OpenRouter Auto Router and Auto Code: One Line Picks the Model

Picking the right AI model for each prompt has become a job in itself. New models land every week, prices shift, and the cheapest option on Monday can be the wrong one by Friday. OpenRouter's answer, launched November 6, 2026, is to stop picking at all: its rebuilt Auto Router chooses the model for you, and a new sibling router, Auto Code, does the same job tuned specifically for coding work. The headline number is striking. Auto Code scores 74.3% on the Deep SWE coding benchmark at $1.98 per task, which OpenRouter says puts it in line with Claude Opus 5.5 at 74% less cost per task. But the fine print matters more than the headline, because this launch also makes the router more expensive than the version it replaces.

OpenRouter is the model exchange that puts more than 600 AI models behind one OpenAI-compatible API and one bill. Its Auto Router has run for nearly three years, routing over 2.5 trillion tokens a month from more than 100,000 people, according to the company's announcement. This release rebuilds that router from scratch on a new technical foundation, adds Auto Code (openrouter/auto-code) alongside the flagship Auto Router (openrouter/auto), and keeps openrouter/auto-beta as the experimental channel where routing changes land first. One small date wrinkle: OpenRouter's Auto Code model page lists November 5, 2026 as the release date, while the announcement went out November 6. Either way, this is a launch from this week, not a rebrand of an old feature.

What actually launched

Three things shipped together. First, the next-generation Auto Router, rebuilt to weigh eleven factors about each request instead of the handful of rough signals, such as task type, that drove the old version. Second, Auto Code, a new router built on the same engine but tuned from the ground up for coding agents and multi-turn software engineering work. Third, the transparency to see what the router is doing: a routing insights panel in OpenRouter's chatroom shows why a model was picked, and the same analysis is available through the API as router metadata.

The routing engine underneath is Jev Router, from TypeSafe, and OpenRouter says it verified the new generation on production workloads through typesafe/jev-router and openrouter/auto-beta while respecting zero data retention throughout. In that beta period, invalid tool calls ran 4 times lower than the previous router and agentic responses were cut off half as often, according to OpenRouter's own figures. Both numbers are company-reported, from OpenRouter's beta, not an independent test.

How the router picks a model

Every request passes through five stages, documented in the Auto Router guide. First, the router narrows the model pool using your guardrails, zero data retention requirements, provider restrictions, bring-your-own-key settings, and any allowed or excluded model lists. Second, a fast decision model rates eleven aspects of the request in a single call: task type, difficulty, whether the task just changed, whether speed matters, whether it needs vision or codebase knowledge, and whether a bigger model, more thinking, or a cheaper model would serve it better. Third, the routing policy weighs that rating against live prices, cache status, and benchmark performance to pick both the model and the reasoning effort. Fourth, the request goes through OpenRouter's normal provider routing, which chooses who serves that model based on provider health, latency, price, and whether a provider already holds your prompt cache. Fifth, the results of each turn inform the next turn, so a struggling session can trigger a switch.

Two design choices in that flow are worth understanding. The router is deliberately restrained about switching models mid-session, because switching forces the new model to rebuild the input token cache, which costs extra. It prefers to change the reasoning effort first and switches only when the gain is worth the cache it gives up; in the beta period it stayed on the same model for nearly nine out of every ten agent turns. It also adapts to failure, so repeated errors or tests failing after an edit raise the odds of a switch.

You steer all of this with one setting, cost_tier, which takes low, medium, or high. Low prioritizes lower estimated cost, medium balances cost and capability, and high favors more capable models when the expected benefit justifies the added cost. The older cost_quality_tradeoff and pin_model settings are deprecated: requests that send them still succeed, but the current router ignores them. You can also save allowed models, excluded models, and a cost preference as account defaults on your workspace's Routing page, with a "prevent overrides" toggle that makes saved values final.

The benchmark numbers, read honestly

OpenRouter published a comparison of the new Auto Router at its default tier, Auto Code, and the previous Auto Router across five benchmarks. The scores and the cost per task tell two different stories, and both matter.

The score improvements are real and large. On τ³-bench Banking, the new Auto Router jumps from 13.7% to 53.1%, nearly four times the old score, which is where OpenRouter's "outperforms the old router by up to 4× depending on the benchmark" claim comes from. Terminal Bench 4.0 rises from 28.3% to 45.7% on Auto Router and 50.0% on Auto Code. The routing decision itself is faster too: 0.65 seconds at p90 against 1.96 seconds before, about three times quicker.

But read the cost column before celebrating. The new routers cost more per task than the old one on four of the five benchmarks. SWE-Atlas QA costs three times as much ($0.36 vs $0.12), Terminal Bench more than twice as much ($1.33 vs $0.61), and τ³-bench Banking nearly seven times as much ($0.60 vs $0.087). Only Deep SWE got cheaper on Auto Router ($0.57 vs $1.37). OpenRouter is open about why: the old router had drifted toward the cheapest models because they were so efficient, leaving stronger models out, so the new default deliberately returns to pricing closer to where Auto Router sat a few months ago. If you preferred the old behavior, set cost_tier to low.

The 74% figure needs the same care. Auto Code costs 74% less per task than single frontier model calls at comparable accuracy, with Claude Opus 5.5 as the comparison point, according to OpenRouter. That is a saving against calling Opus 5.5 directly, not against the previous Auto Router, which was cheaper per task on most benchmarks. All of these figures are OpenRouter's own benchmark runs, published on its model routing benchmarks page. No independent lab has replicated them yet, and the company says it reruns them periodically as new models ship.

OpenRouter also states the limit plainly in its announcement: no router on the market beats the single best frontier model at every point on the cost and quality curve, and the new Auto Router is no exception. A router trades a little peak performance for consistent, cost-efficient performance over time. If you already know which model works best for your specific work, the company says to keep using it. That trade-off is not unique to OpenRouter: Together AI's Polymorphic Inference takes a different route to the same goal, tuning the model itself to the workload instead of switching between models.

What it actually costs

The routers themselves are free. OpenRouter charges no fee for routing: you pay the standard rate of whichever model serves the request, and the response's model field tells you which one it was. That is the whole pricing story for the router, but not for OpenRouter itself, and the platform fees are where the real bill is decided.

OpenRouter does not mark up provider prices. Inference is billed at the provider's list price on every plan, and OpenRouter's fee is charged when you buy credits: 5.5% on the Standard plan and 8% on Business, never on individual requests. So $100 of Standard credit costs $105.50, and the same $100 on Business costs $108. Enterprise pricing is set by contract, based on volume, prepaid credits, and annual commitments, with invoicing. Prices are exclusive of applicable taxes, and VAT or GST may be added on invoices where required.

Using Auto Code's headline figure as an example: $1.98 per Deep SWE task is about £1.48 or C$2.80 at November 7, 2026 market rates (1 USD = 0.7471 GBP, 1 USD = 1.4117 CAD), and the Auto Router's $0.57 Deep SWE task is about £0.43 or C$0.80. Those are per-task benchmark costs, not a subscription, so your monthly spend depends entirely on how many tasks you run and which models the router picks. A provider.max_price setting still applies as a cap: it filters the endpoints of whichever models the router resolves, so you can put a ceiling on what any single request may cost.

There is a free way in. The free tier includes 25 or more free models from four free providers, with free models capped at 20 requests per minute and 50 per day. That daily cap rises to 1,000 requests a day once your account has bought at least $10 of credits, about £7.47 or C$14.12. Failed attempts count toward the daily cap, and a negative balance blocks free models until you add credits.

Two billing details work in your favor. An attempt that errors is not charged for its model tokens, so with fallback routing you pay only for the run that answers. Under Zero Completion Insurance the same holds for a response that returns nothing and no finish reason, though auxiliary services that already ran, such as web search or file parsing, may still be billed. Bring your own provider keys and the first $25,000 of list-price inference per month, about £18,678 or C$35,293, carries no OpenRouter fee; above that allowance OpenRouter charges 5% of the list price, deducted from your credits.

How to use it

Using either router is a one-line change. Set the model to openrouter/auto or openrouter/auto-code in any request, using the same OpenAI-compatible API at https://openrouter.ai/api/v1:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openrouter/auto-code",
    "messages": [{"role": "user", "content": "Add a retry with backoff to fetchUser in api/users.ts"}]
  }'

The response includes a model field naming the model that actually served the request, so you can log what you paid for. To see why the router picked it, send the X-OpenRouter-Metadata: enabled header: the jev-router stage in the response metadata lists the resolved models, the candidates with their reasoning effort and estimated cost, and the reason for the pick. In OpenRouter's chatroom, a routing insights panel shows the same analysis beside each reply. The Auto Router and Auto Code model pages also show a live routing mix, and at launch GPT-6.1 Sol led Auto Router's mix at a 33.8% share of requests.

Sessions matter for coding work. The router recognizes a conversation from an explicit session_id or from a fingerprint of your messages, and it carries signals between turns: whether the model is struggling, what is already cached, and which provider served the last turn. Sessions also keep requests on the same provider, which preserves the prompt cache.

If you want control rather than full automation, the allowed_models and excluded_models settings accept wildcard patterns such as anthropic/* for all Anthropic models or openai/gpt-5* for a model family. The allow list is a hard restriction, and exclusions are applied after it, so an excluded model is never selected even when it matches an allowed pattern. One failure mode to know: if your restrictions leave no eligible model, the request fails with a 404, and the error names the fields to change.

Who should use it, and who should not

Routers earn their place in three situations, and OpenRouter names them: when the work is varied, when an organization needs to hold costs steady, and for people who do not have the time or interest to track the constant stream of model releases. A general agent handling mixed traffic, a coding agent running many different tasks, or an app whose users send unpredictable prompts are the natural fits. Auto Code is the pick for coding agents specifically, since it starts in a higher cost tier, weighs measured pass rate, speed, and cost per solved task, and penalizes failures more harshly so it corrects course faster on work where success and failure are clear cut.

Skip the router if you already know the best model for your workload. A fixed model gives you peak performance on that task, predictable per-token costs, and no routing step adding latency before the first token. The router also requires conversational input, a message list rather than the legacy prompt field, and it is available only in the global data region. Auto Code is not currently available through EU in-region routing, so teams with regional data requirements should check that before switching.

One ownership note: Stripe agreed to acquire OpenRouter in August 2026, reportedly for about $7.5 billion, and the deal was still awaiting regulatory approval at the time of this launch, according to reporting. Nothing in the announcement changes pricing or terms, but an acquisition of that size can reshape a platform's fees later.

Five things to get right

First, set your cost tier deliberately. The new default costs more per task than the old router, by design. If your bills climb after the switch and your quality did not need to, cost_tier: low restores behavior closer to the previous generation.

Second, do not combine openrouter/auto with another router model in the same request. The docs list this as a limitation, and the routers are designed to be the single model string for a request.

Third, treat the benchmark table as a starting point, not a promise. The figures are OpenRouter's own runs, rerun as models change, and your tasks will route differently. The live routing mix on the model pages is a better guide to what the router actually picks day to day.

Fourth, use the transparency tools before you trust the router with a budget. Run a few representative prompts in the chatroom, read the routing insights panel, and check the model field and router metadata on your first API calls. If the router keeps picking a model you would not have chosen, the allow and exclude lists are the fix.

Fifth, mind the free tier's edges. The 50-request daily cap counts failed attempts, and a negative balance blocks free models entirely. Buying $10 of credit lifts the cap to 1,000 requests a day, which is the cheapest way to test the routers at any real volume.

Frequently asked questions

Does Auto Router cost extra?

No. You pay the standard rate of whichever model serves the request, and the response's model field tells you which one it was. OpenRouter's fee is charged when you buy credits, 5.5% on Standard and 8% on Business, never on individual requests.

What is the difference between Auto Router and Auto Code?

Both run on the same routing engine. Auto Router is the general-purpose router for mixed traffic and general agents and favors a cost-efficient tier of models. Auto Code is specialized for coding agents and harnesses, starts in a higher cost tier, and weighs measured pass rate, speed, and cost per solved task.

Can I control which models the router picks?

Yes. The allowed_models and excluded_models settings accept wildcard patterns, and you can save the same lists as account defaults on your workspace's Routing page. If your restrictions leave no eligible model, the request fails with a 404 naming the fields to change.

What happens if a provider goes down mid-request?

OpenRouter's provider routing retries other providers automatically, and an attempt that errors is not charged for its model tokens. Only very high burst traffic is rejected by OpenRouter's own DDoS protection with a 429. You can also configure fallback models, so a failed request tries your next choice instead of returning an error.

Does OpenRouter train on my data?

No. OpenRouter says it does not train on customer data. Provider-side retention can be disabled for the account or per request, and zero data retention routing is available on every plan, including free. The new router was verified on production workloads while respecting zero data retention throughout, according to OpenRouter.

Do I still need to follow new model releases?

That is the point of the router: every one of the 600-plus models on OpenRouter enters the consideration set the day it is added, and benchmarks and popularity signals decide whether it earns a place in the cost-optimal set. You choose once and the router tracks releases for you.

OpenRouter's rebuilt Auto Router and the new Auto Code router solve a real problem: model choice has become a chore, and the wrong pick wastes money or quality on every request. The benchmarks show genuine gains, particularly on agentic and tool-use tasks, and the router itself adds nothing to your bill. The catch is that the new default is more expensive per task than the router it replaces, so the saving is against calling a frontier model directly, not against last month's Auto Router. Try it with cost_tier set deliberately, watch the model field on your first hundred requests, and let the routing data, not the headline, decide whether it stays.

Sources

• Introducing the New Auto and Auto Code Routers — OpenRouter announcement, November 6, 2026
• Auto Router guide — OpenRouter docs, updated November 6, 2026
• Pricing — OpenRouter
• Model routing benchmarks — OpenRouter
• Auto Code model page — OpenRouter
• AI Compass, Friday 6 November 2026 — independent launch roundup

Exchange rates: 1 USD = 0.7471 GBP and 1 USD = 1.4117 CAD, November 7, 2026 market rates, verified at publication. Converted figures are approximations.