OpenCode Data: What Real-World Usage Reveals About Model Choice (Category #AI4SDLC)

After the morning release of The Lightcone, which I wrote about, I found myself absorbed in OpenCode Data. Models are usually compared through benchmarks or impressions on social media. This dashboard exposes a different layer: what developers actually run, how many tokens they consume, what a session costs, and how their choices change over time. It needs to be read carefully, however. This is not a ranking of the best models; it is a behavioral snapshot of one product.

I explored it on July 26, 2026, using the Go and 2 Months filters. OpenCode Go is a separate $10-per-month subscription offering a curated set of open models. The company explicitly says that the plan was created primarily for an international audience. This slice therefore does not represent every OpenCode user, the entire coding-agent market, or developers in general. With that caveat, several observations stood out to me.

1️⃣ Subscription usage is growing

On June 1, Go models processed about 1.9 trillion tokens. By July 20–21, the daily total had reached 3.1–3.4 trillion. The number of daily model users rose over the same period from 91,000 to 167,000–168,000. I am excluding July 26 from the comparison: the dashboard had refreshed at 08:16 UTC, so the day was incomplete.

The chart also shows a clear weekly rhythm. On July 15, for example, usage reached 3.2 trillion tokens and 147,000 model users. On July 18–19, the figures fell to 2.4–2.1 trillion tokens and 119,000–117,000 users.

A similar pattern recurs on other weekends. That looks like work-related rather than purely recreational use of coding agents, although the chart itself cannot establish the cause.

2️⃣ Token volume and user counts tell different stories

DeepSeek V4 Flash leads by volume. In the episode, however, Jay V presents another view: roughly 38,000 unique users for Flash, 31,000 for DeepSeek Pro, and almost 30,000 for GLM-5.2. GLM is therefore already close to one of the DeepSeek models in reach, while the two DeepSeek variants together—and their more intensive use—produce a different picture when measured in tokens.

That does not mean Flash is simply “better.” Jay explains that it is so inexpensive that users switch to it as they approach their daily or weekly limits. Token popularity is a mixture of at least four factors: model capability, price, speed, and quota-spending strategy.

3️⃣ Average session cost can differ by almost 28 times

In the July 26 snapshot, an average DeepSeek V4 Flash session contains 7.9 million tokens and costs about $0.085. A GLM-5.2 session contains 4.6 million tokens and costs around $2.40.

A longer session on the inexpensive model therefore costs far less than a shorter session on the costly model. For a platform team, this matters more than an abstract price per million tokens. The budget is consumed not by the price list alone, but by the actual shape of sessions.

4️⃣ Coding-agent economics depend heavily on caching

The dashboard shows that 96% of DeepSeek V4 Flash input tokens are served from cache, compared with 80% for GLM-5.2.

In a typical agent session, the codebase, system instructions, and conversation history can form a long shared prefix. If the infrastructure can reuse it, millions of logical tokens do not have to be recomputed each time. Comparing models only by input and output prices is therefore insufficient.

5️⃣ The geography looks nothing like the familiar Silicon Valley picture

In this slice, China accounts for 18% of tokens, the United States 12%, Brazil 5%, and Germany, Japan, and Indonesia 4% each.

These are shares of tokens, not users or revenue. The sample is also shaped by the design of Go itself: an inexpensive plan, open models, and infrastructure intended for global access.

Caveats about the analysis

OpenCode publishes the data itself, and the page does not provide an independent audit or a complete aggregation methodology. Tokens are not equivalent to successfully completed tasks, code quality, or team productivity. A large user count may reflect widespread experimentation with a new model; high token volume may reflect a low price, long sessions, or intensive work by a small group. Finally, the dashboard is live, so any undated screenshot becomes stale quickly.

Conclusions

Real-world usage is more useful as a map of trade-offs than as a model competition. A company building an internal coding-agent platform needs similar slices: model, task type, success rate, number of human corrections, latency, session cost, cache share, and switches after a quota is exhausted. Model selection then becomes engineering portfolio management rather than an argument over a single winner.

#AI #AI4SDLC #Agents #Data #Metrics #Engineering #Software