Summary: OpenRouter’s October 31, 2026 analysis of open-weight versus closed-weight models reports that open-weight models went from 27% to 72% of the top 50 models’ tokens in a year, at roughly a quarter the price — while the best closed models still score about 11 points higher. Their practical recommendation: start open-weight, test on your own tasks, and escalate to closed-weight only where quality changes the result. That default flips how many builders should architect their model calls.
What actually changed in the usage data
OpenRouter classified every model in its top 50 by weekly token volume, counting a model as open-weight if its catalog entry links to a Hugging Face repository. In the week of October 6, 2025, open-weight models handled 1.3 trillion tokens against 3.4 trillion for closed-weight. By the week of September 28, 2026, that was 83.5 trillion open-weight versus 33.0 trillion closed. The open share peaked at 77.5% in July 2026 and has stayed above two-thirds since.
Both categories grew — closed-weight volume rose almost tenfold — so this isn’t closed models dying; it’s open models growing 60-fold faster. MIT Sloan research covering May–September 2025 found closed models handled nearly 80% of tokens, and OpenRouter says its own late-2025 data looked similar. The shift happened in 2026.
On both sides, most volume goes to cheap, fast models rather than the highest-scoring ones. The most-used open-weight models over the last four weeks were DeepSeek V4.1 Flash, GLM-5.3 Flash, Hunyuan 4 Preview and DeepSeek V4 Flash; on the closed side, GPT-5.6 Luna led, followed by Gemini 3.8 Flash and GPT-6 Luna.
The price gap is competition, not magic
Over the last four weeks, the token-weighted list price for open-weight models was $0.28 per million input tokens and $1.62 out; closed-weight models averaged $1.27 in and $6.36 out — about four times as much. The reason is structural: anyone can serve an open-weight model, so providers compete on price for the identical model. On October 21, 2026, DeepSeek V4.1 Flash had 31 providers on OpenRouter with input prices from $0.03 to $0.60; GLM-5.3 Flash had 30 providers ranging $0.04 to $1.15. Claude Sonnet 5.5 had five providers, GPT-6 Luna three, Gemini 3.8 Flash two.
The categories also overlap. The cheapest Kimi K3 endpoint costs less than Claude Sonnet 5.5, but its priciest costs about twice as much, and GPT-6 Luna lists at $0.10 per million input tokens — less than any Kimi K3 provider. OpenRouter’s advice: compare specific models, not categories. The wide price ranges reflect provider differences in hardware, quantization, speed and data policy, so the cheapest endpoint isn’t always the one you want.
The quality gap hasn’t closed
On the Artificial Analysis Intelligence Index, the best closed-weight scores were Claude Opus 5.5 at 57.6 and Claude Sonnet 5.5 at 56.0. The best open-weight scores were MiMo-V2.6-Pro at 46.3, GLM-5.3 at 44.8 and Step 5 Preview at 43.7, with Kimi K3 at 43.6. Scores are taken at each model’s highest reasoning-effort setting, and OpenRouter notes that a June 2026 roundup put the gap at about five points on an earlier index version — so treat the 11-point figure as a snapshot, not a constant.
An 11-point quality gap against a 4x price gap is why OpenRouter’s default is to start open-weight and test on your own tasks. Move to closed-weight only for the steps where extra quality changes the result — typically the hardest step in a workflow, like agent planning, not every call.
Routing both in one app
OpenRouter serves both kinds through the same API, so you can mix them without rewriting your app: a closed-weight model first in the models array, an open-weight fallback after it. Fallbacks also trigger on rate limits, downtime, context-length errors and moderation refusals. A price cap is one routing filter — this request skips Claude Sonnet 5.5 and lands on GLM-5.3 Flash:
import os, requests
r = requests.post(
"https://openrouter.ai/api/v1/chat/completions",
headers={"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}"},
json={
"models": ["anthropic/claude-sonnet-5.5", "z-ai/glm-5.3-flash"],
"provider": {"max_price": {"prompt": 1, "completion": 4}},
"messages": [{"role": "user", "content": "In one sentence: what is an open-weight model?"}],
},
timeout=120,
)Remove the open-weight fallback and the same request fails with a 404, because no endpoint meets the cap. This is the same escalation-path thinking we covered in OpenRouter Fusion: what a compound model changes for your fallbacks: design for the step where you actually need frontier quality, and let cheaper models carry the rest.
One caveat before you build: open-weight doesn’t mean unrestricted. Licenses vary widely — DeepSeek V4 Flash shipped under MIT, while MiniMax M3’s community license requires attribution and asks larger commercial products to get written permission first. And since most open-weight models don’t release training data, “open weight” is the accurate term, even though search results and model pages say “open source.” Check the license, then run your own evals against your workload — that’s the decision the data supports.