Token Cost Radar

1Why are AI infrastructure bills rising while token optimization improves?

IDC reported November 6 that worldwide AI infrastructure spending reached $121.4 billion in the second quarter of 2026, an increase of 53% year over year and 34% from the previous quarter. The research firm raised its full-year 2026 forecast from $497 billion to $521 billion. An important detail is that non-accelerated, CPU-based infrastructure grew 61% year over year, compared with 52% growth for accelerated systems. IDC attributes the broader expansion partly to agent orchestration, data processing, component-price increases, and continued hyperscaler investment. These figures describe infrastructure expenditure rather than token consumption or inference revenue. The distinction matters because improving model efficiency does not automatically reduce total capital spending when deployment volume, hardware prices, and supporting infrastructure requirements are simultaneously increasing.

2Anthropic's reported $3.4 billion power deal exposes an overlooked inference cost

The Financial Times reported November 7 that Anthropic agreed to take over approximately $3.4 billion in commitments for 1.7 gigawatts of gas-fired power-generation equipment originally ordered by Texas-based OnePWR Solutions. According to the report, the arrangement gives Anthropic access to Caterpillar generators and associated engineering and construction capabilities as it expands data-center capacity. The agreement was reportedly reached in late September, making this newly reported infrastructure financing rather than a transaction completed within the past 24 hours. The significance for inference economics is the increasing importance of electricity availability, generation equipment, and deployment lead times. These expenses may not appear in advertised API token prices, but they influence the capital requirements and long-term economics of producing inference at scale. The reported commitments should not be interpreted as an immediate operating expense or a measured increase in Anthropic's per-token production costs.

3Token Price Index reaches $1.87 per million tokens in November 6 update

The Token Price Index's November 6 update reports a blended frontier-model inference price of $1.87 per million tokens across 26 models from 12 providers. Its historical series lists $1.89 on October 30, $1.97 on October 23, and $2.05 on October 16. The index reports a 3.6% decline over the preceding 30 days but a 58.2% increase over twelve months. The index uses a geometric mean of blended model prices, so its movement reflects the models included and their relative pricing rather than the exact invoice paid by a particular customer. Changes in the composition of the model basket can also affect comparisons over time. The latest reading complicates the familiar assumption that every measure of AI token pricing must continuously decline. Economy-tier inference can become cheaper while the price of a changing basket of frontier capabilities moves differently.

4How much do cached tokens change the economics of Kimi's latest model?

A November 7 Tech Insider pricing analysis examines Moonshot AI's Kimi K3.1 Pro and the distinction between API consumption, paid memberships, and self-hosted deployment. The article reports advertised API rates of $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens. At those rates, cached input is priced at one-tenth of ordinary input, although realized savings depend on the proportion of requests eligible for caching. The analysis also compares Kimi with lower-priced Chinese alternatives, including DeepSeek, and notes that owning model-serving infrastructure introduces separate hardware and operational costs. These are reported pricing figures rather than independently verified production invoices. The useful lesson is that a single advertised price per million tokens can conceal substantial differences between prompt processing, generated output, repeated context, and infrastructure ownership.

How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks

This 2026 study examines token consumption across eight frontier language models performing SWE-bench Verified coding tasks. The researchers report that agentic coding can consume approximately 1,000 times as many tokens as simpler code reasoning and chat workloads. Repeated executions of the same task sometimes differed by as much as 30 times in token consumption, while greater expenditure did not consistently improve accuracy. The study also found that models were poor predictors of their own eventual token requirements, with reported prediction correlations reaching only 0.39. These findings describe the evaluated benchmark and model configurations rather than every production coding workload.

Why it matters: The results challenge two common budgeting assumptions: that task difficulty reliably predicts inference expenditure and that spending more tokens necessarily improves completion rates. Agent budgets need runtime monitoring, bounded execution, and empirical measurements rather than relying exclusively on model-generated cost estimates.

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

Published October 15, TokenEconomizer investigates recurring token expenditure across sequences of realistic coding tasks. Its external controller selects useful operational instructions, removes persistently irrelevant tools, and intervenes when agents enter unproductive failure loops. Across evaluated coding-agent configurations, the researchers report average token-cost reductions of 18.4% to 30.9% while maintaining comparable task success. Additional experiments involving sequential web development and security auditing report reductions ranging from 6.4% to 21.2%. These are experimental findings, not guaranteed savings for arbitrary repositories or production deployments.

Why it matters: The research shows that agent behavior and tool exposure can create avoidable expenditure independently of the underlying model's advertised token price. It also provides a concrete example of reducing consumption while preserving successful work.

Tiered Super-Moore's Law: Price Evolution, Production Frontiers, and Market Competition in Large Language Model Inference Services

This March 2026 economic analysis combines OpenRouter pricing data for 318 models, Epoch AI records covering 3,237 models, and 62 cross-validated historical observations. The author reports an approximately 600-fold decline in token prices across the studied period but finds substantial differences between market segments. Economy-tier models exhibit an estimated price half-life of 1.10 years and mid-tier models 1.55 years, while flagship pricing follows a less consistent pattern. The paper estimates a reasoning-model premium averaging 31.5 times non-reasoning prices. These findings depend on the study's model classifications, historical dataset, and statistical assumptions.

Why it matters: The research helps explain why inexpensive open models and premium reasoning systems can exhibit different price trajectories. It also provides context for the latest Token Price Index reading, which should not be interpreted as a universal measure of every model category.