The Most-Used Coding Model Is Not the Benchmark King
OpenCode's usage board is led by cheap open models, not benchmark kings. How I read usage, retention, and price before switching models.
TL;DR
Benchmark leaders like GPT-5.6 top SWE-Bench, yet deepseek-v4-flash leads OpenCode's real-world usage, suggesting cost, speed, and context length matter more than peak scores. The author recommends checking usage volume, momentum, and weekly retention as satisfaction signals. Treat official model lists as a starting point, then validate against usage data and a small trial in your own repo before committing.
The number one spot on OpenCode's data page is not Claude. It is not GPT either. deepseek-v4-flash leads with 25T tokens, up 23 percent [1]. I stared at that for a while, because my old habit of picking models ran exactly the other way: open the benchmark, take the top row, done.
When the benchmark king is not the usage king
My first guess was that the usage data must be the weird one. Cross-checking benchmark boards only sharpened the contrast. Vellum lists GPT-5.6 Sol at 96.2% on SWE-Bench, GPT-5.6 Luna at 93%, MiniMax M3 at 80.5% [2]. Meanwhile real-world usage on OpenCode is led by deepseek-v4-flash (25T), muse-spark-1.2-contributor (13T, up 26%), and mimo-v2.5 (9.6T, down 25%) [1]. The priciest, smartest model is not automatically the one people trust with daily work. Those are two scoreboards measuring two different things: peak ability on controlled tasks versus what actually gets used.
That makes sense once you think of aggregate usage as voting with money and time. Cost per token, speed, and context length all shape the vote. GPT-5.6 Luna, for instance, is priced at $0.2 input and $1.2 output per one million tokens with a 1,050,000-token context [2]. For iterative work, numbers like these often matter more than a two-point benchmark gap.
Three signals I now read before switching models
First, volume plus momentum. deepseek-v4-flash is up 23% while mimo-v2.5 is down 25% [1]. Sharp declines usually have a reason, and I would rather not rediscover other people's problems. Second, weekly retention, the share of users who come back the following week. deepseek-v4-flash-vision-exp holds 89.3%, muse-spark-1.2-contributor 85.7%, glm-5.3-flash 84.5%, while mimo-v2.5 sits at 71.9% [1]. Retention is satisfaction, and hype cannot fake it for long; volume can spike for a week, retention cannot. Third, I still read benchmarks, but as a minimum quality bar rather than a destination. DeepSeek V4 Flash posts 91.6% on LiveCodeBench [2] while holding the number one usage spot [1], a combination far more convincing than a high score with no users.
Official recommendations are a starting point, not an answer
OpenCode's documentation says it supports 75+ providers through AI SDK and Models.dev, which is great for experimenting [4]. But its official recommended list is short and explicitly marked as not necessarily up to date [4]. One practitioner wrote that successful users tend to run Anthropic models or open-weight ones like Kimi K2.5 and MiniMax, not necessarily what the docs suggest [3]. I have been burned the same way by a listed model that kept stalling on my own repos. The practical takeaway: use the official list to narrow the field, then validate against usage data and a small trial in your own repo before committing.
One more thing that made me smile: Indonesia ranks ninth globally in OpenCode usage, 14T tokens or about 3% [1]. The local community is clearly not just watching.
All numbers here are a snapshot. The page changes constantly; I read it on September 3, 2026 around 11:38 UTC [1]. Six months from now the order could be fully reversed. That is the point: today's correct numbers guarantee nothing unless you carry the reading method with you.
My habit of verifying benchmark numbers by hand is written up in the benchmark verification protocol, and the operational side of saving model quota shows up in my cron pipeline notes.
Reading it this way is practical because there is no account and no key involved; you just open the page. It also breaks usage down by provider and geography, so one visit gives you three portraits at once. For a technical blog like mine, numbers that link back to an open source are mandatory; if a claim has no linkable source, I write it qualitatively or drop it.