OpenRouter rankings in production: flash defaults, coding-agent traffic, when to escalate

Updated Oct 2, 2026. Weekly rankings through Oct 1: Space Bunny Alpha #1 (~30.9T, +951%), DeepSeek V4.1 Flash #2 (~23.5T, +24%), GLM 5.3 Flash #3 (~10.1T, −47%), MiMo-V2.6-Flash #4 (~9.34T, +412%), GPT-5.6 Luna #5 (~6.96T, −20%), DeepSeek V4 Flash 0731 #6 (~6.83T, −18%), Hy4 preview #7 (~6.81T, −45%). Today (Oct 1): Space Bunny Alpha #1 (~5.17T). Apps still Hermes / Kilo / Codex / Cline heavy. Routing rule unchanged — rankings. Related: Plugin4Shell draft, Batch API draft, Descript model-eval queue, shell + Files API draft.

Mid-September OpenRouter usage is a useful production signal — not because the leaderboard is “truth,” but because it shows where tokens actually go when people ship. Through roughly Sep 16, 2026, two patterns are loud: coding and agentic apps dominate app-side volume, and flash / cheap models dominate model-side volume. Benchmark leaders are a different list. Adoption is not quality.

This expands the earlier stub into a routing note I can use on real stacks: default cheap, escalate on purpose, measure failover, don’t romanticize the model.

Updated Sep 18, 2026. OpenRouter usage through Sep 17 still puts GPT-5.6 Luna at #1 by tokens (~15.8T, ~12%) and DeepSeek V4.1 Flash at #2 with a huge week-over-week surge (>999% on the board), followed by Hy4 preview and GLM 5.3 Flash. Routing rule unchanged: flash default, escalate on purpose — rankings.

What the board is saying right now

On the OpenRouter rankings (usage through mid-Sep 2026), top apps by tokens skew hard toward agentic coding: Hermes Agent (Nous), Claude Code, Kilo Code, Cline — the kind of workloads that burn tokens on tools, retries, and long contexts. That matters if you route for production phone or ops agents: the public “what people buy” curve is not a chat demo curve.

Top models by tokens look different from the apps: GPT-5.6 Luna, Tencent Hy4 Preview, GLM 5.3 Flash, and the DeepSeek V4 Flash family. Volume lives in flash / low-cost tiers. Separately, benchmark leaders (Claude Fable 5.1, Qwen3.8 Max, GPT-6 Astra, and peers) are not the same as usage leaders. If your routing policy chases the eval crown by default, you’re optimizing a different game than the one paying the bill.

Production rule: flash is the default, not the apology

For intake, classification, short tool turns, and most “keep the conversation moving” steps, a flash model is the right call until a turn proves otherwise. I treat “best model” as an expensive default — the kind of default that hides bad prompts and bad tool schemas behind a bigger invoice.

Escalation criteria I actually want logged:

Ambiguity. Caller/user intent still unclear after one clarifying turn.
Money or irreversible action. Booking cancel, refund path, contract-ish language.
Tool failure. Schema miss, empty CRM hit, or retries already burned.
Handoff prep. Warm transfer summary that a human will trust in two seconds.

If none of those fire, stay on flash. If they do, escalate once — then fall back. Sticky “always Opus from here” sessions are how cost curves become folklore.

Coding-agent traffic is a routing hint, not a product brief

Hermes / Claude Code / Cline-style volume tells you that agent loops (plan → tool → observe → retry) are where tokens go. That rhymes with inbound voice work: the expensive path is rarely the greeting. It’s the tool chain and the recovery when the tool chain lies.

So I budget routing for loops, not one-shots. Cheap model for the loop body; stronger model for “decide to escalate / rewrite the handoff / kill the path.” Same idea as coding agents — just fewer PRs and more phone minutes.

What I log per turn (minimum)

Route chosen (model id + reason code: default / escalate / failover).
Latency to first useful token/audio and to tool result.
Tool success / empty / error — not just HTTP 200.
Whether a human re-asked anything after transfer.
Cost for the turn and for the session so far.

Without those five, OpenRouter (or any router) is just a prettier API key. With them, rankings become a weekly sanity check: am I still defaulting where volume lives, and escalating where failures live?

Failover that’s boring on purpose

Primary flash → secondary flash from a different provider → escalate model → graceful degrade (scripted apology + callback / human). No romantic “try the smartest thing three times.” Timeouts should be short enough that callers don’t hear you thinking about vendor politics.

Chinese-origin flash models carrying a huge share of OpenRouter volume is real; so is the need for a non-single-region failover story if your product can’t eat a provider blip. Pick two cheap paths that fail differently, not one hero model.

Side signal: open voice tooling is catching up

Separate from rankings, GitHub daily trending still throws voice-adjacent repos into the feed — e.g. open-source voice studios like jamiepine/voicebox for clone / dictate / create workflows. Useful as a lab for listen-audio and agent voice experiments; not a substitute for telephony evals. The production bar for me stays the same as in the phone-stack teardown: latency, transfers, naturalness under real speech.

Closing

Use OpenRouter rankings as a map of where tokens already go. Default flash. Escalate on ambiguity, irreversible actions, tool failure, and handoff quality. Log the route reason. Treat benchmark kings as optional upgrades, not the home screen.

Related drafts: Agents API + GPT-Live, ripping out the AI phone stack. Comparing notes beats comparing logos. Say hello.