Sports tech analytics is mostly scoreboards, and that is the honest story

Search for sports technology data analytics and you will find a market forecast, a gambling story and a lot of vendor copy. What you will not find, in the material available here, is a single sports organisation publishing how its analytics actually work.

That absence is the finding. The most concrete, verifiable documents in this space describe general purpose data tooling. The sports connection is either historical, commercial or implied. It is worth being blunt about that before repeating a growth number.

The one sports app in the pile does almost no analytics

Slate published a long appreciation of Apple's Sports app in June, and the praise is specifically for what the app refuses to do. There are no news headlines, no streaming guides, no design features built to keep you scrolling. You get team names and logos, records, game times and scores, plus a standings page. Betting odds appear on the scoreboard by default but can be switched off permanently in settings, the piece says. The author had to disable automatic Pittsburgh Penguins notifications during the Stanley Cup Playoffs because the app was beating the streaming feed to the result.

That is a latency claim, not an analytics claim, and it is the most technical thing in the review. The article does not discuss expected goals, tracking data, injury models or anything else that would sit under a sports analytics heading. Its argument is that a product which does one job well beats a product stuffed with features users did not ask for.

Read that against the current fashion in sports apps and the contrast is sharp. Major League Baseball's app has surrounded scores with what the author calls window dressing. ESPN's app now foregrounds its live TV offering and a vertical video feed called Verts. Both are analytics businesses in the sense that they measure you. Neither is selling analysis of the game.

The benchmark that actually measures analytical reasoning

Hex published DataBench, a leaderboard for what it calls agentic analytics, on 13 August. The table is dense and the numbers are the story. Opus 5.5 at Max effort scores 70.5% at an average cost of $3.57 per task. GPT-6 Sol at XHigh scores 61.3% for $0.58 per task. GPT-6 Luna at Max scores 52.3% for six cents.

Those three data points alone describe the trade every data team is now making. The top of the leaderboard costs roughly sixty times more per task than the cheap model that gets about three quarters of the way there. Hex also notes that it updated its judges and rubrics to stop accidentally rewarding misleading responses. That change, it says, made the benchmark significantly harder overall, with Anthropic models especially affected. Opus 5.5 still sits at the top after that change.

The same page carries latencies. Opus 5.5 at Max takes 677.2 seconds per task on average. GPT-6 Luna at Max takes 353.2 seconds. GPT-6 Astra at Low finishes in 88.5 seconds with a 55% score. For anyone building an interactive analytics product, that is the real constraint: the best answer may be minutes away, and the fast answer may be wrong in ways nobody checks.

What the plumbing looks like when you build it yourself

Two open source projects in the same window show the other half of the problem. Adaca, a software consultancy, published a self-hosted replacement for Google Analytics 4 on Cloudflare Workers on 8 September. Daily rollups arrive from the GA4 Data API or a BigQuery export and land in a D1 database the operator owns. Realtime stays on Google. Six dashboards ship out of the box, every number opens into a detail page, and reports can go out weekly or monthly by email or Slack. Setup is deliberately manual: create a Google service account with Viewer access, download the JSON key, press Deploy, put Cloudflare Access or basic auth in front of it.

There is no sign-in, which is either the point or a hazard depending on who is reading the README. It is MIT licensed.

Separately, Lark is a real-time database server written mainly in Rust on Glommio, with a Go edge component terminating WebSocket, WebTransport and REST connections. It is designed to be drop-in compatible with Firebase Realtime Database SDKs, so an app can keep Firebase Auth, Hosting and Storage while moving the realtime tree elsewhere. The repository says the on-disk format will not be broken without a migration path. It also says the 0.x version number reflects the project's age rather than known instability. Lark Cloud runs the same codebase for paying customers.

Why cheaper storage does not settle the argument

Cassis published a piece on 23 August arguing that AI agents do not remove the human cost of making data useful, they redistribute it. The framework it borrows is schema-on-write against schema-on-read, and the working example is a CFO asking how many active enterprise customers renewed last quarter. Someone still has to decide which table holds customers, what active means, what enterprise means and when the quarter ends.

The post splits that cost into three parts: indexing, querying and reliability. A dashboard pays indexing cost upfront, sometimes weeks of work before the first chart renders, and is reliable for the questions it was designed to answer. Ad hoc requests push the whole cost onto a data team at retrieval time. Cassis cites a senior data manager describing a weekly rotation that consumes about 20% of a team of five or six people, permanently.

The question is who does the thinking, and when.

That sentence is the most useful thing in the pile for anyone working in sports. A club tracking player load, a league selling broadcast inventory and a team's commercial department counting renewals all face the same decision about where to put the effort. The tools are improving and the per-task price of machine analysis is falling fast. None of that decides what active means.