ai-market-research

A self-updating intelligence dashboard tracking the AI model and tooling landscape, focused on the decisions an AI infrastructure–oriented developer actually has to make: which subscriptions to hold, which agent harness to run, what to self-host, and how the quota burn really compares across providers and reasoning efforts.

Live dashboard: https://projectious-work.github.io/ai-market-research/ Latest release: https://github.com/projectious-work/ai-market-research/releases/latest (v0.2.0)

---

What it tracks

1. Frontier models — subscriptions, API pricing, context windows, and benchmarks (with SWE-bench Pro front-and-center because SWE-bench Verified is contaminated). Full OpenAI GPT family, Anthropic Claude, Gemini, Mistral, Grok, DeepSeek, MiniMax.
2. Quota burn cross-matrix — model × reasoning-effort cost matrix. Baseline GPT-5.5 medium = 1.00×; surfaces the real cost delta of high/xhigh effort vs. lower tiers.
3. Agent harnesses — Claude Code, OpenCode, Codex CLI, Gemini CLI, Aider, Cline, Roo Code, Cursor, Windsurf, Goose, OMO, OpenClaw, CCR, Hermes.
4. Self-hosting — Vast.ai cloud-GPU configs, MacBook configs (M4 Pro 64 GB, M5 Max 128 GB, Air 32 GB), open-weight models (Gemma 4, Qwen 3, Llama 3.1, MiniMax M2.5, DeepSeek V3.2), quantization formats, frameworks (llama.cpp, vLLM, Ollama, MLX).
5. Strategy — derived recommendations spanning the above.

How it's built

• data/market-state.json — canonical JSON state (the truth).
• src/dashboard.template.html — single self-contained HTML scaffold with a __MARKET_DATA__ placeholder.
• src/scripts/build.py — substitutes the JSON into the template, producing dist/dashboard.html (~132 KB, fully self-contained, no runtime fetches).

The output is one static HTML file. No JavaScript framework, no build chain, no CDN dependency at runtime.

Quickstart

Requires python3, bash, git, and (for deploys) the gh CLI.

# Build the dashboard
bash src/scripts/build.sh

# Validate JSON + rebuild + sanity-check the artifact
bash src/scripts/release-check.sh

# Open dist/dashboard.html in a browser

Deploy

GitHub Pages, fed from the gh-pages branch via a local script (no GitHub Actions). Architecture: DEC-20260517_1455-DeftLynx.

bash src/scripts/deploy.sh

The script builds, stages dist/ as the Pages payload, pushes via git worktree to gh-pages, and idempotently enables Pages on first run. Authentication uses gh auth token, so no global git credential helper is required.

Release

bash src/scripts/release.sh 0.3.0

Runs release-check → annotated tag → push → deploy → gh release create with the dashboard-vX.Y.Z.html asset.

Repository layout

.
├── data/              # market-state.json + daily archives
├── src/
│   ├── dashboard.template.html
│   ├── dashboard-context.md   # project profile + tracked dimensions
│   ├── sources.md             # canonical URLs the briefing checks
│   ├── briefing-prompt.md     # the agent prompt that refreshes data/
│   └── scripts/{build,release-check,deploy,release}.sh
├── dist/              # built dashboard.html (gitignored output)
├── context/           # processkit project context (decisions, logs, …)
├── AGENTS.md          # provider-neutral agent instructions
└── LICENSE

Contributing

This is a personal-research tool maintained for one user's decision-making context, but the code is MIT-licensed — fork it freely if the structure is useful as a template for your own market-watching dashboards.

License

Unless otherwise noted, the copyright holder grants the MIT License for all versions of this repository, including historical commits and tags. The full license text is in LICENSE. © 2026 projectious.