It’s late 2026 so by now we all know that token economics is a big deal. Bragging about token spend was so 2025. Being more efficient about token usage is an individual, team and corporate level question. We all need to be thinking about more efficient use of LLMs from a cost perspective.
Using the right model for the job is increasingly the mantra – using a lighter, cheaper model for simple tasks, and an advanced, expensive model for complex reasoning or coding tasks. Frontier model company pricing leans into this. But competition is what causes prices to fall – one obvious answer to cut token spend is to use open weights models, especially as they become more and more capable. Frontier models can cost 10 to 50 times more per token than lightweight or open-source local alternatives. DeepSeek is crushing it right now – spend any time on Hacker News and you’ll see people lauding its models. Kimi from Moonshot AI is also getting a lot of love. For geopolitical reasons many enterprises are hesitant to use Chinese models but some startups are building SaaS companies that use them, and enterprises are happily working with those startups without realising it.
And then of course there is Jev, the classification model everyone’s going gaga for at the moment. See for example – We swapped our LLMs for Jev. It’s 39% cheaper.
Model routing is now a product and feature in its own right – see companies and projects such as OpenRouter, or model router for Microsoft Foundry. And of course Claude Code offers automated model choice, to help developers use the best Anthropic model for the job at hand.
Another take on the model cost question looks at not overusing LLMs, but instead building deterministic automations that don’t require LLMs in the workflow. Using models to write code that’s repeatable.
Spotify recently published a post with a solid approach to reducing LLM costs by Dimitri Mazmanov, Principal Product Manager at Spotify – Portal by Spotify cut my Claude Code token usage by 90%:
Most of what an AI coding agent does for me isn’t thinking. It’s I/O.
Reading five files to answer a question about one method. Generating a test file that follows the exact same pattern as the twenty test files next to it. Updating docs after a meeting. Thousands of tokens gone and almost zero reasoning. The seat license isn’t what hurts, it’s the tokens. And you’re feeding all of it to a frontier model that’s wildly overqualified. What if you could route the grunt work to something cheaper that handles it just as well, and save the expensive model for the problems that actually need it?
The I/O take is what I wanted to home in on here. Last traffic, fewer tokens, cost savings.
The post goes into some detail about the architecture and it’s worth reading in full. It’s not surprising Spotify would be focusing on cost efficiency. We’ve written before about cost optimisation baked into Spotify’s Backstage developer platform.
According to Mazmanov the use of Spotify Shunt, a model delegation tool, to route some traffic, dramatically cut token spend.
Tested against a Java monorepo across four scenarios, measuring tokens Claude would consume reading files directly vs. consuming the bulk-reader’s summary or writing code via the code-writer. Mean bulk-read savings were around a whopping 90%.
The code-write scenario is harder to measure in tokens because without shunt, Claude both reads the reference files and generates the output as expensive output tokens. With shunt, the code goes straight to disk, Claude never sees it.
The router works with AiKA (AI Knowledge Assistant), the experimental chatbot in Backstage. Mazmanov points to AiKA modes as the key to the approach
“They’re reusable. The same bulk-reader and code-writer modes work across every project and every tool that can shell out to the Portal CLI.
They’re shareable. Both modes are public in AiKA. Anyone can use them today without creating their own.
They’re composable. You could create a doc-writer mode for documentation, a reviewer mode for code review summaries, a translator mode for i18n. Each one is a few clicks away.
They decouple the routing decision from the worker. The plugin decides when to delegate. The mode decides how to respond. Swap Gemini Flash for a cheaper model, change the system prompt, add MCP tools – the plugin doesn’t change.”
So AI enabled routing with an I/O or batch automation focus, reducing traffic and use of Claude tokens. Yeah we’re going to see a lot more of this kind of approach.
No Comments