Today's token-cost story has a useful twist: sometimes the best way to make AI cheaper is to stop using a generative model for work that doesn't require one. GitHub reports reducing secret-scanning serving costs by more than 98% after replacing remote LLM calls with a specialized classifier. Meanwhile, Reuters reports that Anthropic's quarterly revenue has reached $16.8 billion, raising fresh questions about whether extraordinary growth can support extraordinary infrastructure commitments. New research shows agents can reduce token consumption by moving internal reasoning out of text generation, or by executing routine tool sequences without consulting an LLM at every step. The tool-surface problem is getting renewed attention, too, as developers reconsider how many MCP schemas belong in active context. The progression from tokenmaxxing through tokenminimizing to token yield is becoming less philosophical and more measurable. The common thread is that cheaper tokens help, but avoiding unnecessary model execution can change the economics much more dramatically.
Top Developments (Last 24 Hours)
1Why is GitHub spending more than 98% less to scan secrets?
GitHub published October 13 and updated October 14 an engineering account of replacing remote LLM-based secret classification with a specialized ModernBERT classifier. The new pipeline uses deterministic regular expressions to identify potential credentials, then evaluates candidates using an 8-bit quantized classifier that considers file paths, secret types, and surrounding code. GitHub reports more than 98% lower serving costs and 86% lower median post-push scanning time compared with its previous LLM pipeline. Offline evaluation showed 13% higher precision and 19% higher recall, while a customer A/B test produced an 11.5% improvement in repository-owner-weighted alert acceptance. These are GitHub's measurements for a specific production workload, not general-purpose model comparisons. The larger lesson is that a narrowly trained model can outperform a generative LLM on both cost and quality when the task has a well-defined decision boundary.
GitHub Engineering ↗2Can Anthropic's revenue growth outrun its enormous AI infrastructure commitments?
Reuters reported October 14 that Anthropic told prospective investors its third-quarter revenue reached $16.8 billion, up 45% from $11.6 billion in the second quarter and $4.7 billion in the first. Two people attending the company's investor presentation also said Anthropic reported adjusted profitability for the quarter. Those figures were relayed by meeting attendees and are not independently audited results. Reuters notes that Anthropic has committed at least $518 billion over ten years to cloud services, computing, and infrastructure. The company is reportedly targeting a valuation above $2 trillion ahead of a potential IPO. The figures illustrate the scale of the economic challenge confronting frontier-model providers: revenue can grow rapidly while infrastructure commitments remain enormous. Adjusted profitability does not establish that the company's full capital and operating requirements have been covered.
Reuters ↗3Why do falling token prices still produce larger AI bills?
An October 14 TechTarget analysis examines the difference between cheaper inference and cheaper completed workflows. Author Eugina Jordan explains that enterprise agents frequently retrieve records, call tools, reconcile conflicting information, retry failed operations, and verify results before completing a request. Each step can generate additional input, output, and reasoning tokens. Her illustrative calculation shows that a 50% reduction in effective token prices combined with four times the token consumption doubles the token bill. The article recommends measuring cost per correctly completed task, including failed attempts and human correction, and enforcing spending, retry, and runtime limits. It also cites research finding that lower-priced reasoning models produced higher total inference costs in 32% of evaluated model-pair comparisons. The central recommendation is to budget for complete execution rather than individual model calls.
TechTarget ↗From Tokenmaxxing to Token Yield
Tokenmaxxing is the practice of maximizing AI token consumption as a proxy for productivity. Tokenminimizing is the practice of reducing unnecessary token consumption while preserving useful results. Token yield measures useful completed work relative to the tokens consumed. The progression from tokenmaxxing through tokenminimizing to token yield is increasingly visible in production engineering. Tokenminning remains an attested alternative spelling for the efficiency discipline, although today's reporting does not establish a new surge in its adoption. The stronger signal is architectural: developers are reducing unnecessary generation, selecting smaller models for bounded tasks, and keeping information outside active model context until it is actually needed.
Liquid AI
Liquid AI's October 13 release of LFM2.5-24B-A2B provides another example of inference efficiency through selective computation. The open-weight model contains 24 billion total parameters but activates approximately 2 billion per token. Liquid AI reports decoding speeds of 184 tokens per second on an Apple M5 Max and more than 21,000 output tokens per second at high concurrency on a single NVIDIA B300. Its context window extends to 131,000 tokens. The company reports competitive tool-use results against larger mixture-of-experts models, although those competitors retain advantages on several knowledge and coding benchmarks. These are vendor-reported measurements, not independently established cost-per-task savings. The development matters because locally served, sparsely activated models offer an alternative to purchasing cloud inference for every agent operation.
Liquid AI on Hugging Face ↗Zeeshan Mahmood
An October 14 engineering guide revisits the recurring token cost of large MCP tool manifests. The author cites a published example involving 255 tools across 50 servers that produced approximately 70,000 tokens of tool definitions. A compact representation containing only essential identifiers reportedly reduced that particular manifest to approximately 124 tokens. These are referenced benchmark figures, not independently reproduced measurements by the article's author, and they do not represent guaranteed savings across agent clients. The guide recommends keeping complete tool definitions in an authoritative store, exposing compact capability information initially, and retrieving detailed schemas when required. It also emphasizes validating actual token counts and testing whether agents can still discover and invoke tools correctly. The tradeoff is straightforward: smaller resident tool surfaces can reduce context consumption, but deferred discovery introduces additional retrieval and latency costs.
Zeeshan Mahmood ↗VDF AI
An October 14 enterprise pricing analysis highlights the distinction between per-seat licensing and token-based consumption. According to the article's review of Anthropic's published terms, Claude Enterprise costs $20 per seat per month on annual billing, with model usage charged separately at API rates. Claude Team includes usage up to plan limits, while Enterprise separates access fees from consumption. The analysis models annual costs for organizations with 250 and 1,000 seats, showing how developer activity and model selection can dominate the total. It also notes that administrators can apply user and organization spending caps. These are list-price calculations and illustrative workloads, not actual customer invoices. The broader lesson for AI FinOps is that seat counts alone cannot forecast spending when the underlying work is metered separately.
VDF AI ↗Json House
An October 14 update to Json House's weekly model-pricing comparison identifies a change in Claude Sonnet 5.5 cached-input pricing, from $0.20 to $0.10 per million tokens in its collected rate tables. The tracker reports that ordinary input and output prices remained unchanged across its 51 carried-over model entries. The distinction is important because cached-input discounts can alter the economics of long-running agents without changing their published standard token rates. The tracker also compares models from American, Chinese, and other providers, illustrating the widening range of available inference prices. These are the publisher's observations from collected public price lists, not evidence of an industry-wide price reduction on October 14. Production budgets should still use current provider terms and measured cache-hit rates.
Json House ↗Research Watch
Hybrid Reasoning Agents
Submitted October 13 and receiving fresh attention October 14, this paper from researchers affiliated with Stanford, UIUC, and Together AI investigates agents that alternate between internal latent reasoning and ordinary text generation. A lightweight HybridLink component allows reasoning to occur in continuous representations while preserving text-based tool calls and responses. Evaluations across 11 benchmarks using Qwen3-Base models with 4 billion, 8 billion, and 14 billion parameters report an average 10.4% performance improvement, 35.5% to 54.7% fewer tokens, and inference speeds 1.8 to 2.6 times faster than evaluated baselines. The authors estimate 2.7 to 3.7 times lower computational costs. Those estimates exclude some production expenses, including prompt prefill and external tool execution, so they should not be interpreted as measured API-billing reductions.
Why it matters: Conventional agents often express intermediate reasoning as generated text, paying decoding costs for steps that do not need to be communicated. Hybrid reasoning explores whether internal computation can remain compact while retaining explicit text for tool interactions and final answers. It offers a potential improvement in inference efficiency beyond prompt compression.
arXiv ↗ChunkAct: Action Chunking for Agentic Robotics
Submitted October 13, ChunkAct examines the cost of repeatedly invoking an LLM to supervise routine robotic actions. Conventional agentic robotics systems may ask a language model to reason over the entire interaction history before every perception, motion, or tool operation. ChunkAct instead groups established sequences of primitive actions into executable chunks. A single model call initiates a sequence, which proceeds autonomously unless a failure requires intervention. On the LIBERO-Pro benchmark, the researchers report a five-point improvement in task success compared with the previous agentic robotics state of the art while using 2.9 times fewer LLM calls, 4.4 times fewer input tokens, and 2.3 times fewer output tokens.
Why it matters: The principle extends beyond robotics: an agent need not consult a general-purpose reasoning model before every deterministic operation. Compiling routine sequences into executable procedures can reduce repeated context processing and reserve expensive reasoning for decisions that genuinely require it.
arXiv ↗Agent-Reflex: Teaching Language Agents to Act Reflectively
Submitted October 13, Agent-Reflex investigates how agents can recover from noisy tool responses and user corrections without repeatedly making the same mistakes. The framework combines reflection-skill discovery, supervised fine-tuning, and reinforcement learning to teach reusable recovery behavior. Built on Qwen3-8B, the resulting agent scores 58.6 overall across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, outperforming the evaluated reflection and environment-scaling baselines. The paper reports improvements in robustness but does not establish a universal percentage reduction in token consumption or production inference costs.
Why it matters: Retries and failed tool interactions contribute to token inflation even when individual model calls are inexpensive. Improving recovery behavior can reduce wasted execution, although the economic benefit must be measured against any additional reasoning or training overhead introduced by the reflection system.
arXiv ↗History Replays Itself: Unleashing Long-Context Capability in Linear Attention
Submitted October 11 and highlighted in October 14 research coverage, this paper introduces NoteBank, a sparse episodic memory system for linear-attention models. Instead of requiring a fixed recurrent state to preserve every detail of an expanding conversation, NoteBank archives selected evidence and replays relevant material through the model's existing computation. At one million tokens, the researchers report increasing Gated DeltaNet's mean accuracy across six RULER retrieval tasks from 1.7% to 70.3%. They estimate that the added retrieval process requires no more than 0.025% additional analytical FLOPs, while persistent context storage is approximately 2,600 times smaller than a matched transformer's full multi-head-attention KV cache. Results extend to additional recurrent architectures and long-context benchmarks.
Why it matters: Long-context efficiency increasingly depends on separating information that must remain immediately active from information that can be stored and retrieved later. NoteBank demonstrates that selective evidence retention can improve retrieval accuracy without requiring full attention over the entire history. The reported computational and storage advantages are architectural measurements, not commercial token-price discounts.
arXiv ↗Phrase of the Day
“Tool-schema tax”
Tool-schema tax is the recurring context and inference overhead created when an AI agent loads descriptions and argument schemas for tools it may never use.
- AI adoption
- Tokenmaxxing
- Token shock
- Tokenminimizing
- Model routing
- Token discipline
- Token yield
- Tool-schema tax
The October 14 discussion of MCP manifest overhead provides a timely reason to examine this cost separately from ordinary prompt and response tokens. An agent connected to numerous tools may receive extensive descriptions, argument definitions, and usage instructions before it begins solving the user's actual problem. That information occupies context whether the tools are used or not. The economic impact depends on the client, tokenizer, prompt caching, and how frequently definitions are reintroduced. The emerging engineering response is to measure the resident tool catalog, expose relevant capabilities selectively, and retrieve complete schemas when needed. The goal is not to make tools unavailable. It is to avoid making every possible tool part of every request.
- on-demand tool-schema loading
- compact capability directories
- task-scoped tool catalogs
- progressive tool disclosure
- tool-schema token accounting
- cache-aware agent initialization
- permission-filtered tool exposure
- retrieval-based capability discovery
- end-to-end tool-selection evaluation
Having 255 tools available is useful. Making the model read 255 instruction manuals before using one is a different proposition.
Zeeshan Mahmood ↗