Token Cost Radar

1What happens when token price depends on what time you call the model?

DeepSeek announced new peak and off-peak pricing for its V4 lineup, effective August 16. Off-peak rates will be half the peak rate. For V4 Pro, cache-miss input will cost $0.66 per million tokens off-peak and $1.32 at peak, while output will cost $1.98 and $3.96 respectively. InfoWorld notes that some cache-hit rates rise by more than tenfold from current pricing, turning workload scheduling into a new cost-control variable.

2Google cuts the price of its newest Flash model in half

Google introduced Gemini 3.7 Flash with introductory pricing at half the original Gemini 3.6 Flash rate. Google lists introductory Gemini 3.7 Flash pricing at $0.375 per million input tokens and $1.875 per million output tokens through December 31, 2026, reinforcing the trend toward cheaper high-volume models for agentic workloads.

3Writer says the harness can move the bill more than the model menu

Writer says its upgraded agent system cuts average cost per task by 41% across the models it tested while maintaining quality, and by 52% when paired with its new Palmyra X6 model. The company attributes the gains to adaptive reasoning, orchestration, batching, delegation, and tighter control of long-running agent execution, while adding token monitoring, alerts, and hard consumption limits.

TokTier: Exact Stateful Tokenization for Agentic LLM Serving

This recent arXiv paper studies 153,951 agent calls and finds that, even with a 94.1% prompt-cache hit rate, tokenization can consume as much as 64% of time to first token. TokTier incrementally repairs tokenization for appended agent context rather than repeatedly tokenizing the entire transcript, cutting median time to first token by 16% to 34% when integrated with vLLM.

Why it matters: Agent efficiency has another hidden layer below model inference. Repeatedly re-tokenizing long transcripts can remain expensive even when the KV cache is working exactly as intended.

LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers

This recent arXiv paper evaluates more than 16 routing approaches across single-turn, multi-turn, personalized, memory, vision, and other workloads. Learned routers outperformed the strongest fixed-model baseline by 14.6% relatively, while lightweight routers became more competitive as cost constraints tightened.

Why it matters: Modelmaxxing is becoming an engineering discipline rather than an intuition. The useful router must optimize quality and inference cost together, under the actual budget available.

Prompt-Induced Waste in Coding Agents

Across thousands of coding-agent runs, this arXiv benchmark finds that prompt wording can materially change reasoning volume and tool behavior without improving correctness. Instructions to generate and compare multiple approaches increased reasoning tokens by 2.4 to 7.4 times in measured settings, while bounded scope and explicit stopping conditions could reduce waste.

Why it matters: Token efficiency starts before the first model response. A prompt can quietly authorize extra branches, verification, tools, and context growth that never improve the accepted result.

Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets

This arXiv study evaluates 28 tool definitions and reports 44% to 50% tool-schema token savings from conservative compression. At an 8,000-token context limit, full schemas caused the evaluated agent configuration to overflow, while compressed schemas left enough context for retrieval and execution.

Why it matters: Tool-surface bloat is both a cost problem and a capacity problem. Tokens spent describing unused capabilities are unavailable for the evidence the agent actually needs.