Token Cost Radar

1What if the better model costs less per task without costing less per token?

OpenAI released GPT-6.1 Sol on September 29 at $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. On Terminal-Bench Science 0.1 at maximum reasoning effort, OpenAI reports an average cost of $5.47 per task versus $23.80 for GPT-6 Astra, while GPT-6.1 Sol more than doubled the score of GPT-6 Sol. The important unit is no longer just dollars per million tokens. It is dollars required to reach the desired result.

2OpenAI puts an explicit price on speed

OpenAI's new Ultrafast API service tier offers GPT-6 Astra at up to eight times Standard speed and recommends persistent WebSocket connections for tool-heavy agent workloads. GPT-6.1 Sol's API documentation separately lists Fast mode at twice Standard token prices, while Batch and Flex are 50% below Standard. The same model can therefore occupy multiple price points depending on whether the workload values latency or thrift.

3OpenAI turns heavy AI usage into a $500 monthly seat

Business Insider reports September 30 that OpenAI's new Pro 500 subscription costs $500 per month and bundles its highest usage allowances with access to Ultrafast compute in products including ChatGPT Work and Codex. The move is another example of AI spend splitting into two meters: variable API token billing for developers and increasingly expensive capacity or seat tiers for heavy interactive users.

4DeepSeek and Huawei push inference economics down into the chip stack

Reuters reports September 30 that DeepSeek and Huawei are collaborating on open-source programming infrastructure for Huawei Ascend AI chips, including compute and communication libraries and the TileLang programming language. The work accompanies a supernode design using 128 Ascend 950 chips. The project is strategically about reducing dependence on NVIDIA, but economically it is another attempt to improve the software-to-hardware efficiency underneath every generated token.

An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents

This study separates tool-schema filtering, content compression, and history summarization as distinct token-saving levers. Tool-schema filtering removed roughly 21,000 to 57,000 tokens on a typical turn and produced a reproducible linear saving, while compressed file reads accumulated across history and produced approximately quadratic savings until the context window capped the effect.

Why it matters: Tool-surface bloat is unusually attractive to attack because unused schemas impose a predictable recurring tax. Tokenminimizing starts with removing context the agent did not need in the first place.

AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows

AgentRouter assigns individual steps inside an agent trajectory to one of four model tiers rather than routing the entire workflow to a frontier model. Trained on 50,000 annotated trajectory steps, the authors report 72% lower cost than frontier-only execution while retaining 97.3% of frontier-only quality, with less than 5 milliseconds of routing overhead per step on an A100.

Why it matters: Modelmaxxing becomes more precise when the routing unit shrinks from the application to the task and then to the individual step. Planning may deserve frontier inference while formatting does not.

What Does Context Compression Cost an Agent? Interaction Costs Unrevealed by Task-Completion Metrics

This study finds that aggressive context compression can leave task-completion rates apparently unchanged while forcing agents to spend additional tool calls reacquiring information that was removed. At one fivefold compression point, GPT-5.5 completion moved from 80% to 85% without statistical significance while retrieval calls rose from 21.0 to 63.9.

Why it matters: A smaller context is not automatically higher token yield. If an agent must repeatedly fetch discarded state, apparent token savings can migrate into tool calls, latency, and later context.

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

Across eight frontier models on SWE-bench Verified, researchers found agentic coding tasks consuming roughly 1,000 times more tokens than simpler code reasoning and chat tasks. Repeated runs on the same task varied by as much as 30 times in token consumption, higher spending did not reliably produce higher accuracy, and models systematically underestimated their own eventual token use.

Why it matters: Agent budgets need external controls because token consumption is both stochastic and poorly self-predicted. More inference is not automatically more useful inference.