August 31, 2026
Today's token-cost story is moving past the rate card. Fresh analysis argues that enterprises should judge AI workflow by workflow rather than optimize token counts in isolation, Forrester is calling model routing part of an emerging AI consumption layer, and a new audit of OpenAI's coding economics illustrates why successful-task cost is attracting more attention than price per million tokens. The vocabulary arc from tokenmaxxing to tokenminimizing and token yield is settling on a practical middle ground: useful work per token.
IT Brief's August 31 analysis argues that organizations should measure AI workflow by workflow against decisions, margins, failure rates, customer outcomes, and other business effects rather than treating token consumption as the result. The framing pushes token economics toward attribution and return: cheaper execution matters, but only in relation to what the workflow actually changed.
IT Brief ↗Forrester's August 31 analysis of Stripe's planned OpenRouter acquisition argues that the economically important layer sits between applications and model providers, where requests can be routed among models and consumption can be measured and monetized. The larger signal is that model choice, token usage, billing, and orchestration are converging into infrastructure rather than remaining separate developer decisions.
Forrester ↗QUASA's August 31 audit examines OpenAI's claim that GPT-5.6 Terra completed successful Terminal-Bench 2.1 tasks in Kiro at roughly 82% lower cost. QUASA notes that the figure comes from vendor-run testing and that the public announcement does not provide the full baseline or methodology needed for independent reproduction. The metric itself is notable: cost is being attached to successful task completion rather than tokens consumed.
QUASA ↗Tokenmaxxing is the practice of maximizing AI token 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 useful output relative to the resources consumed. The emerging common denominator is useful work per token, which asks whether the consumption helped finish the job.
OpenAI explicitly describes GPT-5.6 as producing more useful work from each token and emphasizes performance per dollar rather than raw consumption. Its August 24 Kiro announcement says GPT-5.6 Terra completed successful Terminal-Bench 2.1 tasks at roughly 82% lower cost in joint OpenAI and AWS testing, while attributing part of the gain to structured requirements and task context.
OpenAI ↗The FinOps Foundation treats token economics as a value-management discipline spanning attribution, cost per inference, token efficiency, token yield, model right-sizing, caching, budgets, anomaly detection, showback, and chargeback. Its framework rejects both maximal and minimal token consumption as goals when neither is connected to useful outcomes.
FinOps Foundation ↗Anthropic's retrieval-based tool discovery remains a primary example of attacking tool-surface bloat. Instead of loading every available tool definition into context upfront, relevant schemas are retrieved when needed. Anthropic reports an 85% token reduction in its evaluation while retaining access to the complete tool library.
Anthropic ↗DeepSeek's V4 pricing separately meters cache-hit input, cache-miss input, generated output, model tier, and peak versus off-peak execution. V4 Flash cache-hit input can cost a small fraction of uncached input, while off-peak execution is half the peak rate, making context reuse and scheduling first-class cost-routing decisions.
DeepSeek API Docs ↗Gartner calls the current dynamic the inference paradox: better unit economics encourage more ambitious agent workflows, so inference cost per completed workflow can rise even while individual tokens get cheaper. Gartner predicts inference cost per agentic workflow will increase more than fivefold through 2028.
Gartner ↗Submitted August 25, AgentSpec targets the efficiency loss that speculative decoding experiences under large batches of agent traffic. It combines structure-isolated drafting, intended to avoid speculation down irrelevant agent paths, with redundancy-aware allocation of dynamically available token budgets and evaluates the design across five agent workloads and four model families in vLLM.
Why it matters: Token budgets exist below the application layer too. Allocating speculative work toward likely useful continuations can reduce wasted inference without asking the agent itself to produce a shorter answer.
arXiv ↗CacheRouter studies a conflict between progressive tool disclosure and prompt caching: changing the visible tool list reduces schema tokens but invalidates reusable prompt prefixes. Its dual-path design keeps a stable core tool surface in the main prompt while handling long-tail discovery separately, producing token-level cache-hit rates above 90% in prototype experiments.
Why it matters: Tokenminimizing can erase its own savings when a smaller changing prompt destroys a valuable cache. Tool-surface optimization therefore has to account for both resident schema size and prefix stability.
arXiv ↗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 preserved invocation performance while producing a 3.655 times time-to-first-token speedup, and the complete system reduced allocated KV-tensor memory by 92.43%.
Why it matters: Tool-surface economics extend beneath visible prompt tokens. Reusing the inference state associated with recurring capabilities can reduce the recurring cost of large agent toolboxes.
arXiv ↗Minima-KV keeps recent and protected KV-cache pages at higher precision while compressing older non-anchor pages without evicting them. In its reported Qwen3.6-27B configuration, deployment accounting showed 3.5 times compression relative to BF16 and 1.75 times relative to FP8, with small measured quality changes on LongBench v2.
Why it matters: Long-context inference has a memory bill as well as a token bill. Preserving useful historical state in cheaper formats can improve inference economics without forcing the application to discard context entirely.
arXiv ↗“Useful work per token”
Useful work per token is the amount of successfully completed, relevant work produced relative to the model consumption required to produce it.
The likely winners are systems that remove unnecessary model calls and context while preserving the evidence, tools, cache state, and reasoning required to finish the task correctly.
Useful work per token gives the meter one small inconvenience: it finally has to answer for the work.
OpenAI ↗Today's tool and cache research reinforces the distinction between removing context and selecting useful context. jCodeMunch reduces code-reading tokens through tree-sitter symbol retrieval and byte-precise context, keeping relevant code evidence available without making repository-scale context the default.
See how the 95%+ cut is measured →