NeuralDash — Energy Board
Live Neuralwatt energy-per-request, throughput, TTFT & Flex pricing inside VS Code.
[!NOTE]
Unofficial. NeuralDash is an independent community extension and is not affiliated with or endorsed by Neuralwatt. It reads only Neuralwatt's public pricing pages and the anonymous /v1/models catalog; it sends no data and needs no account or API key.
- ⚡ Live Energy & Perf Board —
Right now and 7d typical energy (μWh / mWh / Wh), TTFT p50/p95, throughput tok/s, and 7d cache hit %.
- 🔀 3-Mode Sorting — Bang for Buck (
throughput p50 ÷ energy), Fastest, and Lowest Price across both main models and Flex tiers.
- 📈 48h Trend Telemetry Bento — sidebar SVG charts for energy & throughput with average-colored KPIs, right-side indicator tags, and tracking-alias resolution (
alias → backend).
- 🧠 Thinking Effort & Same-Weights Notes — reasoning availability, default effort, and accepted
reasoning_effort levels from /v1/models.
- 🏷️ Live Flex Tiers — flex membership, subsidized energy, and pricing provenance enumerated at runtime with zero hardcoded data.
Contents
Installation
Install NeuralDash from the Open VSX Registry, or download the latest .vsix from GitHub Releases and install via Extensions: Install from VSIX.... Requires VS Code 1.85.0 or newer.
Features
- Tree view
NeuralWatt → Energy Board lists all models with mode-aware metrics (Bang for Buck, Fastest, Lowest Price). Energy is shown in μWh, mWh or Wh to fit the value. Models the board publishes no energy for sit in a collapsed No energy data section.
- Performance metrics — parses live first-token latency (TTFT p50/p95 with fast/mid/slow dot), throughput (p50/p95 tok/s), 7d cache hit %, and Flex dispatch wait from
https://portal.neuralwatt.com/energy-pricing?view=performance.
- 48h trend view — click any model or Flex row (or use the tooltip link) to show the sidebar
48h Trend Telemetry Bento with average-colored KPIs, 48-hour energy & throughput SVG charts with right-side indicator tags, and tracking-alias resolution (alias → backend).
- Rich, sticky tooltips: compact per-band table, colored 48h trend summary, TTFT/throughput/cache hit %, benchmark value, plus
Copy Slug and Open Pricing Page actions.
- Thinking line on model and Flex tooltips, read from Neuralwatt's
/v1/models catalog: whether the model thinks, whether it's always on, on by default or off by default, the default effort, and the effort levels it accepts (verbatim, ready for reasoning_effort). A lane with thinking off also shows what other lanes on the same weights do. Models the catalog doesn't list show no line.
- 3-mode sort (
NeuralDash: Set Sort Order): Bang for Buck (throughput p50 ÷ energy, default), Fastest (throughput p50 desc, TTFT tiebreak), and Lowest Price (energy asc, cost tiebreak). Applies to both main models and Flex tiers.
- Status bar — shows
neuraldash.myModel when set, or pins the #1 model under the active sort mode as ⚡ <name> · <p50> tok/s · <energy> (<mode>).
- Flex tiers — Neuralwatt's flex-tier models retrieved live from the catalog, sorted by the active mode and inspectable in the
48h Trend view at subsidized energy.
- Polling with SHA-256 change detection, 30s timeout, failure backoff (cap 120 min).
Note: NeuralDash is unofficial and reads the public NeuralWatt energy-pricing page directly (energy pricing is HTML-only; the model catalog comes from Neuralwatt's public /v1/models API). If the page's structure changes, parsing may need an update — the board keeps serving its last cached data meanwhile.
Configuration
| Setting |
Default |
Description |
neuraldash.pollIntervalMinutes |
20 (min 5) |
Poll interval in minutes |
neuraldash.tariffPerKWh |
10.00 |
USD per kWh for cost display |
neuraldash.preferredBand |
typical |
Basis for ranking: typical | auto | 0-256 | 256-1k | 1k-4k | 4k-16k | 16k-64k | 64k-256k | 256k-1M |
neuraldash.myModel |
"" |
Model slug shown in status bar |
neuraldash.scoresOverridePath |
"" |
Absolute path to JSON overriding data/scores.json (user-wins per slug) |
Scores
Live scores come from the Epoch Capabilities Index (ECI), read from Epoch AI's benchmark data download (https://epoch.ai/data/benchmark_data.zip, epoch_capabilities_index/eci_scores.csv) at most once a day. Data © Epoch AI, used under CC BY 4.0. Board models are matched to ECI rows by normalized name (slug, then display name, then family); ambiguous or unknown models stay no benchmark. Each tooltip names the ECI row it used.
Bundled and override scores take precedence over live ECI scores. data/scores.json schema:
{
"version": 1,
"scores": {
"deepseek-v4-flash": { "performanceScore": 82.5, "source": "https://example.com/bench", "asOf": "2026-08-20" }
}
}
performanceScore: 0–100 (higher is better).
- Only honest entries with a real public
source and asOf date. Empty scores: {} is valid — models without a benchmark show no benchmark.
- Override file at
scoresOverridePath is merged per-slug, user-wins. Missing/malformed override falls back to bundled.
Commands
NeuralDash: Refresh Board
NeuralDash: Set Sort Order (Bang for Buck / Fastest / Lowest Price)
Pick Best Value
Set My Model
Open 48h Trend Chart (also via row right-click and the tooltip link)
Copy Model Slug (via the tooltip link)
Open Pricing Page
Flex tiers (live catalog)
The Flex tiers section lists Neuralwatt's flex-tier models, retrieved automatically on every board
cycle from https://api.neuralwatt.com/v1/models and the energy-pricing performance view
(per-model flex lanes + preview chips). Membership, pricing, and preview status are enumerated
from those documents — nothing is hardcoded, so the section follows the portal as models rotate.
Pricing provenance is labeled per row: exact alias → models.dev alias → N× derived (multiplier
harvested from the catalog) → portal /models card → TBD with a reason. The tooltip Source
row names the exact documents and fetch times, and NeuralDash: Debug Flex dumps the full
evidence chain. Catalog failures are silent — the board never waits on or reports them.
Development
npm install
npm run compile # tsc -p .
npm test # tsc -p . && node --test out/test/
npm run watch
npx @vscode/vsce package
Commits are scanned by local PII hooks; see CONTRIBUTING.md for setup.
| |