Tokenminning Issue #2

Tokenminning is a newsletter focused on “token engineering”, the discipline of using tokens efficiently in organizations. It is compiled by Rob May from Neurometric.ai, the world’s first token engineering platform.

Coinbase posted a summary of how they optimized token usage over the past quarter.

We route everything through our own gateway: a single endpoint and format for dozens of models, with cross-provider failover, redaction, logging, and cost controls all applied before anything reaches a vendor.

A researcher at NYU recently published the first paper we’ve seen on “AI Tokenomics”.

Tokens have become the practical accounting unit for modern foundation model services, linking information processing, computation, memory use, energy expenditure, pricing, and economic value. This paper develops a framework for AI tokenomics: the study of how tokens are generated, consumed, priced, allocated, and optimized across AI systems. We connect token-level technical costs to workflow-level production functions, enterprise resource allocation, measurement and instrumentation methods, and emerging market-design questions. The framework shows that token expenditure and economic value are distinct: value depends on marginal productivity, workflow position, hidden reasoning activity, risk, and downstream propagation effects. The paper concludes by identifying open research directions in hidden-token measurement, empirical calibration, token productivity, dynamic allocation, and token-based markets.

A Tokenomics Framework For Investors - link

Robonomics on Open Source Models - link

CIO Magazine on Tokenomics in the Enterprise - link

TokenMaxxing is Ending - a blog post from Sedai - link

Inference Optimization in LLMs - a Systems View - link

You build tasks specific SLMs - how do those models help manage inference spend?

Task-specific SLMs help manage inference spend because you stop paying for capability you never use. A frontier model is trained to write code, translate languages, answer trivia, reason through novel problems. You pay for all of that on every call, even when the job is just classifying a support ticket. Our models don't carry any of that. Each one is trained for a single task. On open-source public benchmarks, our classification model runs over 1,800x cheaper than OpenAI's average and over 5,000x cheaper than Anthropic's, on the same task, at higher accuracy. On the QMSum summarization benchmark, 10,000 summaries a day costs about $7.20 with ScaleDown versus $58 with GPT-4.1 Mini, with a quality difference human evaluators couldn't detect.

What's the biggest misperception people have about SLMs?

The biggest misperception is that smaller means worse.  For well-scoped production tasks, the right SLM is faster, cheaper, and more accurate than a frontier model, because it's built for one job instead of all of them. Published academic research shows a half-billion-parameter model hitting 91.7% accuracy on standard classification benchmarks while a 72-billion-parameter model scored 88.6%. Across our independently verified results on public benchmarks, FinanceBench for compression, QMSum for summarization, CLINC OOS for classification, and CUAD for extraction, ScaleDown models average 8-9% higher accuracy than frontier models from OpenAI, Anthropic, and Google, while running orders of magnitude cheaper. The accuracy edge is what makes the cost story real. It's not a trade-off.

How should a potential customer think about what workflows are best for a Scaledown solution?

The best-fit workflows are high-volume, narrowly defined, and currently expensive to run at scale. If a step in your pipeline is compressing long context before passing it to another model, summarizing documents or transcripts, pulling structured data out of unstructured text, or classifying inputs for routing or triage, that's exactly what we're built for. A good gut check: if inference cost is the reason something hasn't shipped, or your team is already looking at open-source models to cut spend, that workflow is almost certainly a fit. The biggest line in most AI budgets is paying frontier prices for work a small model handles better.

How does your pricing structure help lower inference costs?  What's a typical customer improvement?

The pricing is meant to be the floor, not another layer on top. Public API is $0.05 per million input tokens, flat. No tiers, no output token charges. Output tokens are free because most of our models are encoder-only, there's no generative step to bill for. The actual improvement depends on what you're replacing, but our public benchmark results give a real range: QMSum shows summarization running 14x cheaper than GPT-5.4, CLINC OOS shows classification at 113x cheaper than GPT-5.4 Mini at the same accuracy, and FinanceBench shows compression cutting a $10/day pipeline to $4/day at 60% compression. All of it is publicly available and fully reproducible. At millions of calls a month, that gap is often what decides whether something ships or gets cut.

What technology improvements do you see coming that you are most excited about because it will make your solution even more effective?

Three things I'm watching closely: small-model quality ceilings continuing to rise, distillation and fine-tuning tooling maturing, and routing infrastructure getting better. As quality ceilings go up, more workflow types become candidates for a task-specific approach. Tasks that still need a frontier model today for reliability will cross over. Better distillation tooling means we can build and iterate on new models faster with less labeled data. And routing is what makes the whole thing usable: a general model handling the ambiguous parts of a workflow while a fleet of cheap, fast, task-specific models runs the high-volume steps underneath. That architecture is already right. The infrastructure to make it seamless at scale is what's still catching up.

You can learn more at scaledown.ai

If your company would like to be featured in the Tokenminning newsletter, please reach out to [email protected]