Disclosure: CodeRouter is our product. TokenWatt, which ran the benchmark and the cost analysis cited below, is the same team.
On November 6, OpenRouter launched Auto Code, a router "built specifically for coding agents and harnesses." Its announcement says a decision model rates each request on eleven things, such as "what kind of task this is, whether the task just changed, how hard it is," and picks a model and a reasoning effort. CodeRouter has routed coding-agent requests since April 2026, so we're glad to see the idea go mainstream. It also means more teams will be choosing between coding routers soon.
A router's pitch is always the same: Opus-level results for less. Here is what we've learned you have to measure to know whether that's true for your work, with our own numbers, including the ones that didn't flatter us.
1. Compare against a real baseline, not the router's estimate
On our 176-task coding benchmark, CodeRouter's own "saved vs Opus" counter said sending every task to Claude Opus would have cost $3.36. We then actually ran every task on Opus: $2.56. The estimate priced the routed run's tokens at Opus rates. A stronger model uses a different number of tokens and steps, so the estimate overstated the savings.
What to ask: is the "% cheaper" figure from two real runs of the same tasks, or from repricing one run?
2. Measure the session, not the request
In 385 developer-days of our own team's coding-agent traffic, over 90% of the money was input. An agent resends its context on every turn: the median request carried 69K tokens of context. Our most expensive 10% of days had 4.5× the requests of a typical day (915 vs 203) and 4× the cost per request ($0.09 vs $0.02), as context per request grew from 66K to 104K tokens.
One request's cost doesn't tell you what a session costs, because a different model can take a different number of turns to finish the same task.
What to ask: what did whole sessions cost, on tasks like yours, compared with the model you'd otherwise use?
3. Watch what happens to the prompt cache
Agents reuse most of their context, so the cache is where much of the bill goes. Across our developer-days, the share of input served from cache rose from 77% in May to 93% in September. Switching models mid-session throws that cache away. OpenRouter makes the same point: its router "switches only when the gain is worth the cache it gives up," and stayed on the same model for "nearly nine of every ten agent turns" in beta. CodeRouter keeps a coding-agent session on one model and moves up only when a code change fails its tests or the agent stops making progress.
What to ask: how often does the router switch models mid-session, what triggers a switch, and what does your cache hit rate look like afterwards?
4. Check that the benchmark looks like your work
Our public benchmark is 176 single-file Python tasks (Exercism plus Aider's hard set). Routing solved 96–100% of them across runs for 56–70% less than sending every request to Opus. It says nothing about long agent sessions in a large repository, and we say so. OpenRouter reports Auto Code "in line with Claude Opus 5.5 while costing 74% less per task" across five benchmarks, including Terminal Bench 4.0 and Deep SWE.
What to ask: single-turn or agentic? One repository or many? How many runs? Then run your own week of traffic through it.
5. Make sure you can see every decision
You can't trust a router you can't audit. Both routers now explain themselves per request: OpenRouter with a routing insights panel and router metadata in the API, CodeRouter with response headers (model, tier, phase, decider, cost) and a Playground that shows why each answer went where it did.
What to ask: can you get the reason for every request into your own logs, not only in a UI?
6. Know what you pay on top of the tokens
Read the fee schedule, not just the per-token prices: platform fees on credits, markups on usage, subscriptions. For CodeRouter: pay as you go is model list price +8% from a prepaid balance, with smart routing or Direct (you pick the model), and no monthly fee; Team is $19 per developer a month plus usage at list +8%. Pricing.
What to ask: what's the total, fees included, for a month of your team's traffic?
The short version
| Check | Red flag |
|---|---|
| Baseline | Savings computed from one run repriced at another model's rates |
| Unit | Cost per request instead of cost per finished session |
| Cache | Frequent mid-session model switches |
| Benchmark | Single-turn tasks used to sell an agent product |
| Transparency | No per-request reason you can log |
| Fees | Token prices quoted without platform fees |
We hold ourselves to the same list. If you want to see a router decide before you sign up, paste a task into the route preview: no account needed, no model runs, nothing billed.
Limits: the numbers here are our router and our own team's traffic (May–September 2026, list prices); the benchmark is single-file and single-turn. OpenRouter quotes are from its November 6 announcement and documentation, read on November 9.