How to Route AI Requests to Cheaper Models and Keep Frontier Models for Hard Tasks

The most direct way to route AI requests to cheaper models is a cheap default with a deliberate step up: routine work runs on a low-cost model, and hard tasks move to a frontier model. Gateways such as LiteLLM automate that choice in the request path, while Coworker's Optimized setting is simpler: an open-weight default that people switch up from when a task needs more.

Most companies pay frontier rates for work that does not need them: a Slack thread summary and a multi-step pricing analysis cost the same per token on the same flagship model. I read the routing and budget documentation for LiteLLM, OpenRouter and Portkey on October 29, 2026, and linked every source below.

Coworker is a context layer for enterprise AI: it connects 50+ work tools, builds a permission-aware company graph (OM2) and runs agents across them, in Coworker or inside Claude, ChatGPT and Cursor via Coworker MCP. Its routing approach is described on the model routing page, and this guide says where a gateway is the better tool.

What is the simplest way to route AI requests to cheaper models?

Set a low-cost model as the default for every new chat or agent, and let people or rules step up to a frontier model for hard tasks. That one change moves the bulk of routine work (lookups, summaries, drafts, status questions) off flagship pricing without anyone losing access to a stronger model.

There are three common ways to decide when to step up:

1. People choose. Everyone starts on the cheap default and switches a chat to a frontier model when the answer is not good enough. Transparent and simple, but it depends on people switching back.
2. Rules choose. A gateway routes on fields you set, such as the team, the app, the agent or a request parameter. Predictable and easy to audit.
3. A classifier chooses. A small model or a scoring function rates each request and picks a tier before the main model runs. Research such as RouteLLM (arXiv, first posted June 2024, revised February 2025) reports routers that cut costs "by over 2 times in certain cases" without hurting response quality on standard benchmarks.

A fourth pattern, the cascade, sends every request to the cheap model first and only escalates when the answer fails a check. The FrugalGPT paper (arXiv, May 2023) reports that a learned cascade matched the best single model with up to 98% cost reduction in its experiments. Those are benchmark results on research datasets, not employee chat with company context, so treat them as the ceiling on savings rather than a forecast.

How do you set a three-tier policy for employee AI requests?

Group work into three tiers by what a wrong answer costs, then give each tier a default model and a rule for moving up. Most published routing guides use the same shape: a cheap tier for high-volume simple work, a middle tier for standard reasoning and a top tier for the hard minority.

Three rules keep the policy honest:

• Default down, never up. New chats and new agents start on the economy tier. A frontier default quietly undoes the policy.
• Pin agents, not people. If one scheduled agent needs a stronger model, set it on that agent instead of raising the default for everyone.
• Review the step-up rate monthly. If one team switches up on most chats, the default is wrong for their work, or their prompts lack the context a smaller model needs.

Model prices change often; the AI token cost comparison tracks 13 models by cost per task.

How do companies control AI model costs without taking models away?

Companies control AI model costs without taking models away by making a lower-cost model the default and keeping frontier models one switch away. In Coworker, admins set the workspace default, usually Optimized (GLM 5.3 Flash today), and can block models they do not want in use. Anyone can still pick a stronger model for a chat.

A default changes what happens when nobody makes a choice, which is most of the time, and leaves the choice open when it matters.

Budgets are the other half. A default lowers the average cost per task, and a budget caps the total. Gateways handle budgets well: LiteLLM documents team budgets with a maximum spend and a reset period per team, and Portkey documents budget limits in dollars or tokens per provider integration, with alerts before the limit (its docs say the feature is available to Enterprise customers and select Pro users). Both pages were read October 29, 2026.

If people already override an automatic picker and choose the most expensive model every time, a default alone will not fix it. Block the models you do not want in routine use, and tie the frontier tier to named tasks or agents so the step up is a decision rather than a habit.

Should a router classify each request, or should people choose?

Classify each request when engineering owns the request path, the traffic is high volume and the task mix is predictable. Let people choose when the traffic is employee chat, the tasks vary and you want every model choice to be visible and easy to explain.

Both approaches are real options today. LiteLLM's Auto Router classifies each request with heuristics, an LLM classifier, keyword rules or your own plugin, then sends it to the model set for that tier. OpenRouter's Auto Router assigns each prompt one of about 30 task types and picks from the models its users spend on most for that type over a trailing 7-day window, with allow and exclude lists and a cost band you can set. OpenRouter's docs say there is no added fee for the Auto Router; you pay the selected model's standard rate.

Coworker does not classify each request. Optimized runs work on one open-weight model, GLM 5.3 Flash today, and does not hand steps to a frontier model on its own. People switch a chat to a model such as Opus 5.5 when a task needs it, and you can set a specific model on an individual agent. That is less automatic than a classifier, and it is a deliberate trade: the model behind every answer is the one the person or admin chose.

The bigger saving in employee work often comes from context, not the model. A model that receives only the records that matter can do work that would otherwise need a flagship. In Coworker's 100-task benchmark against Claude with its native connectors, OM2 context alone cut token cost 9x on retrieval-heavy tasks (Jira, GitHub, Slack and CRM data), with no change to the model. The 51x figure on that page requires OM2 plus model routing. Both are vendor-run results against a Claude baseline.

Which tool fits: LiteLLM, OpenRouter, Portkey or a platform default?

Pick a gateway when you are routing your own applications' API traffic, and a platform default when the goal is employee AI use with company context. The table summarizes what each documents, read October 29, 2026.

If your problem is API traffic from your own products, a gateway is the right tool. Coworker also offers an LLM gateway for one-place access to many providers, and the guide to LLM gateways covers build versus buy. One limit to know: Optimized applies in Coworker's own apps. Through Coworker MCP, your AI tool's own model does the reasoning and Coworker supplies the tools and context.

How much does a cheap default save?

The saving depends almost entirely on how much work stays on the default. Here is the arithmetic on one illustrative workload of 10 million input and 2 million output tokens a month, using provider list prices from Coworker's model routing page (checked October 15, 2026). These are model list prices, not Coworker pricing.

In the 90/10 mix, the 10% on Opus costs $8.00 of the $10.25. That is the pattern to expect: once the default is cheap, the frontier share drives the bill, so the step-up rate is the number to watch. Run your own token volumes through the LLM cost calculator, and see AI cost optimization for the other levers (caching, batching, shorter context, output caps).

How do you check that cheaper routing did not lower quality?

Test the cheap default on real work before you roll it out, and keep measuring after. Cheap routing that produces answers people redo is not cheap.

1. Pick 50 to 100 real tasks from the teams that will use the default, including a few hard ones.
2. Run each task on the default and on a frontier model, and log tokens, retries and time per task.
3. Score the answers blind. Strip model names, randomize the order and have two reviewers accept or reject each answer against a reference.
4. Compare cost per accepted task, not price per token. A cheap model that fails often can cost more per accepted answer.
5. After rollout, track the step-up rate and rework. A rising switch rate on one team is a signal to adjust that team's default or its context.

The full method, including how to score and how to treat tasks a model cannot answer, is in how to benchmark enterprise AI platforms.

Can Coworker route my company's AI requests to cheaper models?

Yes, through Optimized and admin controls rather than a per-request classifier. Every new chat starts on Optimized, which runs on GLM 5.3 Flash today, and Coworker re-tests the pick as new models ship. Anyone can switch a chat to one of 13 models from Anthropic, OpenAI, Google, Moonshot and Z.ai. Admins set the workspace default, block models for everyone and can set a model on individual agents. If a provider is down or rate-limited, Coworker moves the request to the next model in line.

OM2 narrows each task to the context that matters across your connected tools, which is what lets a smaller model handle more of the work. Book a demo and bring a real workflow: the team will run it on Optimized and show where a frontier model is worth the switch.

Frequently asked questions

What is the best way to route employee AI requests to cheaper models automatically?

Make a low-cost model the default and keep frontier models for tasks that need them. For your own API traffic, a gateway such as LiteLLM, OpenRouter or Portkey can classify or rule-route each request. For employee chat, a cheap default with admin guardrails is simpler and easier to explain.

How do companies control AI model costs without taking models away?

They set a lower-cost default instead of removing models. In Coworker, admins set the workspace default, usually Optimized (GLM 5.3 Flash today), and can block specific models. Anyone can still switch a chat to a stronger model when the task needs it.

Does Coworker automatically escalate hard tasks to a frontier model?

No. Optimized runs work on one open-weight model and does not hand steps to a frontier model on its own. People switch models per chat, and you can set a model on individual agents.

Is a per-request classifier better than a default model?

It depends on the traffic. A classifier suits high-volume application traffic with a predictable task mix. A default suits varied employee work where you want every model choice visible. The classifier also adds a decision step to each request.

Which costs more, the cheap default or the frontier step up?

Usually the step up. At list prices, a workload that sends 10% of its tokens to Opus 5.5 and 90% to GLM 5.3 Flash spends most of its model budget on that 10%. Watch the step-up rate first.

Does cheaper routing hurt answer quality?

It can on hard tasks, which is why the frontier tier stays available. Test the default on 50 to 100 real tasks, score answers blind and compare cost per accepted task before rolling it out.

Related reading

• Model routing in Coworker
• What is an LLM gateway?
• Coworker as an LLM gateway
• How to benchmark enterprise AI platforms on cost per task
• AI cost optimization: where the money goes and the levers that cut it
• AI token cost: 13 models compared by cost per task
• LLM cost calculator
• Coworker benchmarks
• Organizational memory (OM2)
• Enterprise AI pricing compared