The AI spending debate has acquired an unusually sharp contrast. Reuters reports that Apple is cutting employee AI usage budgets, while Revolut says self-hosting a Chinese open-weight model reduced its AI coding inference costs by roughly 90% and allowed engineers to stop rationing tokens. Menlo Ventures' new State of Enterprise AI report adds scale to the discussion: agents now account for approximately 80% of tokens flowing through OpenRouter, and the firm estimates annual enterprise AI deployment spending at $176.8 billion. Meanwhile, the open-source llm-d project is bringing production-grade routing, caching, and multi-tenant controls to distributed inference. The emerging question is not simply how to buy cheaper tokens. It is whether organizations should restrict consumption, reduce the cost of producing it, or measure the useful work those tokens accomplish.
Top Developments (Last 24 Hours)
1Why is Apple cutting employee AI budgets while other companies are removing token caps?
Reuters reported November 4, citing Bloomberg News, that Apple has reduced employee budgets for AI usage as part of a broader cost-control initiative. The report describes a tension between Apple's efforts to expand AI adoption and its decision to tighten spending allowances. Specific budget reductions, affected employee groups, and tool-level restrictions were not established in the available reporting. The development arrives as enterprises debate whether per-employee caps prevent waste or discourage valuable experimentation. Spending limits can provide predictable expenditure ceilings, but they do not establish whether the work being restricted would have generated more value than it cost. Organizations considering similar policies need visibility into usage, task outcomes, and the cost of alternative ways to complete the same work.
Reuters ↗2Menlo Ventures finds agents consuming 80% of OpenRouter tokens
Menlo Ventures' November 4 State of Enterprise AI report estimates that agents now account for approximately 80% of token consumption on OpenRouter, representing about 130.9 trillion tokens weekly and roughly seven times human-driven usage. The firm estimates annual enterprise AI deployment spending at $176.8 billion, up 5.4 times year over year. Its broader accounting distinguishes spending on deploying AI from expenditure on developing foundation models. Menlo also estimates that Anthropic, OpenAI, and Google capture 69% of enterprise AI deployment spending. These are Menlo's market estimates and observations from selected platforms, not a comprehensive census of global token consumption. OpenRouter traffic is particularly relevant to developers and agents, but its usage distribution should not be assumed to represent every enterprise deployment. The report's central implication is that autonomous software, rather than human chat alone, is becoming a major driver of inference demand.
Menlo Ventures ↗3Revolut says self-hosted GLM reduced AI coding inference costs by 90%
Reporting published November 3 and updated November 4 describes Revolut's transition from predominantly commercial coding models to self-hosted GLM models developed by China's Z.ai. Revolut infrastructure lead David Abitbol told an October AI Engineer conference that the company reduced AI coding inference expenditure by approximately 90%. Its deployment serves roughly 2,000 engineers using 56 NVIDIA B200 GPUs, with GLM-5.3 replacing GLM-5.1 as the default model. Abitbol cited approximately $0.30 per million tokens for the earlier configuration, around one-tenth the cost of the commercial services Revolut previously used. The company continues routing some workloads, including vision tasks, to other models. Revolut also reported improvements in coding and delivery times, although these are internal observations rather than controlled comparisons. The quoted inference economics depend on substantial infrastructure utilization and engineering expertise and should not be treated as a universally available API price. The most interesting operational consequence is that lower serving costs reportedly allowed Revolut to relax restrictive token budgets and expand AI usage beyond code generation into debugging and administrative development work.
The Register ↗4AI infrastructure inflation adds costs that token-price comparisons miss
VentureBeat reported November 4 that the AI infrastructure race is increasing expenses across several categories, including memory, reserved cloud GPUs, and enterprise software. The article examines how demand for computing infrastructure affects companies purchasing AI capacity as well as organizations deploying AI applications. These costs sit outside the advertised input and output token prices that dominate model comparisons. An enterprise may obtain lower API rates while simultaneously facing higher infrastructure reservations, software subscriptions, or hardware expenses. The report reinforces the need to distinguish the cost of producing inference from the price charged for consuming it. It does not establish that every organization's total AI bill is rising or that particular cost increases can be attributed exclusively to AI demand.
VentureBeat ↗From Token Rationing to Sustainable Inference Economics
Tokenmaxxing is the practice of maximizing AI token consumption, sometimes treating usage volume as evidence of productivity. Tokenminimizing is the practice of reducing unnecessary token consumption while preserving useful results. Token yield measures useful completed work relative to tokens consumed. Today's reporting illustrates the progression from tokenmaxxing through tokenminimizing toward token yield without establishing that any one approach is universally correct. Apple's reported spending restrictions and Revolut's decision to make tokens inexpensive enough for broader use represent different responses to the same financial problem. Menlo's findings suggest that agent consumption is growing rapidly enough to make both approaches consequential. No convincing new evidence of widespread adoption of the alternative spellings tokenminning or tokenmining emerged in today's coverage.
Menlo Ventures
Menlo's November 4 report finds that open-source models increased their share of enterprise token consumption from approximately 11% to 19%, while their share of spending declined from 6% to 4%. The firm attributes the divergence partly to falling prices, estimating that average open-source token prices declined 3.3 times to approximately $0.13 per million tokens. Chinese models account for an estimated 89% of open-source model spending in its analysis, led by Kimi, DeepSeek, GLM, and Qwen. The report also highlights the importance of cached tokens in agent workloads, noting that cache reads can account for 95% or more of tokens in some sessions. These are estimates derived from Menlo's market analysis and selected usage data, not universal prices or workload distributions. The findings show why token volume and token expenditure can move in different directions.
Menlo Ventures Research ↗llm-d
The open-source llm-d project's version 1.0 release is receiving renewed attention following November 4 coverage of its production-readiness improvements. The project combines distributed inference components around Kubernetes, vLLM, and SGLang. Its documented capabilities include prefix-cache-aware routing, load-aware scheduling, KV-cache offloading, disaggregated prefill and decode, and multi-tenant quality-of-service controls. Version 1.0 emphasizes stable configuration interfaces, high availability, zero-downtime operations, and agent-aware serving. These features can help infrastructure operators improve utilization and avoid redundant computation, but the release does not establish one universal percentage reduction in inference costs. Actual economics depend on hardware, workload composition, latency requirements, and operational overhead.
llm-d GitHub ↗Microsoft
Microsoft's engineering guidance on tool search remains relevant as enterprise agent catalogs expand. The July 29 analysis explains how agents can retrieve relevant tool definitions instead of loading every available schema into their context. The approach reduces initial tool-description overhead while preserving access to a larger capability catalog. Its effectiveness depends on reliable discovery, accurate tool descriptions, and the additional cost of retrieval operations. The article is not a November 5 announcement, but the architecture directly addresses the growing problem of tool-surface bloat. Organizations should measure complete task costs rather than assuming that reducing initial schema tokens automatically improves end-to-end efficiency.
Microsoft ↗Anthropic
Anthropic's established research on code execution with MCP explains how agent systems can reduce context consumption by loading tools on demand and processing intermediate results outside the language model's context. Instead of repeatedly sending large tool outputs back to the model, an execution environment can filter and transform information before returning only the required result. The technique can reduce unnecessary token consumption and improve control over sensitive intermediate data. Its benefits depend on secure execution, reliable tool access, and the overhead of the execution environment. The research provides architectural context for today's enterprise spending debate rather than evidence of a newly announced product change.
Anthropic ↗Research Watch
TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents
Published October 15, TokenEconomizer studies recurring token expenditure across sequences of realistic coding tasks rather than isolated benchmark questions. The researchers introduce an external controller that identifies useful operational instructions, removes persistently irrelevant tools, and intervenes when agents enter unproductive failure loops. Across evaluated configurations using Claude Code, Codex, and GitHub Copilot CLI, the system reduced average token costs by 18.4% to 30.9% while maintaining comparable task success. Additional experiments covering sequential web development and security auditing reported reductions of 6.4% to 21.2%. These are experimental findings, not guaranteed savings for arbitrary repositories or production deployments.
Why it matters: The paper connects token efficiency to agent behavior over time. Preventing repeated mistakes, unnecessary tool exposure, and unproductive execution loops can reduce expenditure without requiring changes to the underlying model.
arXiv ↗From Token-Max to Outcome-Max: How You Use AI Determines Its Productivity
Published October 5, this research examines the consequences of rewarding AI consumption rather than independently verified task outcomes. The authors distinguish token-max behavior, which encourages additional computation, from outcome-max behavior, which rewards successful work relative to its cost. Their simulation framework examines immediate interactions, longer-term behavioral adaptation, and organizational collaboration. Across these modeled settings, outcome-max improves efficiency while largely preserving verified task performance. The researchers also propose an incentive mechanism called OutcomeShare intended to align the interests of workers, organizations, and model providers. The results are theoretical and simulation-based rather than evidence from a controlled enterprise deployment.
Why it matters: The research provides a formal foundation for the vocabulary shift from maximizing consumption toward measuring useful work. It also explains why spending caps and unrestricted token allowances can both produce undesirable incentives when disconnected from outcomes.
arXiv ↗The Token Efficiency Index: A Peer-Benchmarked Composite Indicator for AI Token Efficiency
Published August 7, this paper proposes a standardized indicator for comparing AI token-spending efficiency across organizations. The Token Efficiency Index combines cache hit rate, cache amortization ratio, and premium-model usage into a score ranging from zero to 100. The authors describe normalization methods, peer comparisons, and a data-envelopment-analysis approach intended to identify improvement opportunities. The framework aims to make usage patterns more comparable across providers and organizations. It is a proposed measurement methodology, not an established industry standard or independently validated measure of business productivity.
Why it matters: Enterprises need better ways to compare token expenditure, but efficiency scores can miss differences in task difficulty and output quality. The proposed index offers a useful starting point for operational benchmarking while highlighting the need to connect consumption metrics to verified outcomes.
arXiv ↗Phrase of the Day
“Outcome-max”
Outcome-max is an approach to AI optimization that rewards independently verified task completion relative to cost rather than maximizing token consumption. The term appears in the October 2026 research paper From Token-Max to Outcome-Max, which examines how different incentives affect AI productivity.
- AI adoption
- Tokenmaxxing
- Token shock
- Tokenminimizing
- Token discipline
- Token yield
- Outcome-max
- Return on token spend
- Verified business value
Today's developments illustrate why outcome-max is a useful concept. Apple is reportedly tightening employee AI spending allowances, while Revolut says dramatically cheaper self-hosted inference allowed its engineers to use AI more freely. Neither policy is inherently optimal. A strict cap can prevent waste but may also block productive work. Cheap and abundant tokens can encourage experimentation but do not guarantee that the resulting work is useful. Menlo Ventures' estimate that agents account for approximately 80% of OpenRouter token consumption makes the distinction increasingly important. Autonomous agents can consume substantial resources without continuous human supervision, so organizations need mechanisms that connect expenditure to successful outcomes. The outcome-max research proposes rewarding verified work rather than raw usage. Its findings are based on simulations, but the underlying measurement principle is practical: define the task, establish acceptance criteria, measure the resources consumed, and compare the result with alternative ways of completing the work. Token efficiency remains important, particularly when unnecessary context, repeated tool calls, and failure loops inflate expenditure. However, the goal is not to minimize every token regardless of consequences. It is to eliminate computation that does not improve the probability or quality of successful completion. The strongest economic case belongs to systems that produce reliable results at a sustainable total cost.
- verified task completion
- quality-adjusted token efficiency
- cost per successful task
- selective context retrieval
- reduced failure loops
- appropriate model selection
- cache-aware inference
- bounded autonomous execution
- measured productivity
- transparent spending attribution
- sustainable infrastructure utilization
- business outcomes per dollar
The objective is not to win the token-eating contest. It is to finish the work.
arXiv ↗The jCodeMunch read
Today's research on coding-agent efficiency reinforces the value of reducing unnecessary context processing. TokenEconomizer demonstrates that pruning irrelevant tools and interrupting unproductive execution can lower consumption while preserving task success. jCodeMunch addresses a complementary source of overhead by reducing code-reading tokens through tree-sitter symbol retrieval and byte-precise context. As organizations reconsider token caps and inference infrastructure, efficient retrieval remains useful because it reduces the work required from whichever model ultimately handles the task.
See how the 94.5% per-lookup figure is measured → ← All editions