Token Cost Radar

1What happened to the Sonnet 5 price hike scheduled for today?

Anthropic's live pricing documentation confirms that the planned September 1 increase for Claude Sonnet 5 will not occur. The $2 per million input tokens and $10 per million output tokens introduced as temporary launch pricing are now the standard rates. Anthropic also notes that its newer tokenizer produces approximately 30% more tokens for the same text, so effective task cost still depends on workload shape rather than rate card alone.

2Employee token spend gets a headcount-style approval workflow

TeamOhana's Token Spend Management becomes generally available September 1. The company says the system attributes AI dollars to the employee they serve and routes requests for additional AI spending through approval workflows similar to those used for headcount. Its launch follows a summer in which Ramp and Rippling both introduced tools for tracking and controlling employee AI consumption.

3Model routing attracts another seed round

Axios reports that TrustedRouter raised $1.25 million to build routing infrastructure that sends AI requests toward models based on factors including cost, capability, and privacy requirements. The funding is small, but the signal is larger: model selection is increasingly being treated as a dynamic execution decision rather than a one-time vendor choice.

4Boards start asking whether the token bill adds up

The Australian reports that companies including CBA, Suncorp, Telstra, Qantas, and BlueScope Steel are putting greater financial scrutiny around AI spending as cloud and inference costs rise. The article describes real-time monitoring and explicit cost-benefit tests as businesses move from broad experimentation toward proving measurable productivity and financial return.

One Model, Many Minds: Unlocking Multi-Agent Synergy in a Single Agent via Mixture of Roles

Submitted August 27, this arXiv paper replaces multi-agent text exchanges with a single model that dynamically combines learned latent roles. Across its reasoning and personality benchmarks, the approach performed on par with multi-agent systems while reducing token cost by 20 times.

Why it matters: Multi-agent architectures can spend substantial tokens having models explain work to one another. Some of that specialization may be achievable inside one inference pass without paying the communication tax.

CacheRouter: Tool Discovery Without Sacrificing Prompt Caching

CacheRouter studies a conflict between progressive tool disclosure and prompt caching. Changing the visible tool list saves schema tokens but can invalidate reusable prefixes. Its dual-path architecture keeps a stable core tool surface while routing long-tail tools separately, reaching token-level cache-hit rates above 90% in prototype experiments.

Why it matters: Tokenminimizing can erase its own savings when a smaller but constantly changing prompt destroys cheap cached context. Tool-surface optimization has to preserve cache geometry as well as reduce resident schemas.

Paritok-4B: Intent-Conditioned Context Compression for Coding Agents

Paritok-4B compresses coding-agent file reads and tool outputs by selecting exact spans rather than paraphrasing them. Across all 300 SWE-bench Lite instances, it reduced context to 25.7% of the original size while retaining 86.5% of uncompressed single-shot solve quality. The authors also found that using GPT-5 as the compressor was economically negative at list prices.

Why it matters: Compression itself has a bill. A system only improves token economics when the cost of deciding what to remove is lower than the downstream inference it eliminates while preserving the evidence needed to solve the task.

ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents

ReCache independently caches recurring tool and skill representations so changing their order or combination does not force the model to rebuild all of their KV state. Across seven tool and skill datasets, it matched dense invocation performance closely while producing a 3.655 times time-to-first-token speedup and reducing allocated KV-tensor memory by 92.43%.

Why it matters: Tool-surface economics exist below visible prompt tokens. Reusing the computed representation of recurring schemas can reduce the cost of broad agent capability without making every tool permanently resident in fresh context.