Best AI gateway / LLM router for coding agents (2026)
For coding agents, Standard Compute is the cost-optimized pick: it chooses an efficient model for each request and bills a flat monthly plan whose compute budget stops instead of overspending. OpenRouter is the pick for catalog breadth and exact model pinning, Vercel AI Gateway for tokens at list price with no markup, LiteLLM for self-hosting, and Requesty or Opper when an EU endpoint matters most.
Gateway or router: what you are choosing
An AI gateway puts one API and one key in front of many models, with logging, fallbacks and spend limits. A router goes one step further and decides which model answers each request. For a coding agent that decision is most of the bill: a loop that sends every file read and test rerun to a frontier model costs several times one that saves the frontier model for the hard steps.
Most gateways leave that choice to you. A few pick the model for you, and they differ in what they optimize: popularity, a quality score you set, or the cost of your plan.
The main options, compared
| Gateway | Who picks the model | How you pay | Best for |
|---|---|---|---|
| Standard Compute | The router, per request, from up to five models you select | Flat monthly plan from $19 with a published compute budget; requests stop at it | Coding agents on a fixed monthly cost, teams and agencies with a plan per key |
| OpenRouter | You, or its Auto Router, which ranks models by what other users spend on each task type | Provider price plus a 5.5% fee on credit purchases | The widest catalog and exact model pinning |
| Vercel AI Gateway | You; Vercel's own guide says the classifier is the part you own | List price, no markup on tokens | Apps already on Vercel, tokens at list price |
| Kilo Gateway | You, or kilo-auto/efficient | List price plus a 5% payment fee, or Kilo Pass credits | Kilo Code users |
| Requesty | You, through fallback, load-balancing or latency policies | Provider price plus 5%; 0% with your own keys | A Frankfurt EU endpoint on every plan |
| Opper | You, through pools and routes you build | Provider price plus 3% on credit purchases (5.5% on Control Plane) | EU-hosted gateway with ISO 27001 |
| Cloudflare AI Gateway | You, through Dynamic Routing flows, or its Auto Router for Chat Completions and Responses requests | Free gateway; 5% fee on Unified Billing credits, or your own keys | Teams already on Cloudflare |
| Portkey / Prisma AIRS | You, through routing configs | Your own provider keys plus a plan priced by logs | Enterprise security and governance |
| LiteLLM | You, or its Auto Router add-on, which sorts requests by difficulty tier on the proxy you run | Free open source; you host it and pay providers | Platform teams that self-host |
Checked on each vendor's own pricing and docs on October 13, 2026. Each name links to our full comparison with sources.
How to choose
- You want the monthly cost fixed and the model choice made for you: a flat plan with routing, such as Standard Compute.
- You need one exact model version, many providers, or light pay-as-you-go use: OpenRouter or Vercel AI Gateway.
- Prompts and keys must stay in your own infrastructure: LiteLLM, or Portkey's hybrid deployment on its enterprise plan.
- The gateway itself must be in the EU: Requesty's Frankfurt endpoint or Opper in Stockholm.
- Your agent is Claude Code or Codex: check that the gateway serves Anthropic's Messages API and OpenAI's Responses API. Kilo Gateway and ClinePass document Chat Completions only.
Gateways that changed in 2026
Several gateways changed course this year: Roo Code shut down its Router on May 15, TensorZero is no longer maintained, OpenPipe stopped serving inference on July 30, Unify moved away from routing, Glama narrowed its gateway to a curated set of providers, and Portkey became part of Palo Alto Networks. If you used one of them, our guide to discontinued AI gateways lists what changed and where to move.
The catch
Watch what the router optimizes. A gateway that bills per token earns more when you use more tokens, and a router that ranks models by popularity sends you where everyone else goes. A flat-plan router earns more when it handles your work with efficient models, which is your goal too, as long as hard tasks still reach a frontier model. Test any router on your own repository before moving a team.
Where Standard Compute fits
Standard Compute is one key for Claude Code (Anthropic Messages API), Codex (Responses API) and every OpenAI-compatible agent. The router picks an efficient model for each request from the lineup you select, plans run from $19 to $249 a month with a published compute budget, and requests stop at the budget with no overage. Companies and agencies get a plan or limit per key, and Europe or US region rules per key, team or company.
FAQ
What is the difference between an AI gateway and an LLM router?
A gateway gives you one API in front of many models, with keys, logs, fallbacks and spend limits. A router decides which model handles each request. Most gateways route only between providers of the model you named; a few, including Standard Compute, OpenRouter's Auto Router, Cloudflare's Auto Router, Kilo's auto models and LiteLLM's self-hosted Auto Router add-on, choose the model itself.
Which AI gateway is cheapest for coding agents?
Per token, the zero-markup gateways: Vercel AI Gateway at list price and LiteLLM if you host it yourself. Per month, what matters more is which model handles each step. A router that sends file reads and small edits to efficient models, inside a plan with a fixed budget, usually costs less than list-price frontier tokens for the same agent work.
Is OpenRouter a smart router?
It has one. OpenRouter's Auto Router sorts each request into a task type and ranks models by the share of spend other users give them for it, and its Pareto router picks the cheapest model above a coding score you set. Many users still pin models directly. Billing is pay-as-you-go plus a 5.5% fee on credit purchases.
Does smart routing hurt output quality?
It can if the router saves money at any cost. Judge a router by whether hard tasks still land on frontier models: run the same task from the same commit through it and through your current setup, and compare the result, retries and cost.
Which gateways work with Claude Code and Codex?
Standard Compute, OpenRouter, Vercel AI Gateway, Requesty, Opper, Portkey and LiteLLM all document Claude Code. Codex needs OpenAI's Responses API, which most of them serve. Two exceptions: Cloudflare's Codex setup covers OpenAI models only, and Kilo Gateway documents Chat Completions only.
Is smart routing on a flat subscription a real product yet?
Yes. Standard Compute is one OpenAI- and Anthropic-compatible endpoint with smart routing included in a monthly plan from $19: send the model standardcompute and each request goes to a cost-efficient model among the ones you select, such as Qwen 3.8 27B or GPT-6 Luna for routine steps and GPT-5.6 Sol for hard ones, and the plan price is the whole bill unless you turn on Pay as you go. Codex connects through the Responses API with standardcompute launch codex. The router decides per request and does not learn your repository, so check in the request log that hard tasks still land on the strong model.