Token Cost Radar

1Can you cut the price of repeated agent context by 90%?

DigitalOcean announced October 23 that prompt caching is generally available across supported models in its Serverless Inference catalog. Cached input receives discounts of up to 90%, while the company reports reductions in time to first token of up to 80% in its testing. The important engineering improvement is cache-aware routing. Repeated requests are directed toward replicas already holding relevant prompt prefixes instead of being distributed without regard to cache state. DigitalOcean reports that this increased cache-hit rates from approximately 35% to more than 90% in some configurations. GLM-5.3 improved from single-digit cache-hit rates to 88% to 95%, while Kimi K3 achieved approximately 84% to more than 92% prefix reuse in evaluated workloads. The company also introduced optional cache-isolation keys, expanded context windows, and dashboards showing cached tokens and estimated savings. DigitalOcean says more than 90% of input tokens submitted by agents on its platform have already been processed in previous requests. That figure describes its own traffic, not the entire industry. The discounts apply to eligible cached input rather than the complete inference bill, and actual savings depend on model selection, workload structure, and cache reuse.

2Frontier token prices fall again, but the annual trend tells a different story

The Token Price Index published its October 23 reading at $1.97 per million blended tokens, down from $2.05 on October 16 and $2.13 on October 2. The index reports a 7.4% decline over the preceding 30 days but a 67.3% increase over 12 months. Its methodology calculates a geometric mean of blended inference prices across 26 selected models from 11 providers. These results differ from transaction-weighted spending indexes that track what customers actually purchase. An index of selected frontier-model list prices can rise over a year even while customers reduce their effective costs by switching models, increasing cache reuse, or negotiating discounts. The changing composition of the index also complicates comparisons across longer periods. DeepSeek's time-dependent pricing introduces another qualification: the index uses its off-peak rates rather than modeling every possible execution hour. The latest reading illustrates why claims about falling token prices require a precise definition of the market being measured. Published model prices, actual customer expenditure, and cost per successfully completed task are different economic quantities.

3Kochava moves AI cost control into marketing workflows

Kochava announced October 22 that its StationOne Enterprise platform now includes expanded token-efficiency and spending-governance capabilities. Its Auto Router selects models according to task complexity and enterprise policy. Administrators can assign token-based or dollar-based budgets to individual users and teams, configure overrides, and receive alerts as spending approaches established thresholds. The platform also links token consumption to specific business tasks rather than reporting only conversation-level usage. Particularly relevant is its emphasis on deterministic execution. A new Audience Management Plugin performs operations including deduplication, segmentation, and data reduction through conventional code instead of requiring a language model to process the underlying datasets. Another plugin handles selected calibrated decision and scoring operations outside full-model inference. StationOne also supports compatible models running locally on customer hardware. Kochava has not published independent benchmarks establishing the resulting cost savings. The development nevertheless illustrates an important direction in enterprise AI architecture: use language models for tasks requiring language-model capabilities, and use simpler execution methods when they are sufficient.

What Does a Harness Buy? Tokens, Mostly

Originally submitted October 3 and receiving renewed attention through October 23 reporting, this paper isolates the effect of coding-agent harnesses on task success and inference expenditure. The researchers compared Claude Code, OpenCode, and mini-SWE-agent across five language models using SWE-bench Verified tasks. On a 447-task evaluation involving two models, the most elaborate and simplest harnesses were statistically equivalent within five percentage points. On a separate 45-task difficult subset, switching harnesses changed which tasks passed approximately as often as rerunning the same configuration, with a median disagreement rate of 13%. Despite similar performance, inference costs varied by up to threefold. Initial system prompts and tool schemas were major contributors, with substantially different fixed-context requirements across harnesses. The authors also found that prompt-cache pricing changed absolute costs without eliminating the underlying cost differences.

Why it matters: The paper provides direct evidence that agent infrastructure can consume substantially more tokens without producing proportionately better results. It also demonstrates why one-run benchmark comparisons can be misleading. A more expensive harness may provide useful capabilities, but its additional context and execution overhead should be justified through repeatable improvements in actual task performance.

TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents

Submitted October 15, TokenEconomizer investigates reducing coding-agent token consumption across sequences of related repository tasks rather than optimizing each task independently. The researchers introduce an external controller that learns recurring characteristics of an execution environment and adjusts how agents operate. Its interventions include supplying concise operational instructions, removing persistently irrelevant tools, and interrupting unproductive failure loops. Across evaluated configurations involving Claude Code, Codex, and GitHub Copilot CLI, the authors report average token-cost reductions of 18.4% to 30.9% while maintaining comparable task success. Additional evaluations involving web development and security auditing produced reductions of 6.4% to 21.2%, depending on workload structure. These are research measurements from the tested environments, not guaranteed reductions in commercial customer bills.

Why it matters: Agent inefficiencies often recur across tasks in the same environment. A controller that learns which tools are unnecessary and which execution patterns waste tokens can reduce consumption without changing the underlying model. The findings complement the newly prominent harness research by showing that context and execution policies can be optimized independently of model capabilities.

Rethinking Sequence Compression with Test-time Regression

Submitted October 22, this paper introduces TTR-Compressor, a method for compressing token sequences using regression coefficients computed during inference. Rather than relying exclusively on conventional pooling or learned projection layers, the researchers formulate sequence compression as a regression problem whose parameters summarize the input. The method is trained end to end, while the compressed representation is calculated efficiently during the forward pass. Evaluations cover block-wise sparse attention and visual-token merging in multimodal language models. In sparse-attention experiments, the researchers report substantial improvements on RULER compared with evaluated attention-based compressors. In multimodal experiments, the method matched the performance of a conventional MLP-based visual-token merger while using 55.6% fewer visual tokens. The paper does not establish corresponding reductions in commercial API prices or complete inference costs.

Why it matters: Reducing token counts is valuable only when the compressed representation preserves information needed for the task. This research demonstrates a possible way to represent sequences more compactly without proportionately sacrificing performance. Its direct relevance is architectural efficiency rather than immediate savings available through existing commercial model APIs.

Your Token Is Gone: A Token-Consumption Countermeasure for Disrupting LLM-Based Scam Systems

Submitted October 22, this paper examines token expenditure from an unusual perspective: using the resource requirements of malicious AI systems as a defensive opportunity. The researchers introduce TokenTrap, a system intended to redirect suspected automated scam agents toward non-sensitive, low-value computational activity while withholding exploitable information. The work evaluates different agent coordination architectures and uses explicit limits on execution time, rounds, and resource consumption. Across the tested 50-round interactions, the authors report large increases in the target systems' token consumption relative to their baseline configurations. Their architecture classifier achieved 85.67% accuracy across 300 labeled prompts. These results come from controlled experiments and do not establish real-world effectiveness against criminal operations. Deliberately increasing another system's resource consumption also raises questions about authorization, collateral effects, and appropriate defensive boundaries.

Why it matters: The study illustrates that token consumption is not merely a billing metric. It can also become a resource-exhaustion vulnerability when agents accept untrusted tasks and perform repeated reasoning or coordination without effective limits. For legitimate agent operators, the defensive lesson is to enforce bounded execution, evaluate tool requests, and avoid allowing untrusted inputs to trigger uncontrolled inference expenditure.