# OpenCodeGoEvals

Benchmarks and plain-language notes for every model in the **OpenCode Go** catalog:
a task-by-task performance matrix, cost/value analysis, and a description of what each
model is actually targeted for (agentic work, coding, data analysis, image understanding,
deep research, high-volume chores).

Static site, no build step — it is plain HTML/CSS/ES modules, deployed to **Cloudflare Pages**.

## What's here

| Path | What it is |
|---|---|
| `site/index.html` | The site: performance matrix, price/value charts, model guide |
| `site/assets/data.js` | Model catalogue — pricing, context, modality, target descriptions, caveats |
| `site/assets/evals.js` | Benchmark scores (0–100) per model per task family |
| `site/assets/app.js` | Rendering logic for the matrix, charts and guide |
| `site/assets/style.css` | Styling |
| `MODELS.md` | The same model descriptions in Markdown |
| `scripts/sync-catalog.mjs` | Weekly refresh: models.dev catalogue + Artificial Analysis → `catalog.js`, `evals.js`, `snapshot.json` |
| `scripts/build-models-md.mjs` | Regenerates `MODELS.md` from the three data layers |
| `scripts/notify-new-models.mjs` | Files a Gitea issue when a new model needs a write-up |
| `scripts/deploy.sh` | Manual Cloudflare Pages deploy |
| `.gitea/workflows/weekly-eval.yml` | Weekly scheduled evaluation refresh |
| `wrangler.toml` | Cloudflare Pages project configuration |
| `.gitea/workflows/pages.yml` | CI: builds and deploys to Cloudflare Pages on every push to `main` |

## Data sources

- [OpenCode Go docs](https://opencode.ai/docs/go) — pricing, monthly allowances, privacy/retention, client support
- [OpenCode Data](https://opencode.ai/data) — catalog metadata, context windows, usage and retention stats
- [Artificial Analysis](https://artificialanalysis.ai) — benchmark scores and the Intelligence Index
- Vendor announcements — Google (Gemini 4 Argon), Reflection AI (Beam), Meituan (LongCat),
  MiniMax (M3.1 Flash), xAI (Grok 4.7), StepFun (Step 5), Qwen, Zhipu, Moonshot, DeepSeek, Xiaomi, Tencent

Data snapshot date: **2026-10-21**. Re-run the collection script to refresh.

## Publishing to Cloudflare Pages

### 1. Create the Pages project

```bash
npm install
npx wrangler login                       # or export CLOUDFLARE_API_TOKEN
npx wrangler pages project create opencodogoevals
```

### 2. Give CI the credentials

Add these as repository secrets (Gitea: *Settings → Actions → Secrets*):

| Secret | Value |
|---|---|
| `CLOUDFLARE_API_TOKEN` | Cloudflare API token with **Cloudflare Pages: Edit** (and Account Settings: Read) |
| `CLOUDFLARE_ACCOUNT_ID` | The Cloudflare account that owns the Pages project |

The workflow in `.gitea/workflows/pages.yml` then:

1. checks out the repo and installs Wrangler,
2. stages `MODELS.md` next to the site,
3. creates the Pages project if it does not exist yet (idempotent),
4. deploys `site/` on every push to `main`, on a manual *Run workflow*, and hourly.

A runner must be registered on the Gitea instance for the workflow to execute.

### 3. Deploy manually (optional)

```bash
./scripts/deploy.sh     # uses env vars or your wrangler login session
npm run deploy          # same thing, wrangler only
npm run dev             # local preview on http://localhost:8788
```

### Custom domain

Set it in Cloudflare Pages → *Custom domains*, then add a CNAME record pointing at
`<project>.pages.dev`.

## Local development

The site is static — open `site/index.html` directly, or use `npm run dev` for a
Cloudflare-compatible local server.

## The weekly cycle

Two workflows run in Gitea Actions:

| Workflow | Trigger | What it does |
|---|---|---|
| `weekly-eval.yml` | Mondays 05:23 UTC (cron), or manual | Re-scrapes the catalogue and benchmark scores, rebuilds `MODELS.md`, opens an issue for any new model needing a write-up, and commits the result |
| `pages.yml` | every push to `main` | Deploys `site/` to Cloudflare Pages |

The chain is: **schedule → data refresh → commit → Pages deploy**, so the live site
updates itself without anyone touching it. The weekly run is idempotent — if nothing
changed, nothing is committed and nothing is redeployed.

A run leaves behind a trail you can audit:

- `site/assets/snapshot.json` — when the data was taken, which models are new or gone,
  and every benchmark score that moved since the previous run
- the commit `chore: weekly evaluation refresh <date>` in the history
- the Actions run log for the week
- a Gitea issue when a new model appears in Go without a write-up

### What is automated vs. hand-written

| File | Owner | Contents |
|---|---|---|
| `site/assets/catalog.js` | auto-generated | every model in Go: pricing, context, modalities, release date |
| `site/assets/evals.js` | auto-generated | benchmark scores per model per task family |
| `site/assets/data.js` | **hand-written** | what each model is targeted for, caveats, tags |
| `site/assets/snapshot.json` | auto-generated | timestamp, added/removed models, score deltas |

When you want to change a description, edit `site/assets/data.js` and run
`node scripts/build-models-md.mjs`; the weekly refresh will never overwrite it.

### Running it by hand

```bash
node scripts/sync-catalog.mjs        # refresh catalogue + scores
node scripts/build-models-md.mjs     # rebuild MODELS.md
node scripts/notify-new-models.mjs   # file the "needs write-up" issue
```

### If the scheduler is not firing

Gitea only runs scheduled workflows when a runner is registered and awake. If Mondays
slip past, trigger it manually from the Actions tab (*Run workflow*), or run the three
commands above on any machine with network access and push the result.

## Caveats

- Task-matrix cells use the best published evaluation per task family; when a model has
  not been run on a benchmark the column falls back to another one, so the underlying
  benchmark can differ between rows in a column. Hover a cell to see which one it is.
- Free-tier models (MiniMax M3.1 Flash Free, Beam Free, LongCat 2.5 Preview Free) are
  limited-time offers.
- Contributor-tier models (Gemini 4 Argon, Muse Spark 1.3) may use your prompts and
  completions for training — check the Go docs' privacy table.
- No Go model generates images; they are all text-output.

## License

MIT — see [LICENSE](LICENSE).