input
AllAIModel's September 6 price index lists GLM-5.3-Flash at $0.075 per million input tokens and $0.25 per million output tokens, DeepSeek V4 Flash Vision Experimental at $0.14 and $0.28, Gemini 3.8 Flash at $0.75 and $3.75, and GPT-6 Astra and Claude Fable 5.1 at $10 and $50. That spread keeps modelmaxxing economically relevant because model choice can outweigh modest prompt-level savings.
AllAIModel ↗4Live agent traces show cache reads swallowing the token count
Tokens.ci's September 6 public trace for one Codex user records 118.6 million tokens at $92.51 across 1,038 messages. Of those tokens, 106.2 million were cache reads, versus 12.0 million fresh input and 343,800 output. Individual traces are not industry benchmarks, but they make the composition of modern agent token burn unusually visible.
Tokens.ci ↗From Tokenmaxxing to Selective Tokenminimizing
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 choosing the best-value model capable of completing each task. Token yield measures useful output relative to the resources consumed. Today's signal is that tokenminimizing increasingly means minimizing fresh computation rather than blindly minimizing context.
Momo Advisors
A September 5 analysis of Anthropic's new cache pricing argues that the 75% Fable 5.1 cache-read reduction only delivers its advertised economics when agents preserve stable prompt prefixes. Agents that continually rewrite instructions or reorder context can destroy cache reuse and leave much of the discount uncaptured.
Momo Advisors ↗Nutanix
Recent reporting says Nutanix spent $20 million on an internal AI cluster partly to manage token costs and expects the investment to pay back within a year. CEO Rajiv Ramaswami said the company expects open-weight models on its own infrastructure eventually to handle as much as 80% of internal needs, leaving frontier services for the remaining workloads.
SDxCentral ↗FinOps X
Flexera's FinOps X recap describes the progression from tokenmaxxing into a broader token panic as contexts and agent loops expanded. It highlights routing and governance as a tokenomics layer where systems can downgrade unnecessary premium-model calls while enforcing budgets and connecting usage to business applications.
Flexera ↗Anthropic
Anthropic's retrieval-based tool discovery remains a primary reference for tool-surface bloat. Instead of loading every available tool definition into context upfront, relevant schemas can be retrieved on demand. Anthropic reports an 85% token reduction in its evaluation while retaining access to the complete tool catalog.
Anthropic ↗DeepSeek
DeepSeek's pricing separately meters model tier, cache-hit input, cache-miss input, generated output, and peak versus off-peak execution. That makes model selection, context reuse, and scheduling independent cost-routing decisions rather than one headline cost per million tokens.
DeepSeek API Docs ↗Research Watch
CacheRouter: Tool Discovery Without Sacrificing Prompt Caching
CacheRouter identifies a collision between progressive tool disclosure and prompt caching. Changing the visible tool list saves schema tokens but can invalidate the reusable request prefix. Its dual-path design keeps a small stable core tool surface while discovering long-tail tools separately, reaching token-level cache-hit rates of 90.99% and 95.2% in prototype experiments.
Why it matters: Tokenminimizing can increase the bill when a smaller changing prompt destroys cheaper cached context. Tool-surface optimization therefore has to minimize resident schemas while preserving cache stability.
arXiv ↗ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents
ReCache independently caches recurring tool and skill representations so they remain reusable when resources appear in different combinations or orders. Across seven public tool and skill datasets, its resource-wise attention produced a 3.655 times time-to-first-token speedup while the complete system reduced allocated KV-tensor memory by 92.43%.
Why it matters: Tool-surface cost exists below visible prompt tokens. Preserving the computed representation of recurring capabilities can reduce repeated inference work without forcing the entire toolbox into fresh context.
arXiv ↗One Model, Many Minds: Replacing Multi-Agent Chatter with Mixture of Roles
This recent arXiv paper replaces repeated text exchanges among specialized agents with multiple learned roles combined inside one model invocation. Across its reported reasoning and personality evaluations, the approach matched multi-agent-system performance while reducing token cost by 20 times.
Why it matters: Agent-to-agent communication has its own token bill. Useful specialization does not necessarily require repeatedly serializing intermediate work into another model's context.
arXiv ↗Dual-Pool Token-Budget Routing for Cost-Efficient LLM Serving
This serving study routes requests into short-context and long-context infrastructure pools according to estimated total 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 fleet-scale annual savings of $2.86 million.
Why it matters: Token budgeting is also an infrastructure decision. A short request should not inherit the memory allocation and serving cost of the longest request the system might ever receive.
arXiv ↗Phrase of the Day
“Tokenminimizing”
Tokenminimizing is the practice of reducing unnecessary AI token consumption while preserving the quality and usefulness of the required result.
- AI adoption
- Tokenmaxxing
- Token shock
- Tokenminimizing
- Modelmaxxing
- Token discipline
- Token efficiency
- Token yield
The likely winners are systems that minimize fresh and unnecessary consumption without throwing away useful evidence or reusable context, combining routing, caching, selective tools, and explicit budgets.
- AI FinOps teams
- model routers
- AI gateways
- agent budget controls
- cache-aware inference systems
- retrieval-based tool loading
- task-level cost observability
Tokenminimizing is getting smarter: the goal is not to starve the model, just to stop buying it the same lunch twice.
The Next Web ↗