Playbooks · Coding tasks
OpenRouter Auto Code: Opus quality, 74% cheaper?
Written November 8, 2026
OpenRouter's new coding router claims Opus 5.5-level scores at 74% lower cost per task — and changed what openrouter/auto defaults to overnight.
What shipped
On November 5, OpenRouter released two router IDs: openrouter/auto, a rebuilt general-purpose router, and openrouter/auto-code, a new coding-specific route. OpenRouter's announcement says the rebuilt router weighs eleven factors about each request, combined with live pricing, provider health, and cache status, and that Auto Code treats every request as part of an agent session, with model choices guided by coding-task pass rate, completion time, and cost per solved task. Auto Code starts in a higher cost tier than the general router and weighs speed more heavily, because responsiveness matters while you build. Your guardrails, zero-data-retention requirements, provider restrictions, allowed and excluded model lists, and BYOK keys still decide which models the router may pick from. There is no routing fee: each request is billed at the selected model's standard rate, and the response's model field tells you which model served it.
The headline claim is that Auto Code scores in line with Claude Opus 5.5 across five benchmarks while costing 74 percent less per task. The five are Deep SWE, SWE-Atlas QA, Terminal Bench 4.0, τ³-bench Banking, and MMMU Pro Vision. On OpenRouter's routing-benchmarks board, the aggregate comes to $1.64 per task for Auto Code against $6.43 for Opus 5.5 at high effort. The routing step itself is faster too: 0.65 seconds at p90 in OpenRouter's live test, down from 1.96 seconds. In the beta period, invalid tool calls ran at one-quarter the old rate, agentic responses were cut off half as often, and the router stayed on the same model for nearly nine in ten agent turns, which protects the prompt cache. Every one of those figures is OpenRouter's own measurement, on OpenRouter's own benchmarks. None of them has an independent test behind it yet.
The overnight change that matters more
The claim is the headline, but the change that bites first is the default. The openrouter/auto slug is unchanged, and the cost_tier, allowed_models, and excluded_models settings still work. What changed is what the slug defaults to. OpenRouter says the old router had drifted toward the cheapest models, leaving out many strong ones, so with this release it returned pricing closer to where Auto was a few months ago, to bring in a greater variety of models. Existing callers get the rebuilt router with no slug change, and the unchanged slug now defaults to a more expensive model mix. If you preferred the old behaviour, set cost_tier to low to get similar cost performance. That is the setting to check before your next invoice, not after it.
Two more facts shape what the router can do for you. The model pages advertise a two-million-token route ceiling, but a router is not one model: the context a request actually gets depends on the model that serves it, so treat the ceiling as a ceiling, not a promise. And Auto Code is not available through EU in-region routing; the router is available only in the global data region.
Strengths and weaknesses, as OpenRouter states them
The strength is the trade the router is built for. If your work is varied, if you need costs held steady, or if you do not want to track every model release, a router that picks per request and adapts when a session goes wrong is worth more than a single pinned model. OpenRouter says the router adapts when the same error comes back or tests fail after an edit, raising the odds of a switch instead of leaving a session stuck with a wrong early pick. It also says new models enter the consideration set the day they are added, with benchmarks and popularity signals deciding whether they earn a place in the cost-optimal set.
The weakness is one OpenRouter states itself: 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. Choosing Auto Code over a single model trades a little peak performance for consistent, cost-efficient performance over time. The benchmark claim is also narrow in a way the headline hides: five benchmarks, chosen by the vendor, scored by the vendor, with the cost figure measured per task on the vendor's setup. A router that scores in line with Opus 5.5 on those five may not do so on your codebase, your task mix, or your definition of done. And the router's judgment is only as good as its pool: if your guardrails narrow the pool to a few models, the router has little left to choose between.
The run that would settle it
The claim worth testing is the cost-per-solved-task one, and the test is simple. Take a frozen set of coding tasks with a fixed definition of done, run it through openrouter/auto-code, and run the same set through Opus 5.5 directly. Record three things per task: what it cost, whether the task passed, and which model the router actually picked. The third column is the one most people skip, and it is the one that tells you whether the savings come from smart routing or from serving cheaper models on tasks where the answer was easy anyway. Run the set more than once: a router that switches models between turns will not give the same answer twice, and a single pass hides that.
Until that run exists, treat the 74 percent figure as a vendor claim to check, not a price you can budget on. Start with a small, non-critical task set, pin cost_tier explicitly rather than trusting the new default, and read the model field on every response for a week before routing real work through it.
Limits
- This page is a read of OpenRouter's announcement, its Auto Code model page, and its routing documentation. No frozen-task run was done here, and no request was sent through either router.
- The 74 percent figure, the $1.64 and $6.43 board figures, the 0.65-second routing time, and the beta reliability figures are OpenRouter's own measurements. None of them was independently tested.
- The two-million-token figure is a route ceiling. The context a request gets depends on the model that serves it.
- Auto Code is not available through EU in-region routing.
Related playbooks
- Claude Code 2.1.328 makes auto mode the cloud default
- Antigravity CLI 1.5.4: ACP Auto mode
- Antigravity CLI 1.5.7: Auto becomes the default
- Run Sonnet with an Opus advisor in Claude Code
Sources
- Introducing the New Auto and Auto Code Routers — OpenRouter — the November 6 announcement: the 74 percent claim, the five benchmarks, the eleven factors, the cost_tier change, the beta reliability figures, and the no-fee shape
- Auto Code — OpenRouter — the Auto Code model page: the November 5 release date, the per-request session treatment, the EU routing exclusion, and the two-million-token ceiling
- Auto Router — OpenRouter documentation — how the router family works, the three members, and the global data region limit
Back to: Coding playbooks hub · Also: Run three independent lines of research at the same time in Claude Code, then combine the short summaries · Model prices: what the major models cost · Plan allowances: what each agent plan lets you do