OpenRouter claims 24x inference-cost savings with MCP model routing

• OpenRouter's cost-saving post claims an OpenRouter MCP workflow cut inference spend 24x at comparable quality for an image-extraction and recipe-generation workload.
• The modeled per-call ladder in OpenRouter's screenshot puts Claude Sonnet 4.6 at about $0.0125, Claude Haiku 4.5 at about $0.0042, and Gemini 2.5 Flash Lite at about $0.0004.
• OpenRouter's routing reply says the model chooses using codebase context plus OpenRouter data APIs, including benchmarks, aggregate usage data, and live performance stats.
• emollick's routing post and levie's context post both frame the same pattern: a stronger model plans or reviews while cheaper models handle volume work.

The Reddit post OpenRouter amplified includes the useful part most routing demos skip: the actual per-call cost table and a claim of A/B split testing for quality. OpenRouter separately pointed readers to the Reddit story, its MCP explainer, and the getting-started link. The sharp edge is not provider choice by itself, it is letting the model inspect the task and decide when Sonnet-level spend is unnecessary.

The workload in OpenRouter's screenshot was image extraction and recipe generation. The Reddit post says the author used an MCP tool to optimize API usage, then used A/B split testing to check quality.

The modeled per-call table gives the cost ladder:

• Claude Sonnet 4.6: about $0.0125, baseline 1x.
• Claude Haiku 4.5: about $0.0042, roughly 3x cheaper.
• Gemini 2.5 Flash Lite: about $0.0004, roughly 30x cheaper.

OpenRouter's headline number, 24x, lands between the table's modeled endpoint and the Reddit post's claimed observed savings.

OpenRouter's reply makes the routing loop explicit: the model decides. The inputs OpenRouter named were:

• the user's codebase,
• OpenRouter benchmark data,
• real-world aggregate usage data,
• live performance stats,
• other OpenRouter data APIs.

That turns MCP into a policy surface for model selection, not just a way to expose tools. OpenRouter's MCP follow-up links to more detail about the MCP, while OpenRouter's setup post points to the getting-started path.

The same routing story shows up one layer below model choice. wafer_ai's OpenRouter screenshot shows OpenRouter's provider table for Z.ai GLM 5.2, sorted by throughput, with Wafer Fast at 169 tps, Fireworks Fast at 98 tps, Weights & Biases at 91 tps, and several providers clustered between 62 and 85 tps.

The follow-up in wafer_ai's GLM post claimed GLM 5.2 on AMD MI355X reached 2626 tok/s/node and 213 tok/s single stream at more than 2x lower cost than Blackwell. For routing systems, model slug, provider, hardware path, latency, throughput, uptime, and price all become selection variables.

emollick's post puts the thesis cleanly: start with a smart planner, then let it delegate work to cheaper models. levie's post describes the applied AI layer doing the same thing at product scale, using frontier models for planning, orchestration, and review, then cheaper open or closed models for the high-volume work between those steps.

This is the cost model behind the 24x anecdote. The expensive model earns its keep when it decides what work is worth doing, what context is relevant, and which cheaper worker can finish the call.

torchcompiled turned the example into an enterprise-spend argument, saying companies spending millions to billions on employee token usage are missing savings from tricks like this. The same post named Composer, Kimi, and GLM as viable cheaper options.

The claim is broad, but the mechanism is concrete: route routine calls away from the default premium model when the task and quality bar allow it.

Several primary posts around the same window were link-only dispatches rather than commentary. dejavucoder's post adds one implementation-adjacent artifact: an attached systems sketch labeled broadcast, send, reduce, tensor parallelism, allreduce, expert parallelism, allgather, context, and reduce-scatter.