Hidden Caching Costs Make Sonnet 4.6 Cheaper Than GPT-4.1

On July 15, 2026, IBM Research published a technical analysis on Hugging Face detailing why production-scale model routing systems frequently fail to deliver expected cost savings. The study reveals that relying on base token prices to route tasks between models ignores the operational realities of caching, invisible task complexity, and infrastructure latency.

While academic frameworks like RouteLLM and FrugalGPT suggest that directing simple tasks to small models and complex tasks to frontier models can reduce costs by 60 to 80 percent, real-world execution introduces massive variables. As the price spread between models like DeepSeek V4 at $0.44 per million tokens and GPT-5.5-pro at $30 per million tokens reaches nearly 100x, routing logic has become the primary cost lever for developers.

The Caching Factor in Agentic Workloads

To demonstrate the flaw in static sticker-price routing, IBM tested a CodeAct agent across 417 tasks on the AppWorld Test Challenge. The test compared Anthropic’s Claude Sonnet 4.6 against OpenAI’s GPT-4.1.

Model	Total Cost	Cost Per Task	Reasoning Steps
Claude Sonnet 4.6	$79.00	$0.19	3x baseline
GPT-4.1	$155.00	$0.37	1x baseline

Despite GPT-4.1 possessing a lower published per-token sticker price, Sonnet 4.6 was nearly 50 percent cheaper to operate in an agentic workflow. Sonnet required three times as many reasoning steps to complete the tasks, which intuitively should have driven the cost higher.

The inversion occurred because of context caching. Agentic workloads continuously reuse large chunks of context across multiple steps. Sonnet 4.6 features significantly lower cache-read pricing, which provided a compounding advantage over multiple agent iterations. A standard routing algorithm evaluating how to reduce LLM API costs in production based solely on the base token price would have incorrectly defaulted to the more expensive execution path.

Invisible Complexity and Infrastructure Constraints

The IBM study also highlights that task difficulty is often invisible at the moment a routing decision is made. A prompt such as “summarize this contract” appears trivial but frequently triggers high-complexity sub-tasks like retrieval, compliance checking, and multi-round refinement. Conversely, highly technical prompts are sometimes resolved instantly by smaller, specialized models.

Because of this, routing must function as a multi-objective systems optimization problem rather than a basic text classification task. Serving infrastructure often overrides raw model speed in production environments. Enterprise implementations require routers to balance latency, cost, and quality alongside strict governance constraints like data residency and approved-model lists.

This operational complexity is visible in enterprise gateway configurations like LiteLLM Enterprise, where governance constraints often override raw performance metrics when implementing multi-agent coordination patterns.

Production Implementations

IBM has integrated these routing insights into its own internal tools. IBM Bob, an agentic software development platform updated on July 9, 2026, uses this dynamic routing approach across Anthropic’s Claude, Mistral, and IBM Granite models.

Similarly, IBM Power Autonomous Operations, launched concurrently with the research on July 15, applies this internal routing logic to monitor and resolve Power system capacity issues. The platform operates up to 15 times faster than manual operations by optimizing the model selection for each specific diagnostic step.

If you build autonomous systems, you must calculate task costs based on the complete execution loop, not the initial prompt. Implement routing logic that accounts for cache-read discounts, infrastructure latency, and data governance requirements rather than relying on static model pricing sheets.