{
  "@context": "https://schema.org",
  "@id": "https://aggr.aymericbeaumet.com/items/ampcode-com/2026-06-18-a-faster-librarian/#webpage",
  "@type": [
    "WebPage",
    "ArchiveComponent"
  ],
  "articleBody": "The [Librarian](https://ampcode.com/news/librarian) is now ~3x faster and 43% cheaper, with the same quality.\n\nIt now runs on GPT-5.5 (no reasoning) with [websocket mode](https://ampcode.com/news/faster-deep-rush) and an updated system prompt that encourages more parallel exploration. The Librarian fires ~8 tool calls in parallel per turn, up from ~3 with Sonnet, and wraps up a search in ~5 turns instead of ~15.\n\nIn our internal eval, about a quarter of that speedup comes from OpenAI's websocket mode and the rest from switching to GPT-5.5 with no reasoning:\n\n|                       | Sonnet-4.6 (medium) | GPT-5.5 (none)    |\n| --------------------- | ------------------- | ----------------- |\n| Latency (mean)        | 237s                | 81s (2.9x faster) |\n| ↳ gain from websocket | —                   | \\~1.3x            |\n| ↳ gain from model     | —                   | \\~2.2x            |\n| Quality (F1, mean)    | 0.47                | 0.48              |\n| Average cost          | $1.21               | $0.69             |\n\nHere's a comparison:\n\n> How does Kubernetes' HorizontalPodAutoscaler handle missing pod metrics when scaling down — does it assume missing pods are at 100% of their resource requests, or 100% of the target utilization? Cite the function and logic in the source.\n\nSonnet 4.6 (left) took 2 minutes and cost $1.08, while GPT-5.5 (right) took 40 seconds and cost just $0.47.\n",
  "authors": [],
  "captured_at": "2026-10-03T23:09:50.514863041+00:00",
  "category": "programming",
  "content_html": "<p>The <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://ampcode.com/news/librarian\">Librarian</a> is now ~3x faster and 43% cheaper, with the same quality.</p>\n<p>It now runs on GPT-5.5 (no reasoning) with <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://ampcode.com/news/faster-deep-rush\">websocket mode</a> and an updated system prompt that encourages more parallel exploration. The Librarian fires ~8 tool calls in parallel per turn, up from ~3 with Sonnet, and wraps up a search in ~5 turns instead of ~15.</p>\n<p>In our internal eval, about a quarter of that speedup comes from OpenAI's websocket mode and the rest from switching to GPT-5.5 with no reasoning:</p>\n<table>\n<thead>\n<tr>\n<th></th>\n<th>Sonnet-4.6 (medium)</th>\n<th>GPT-5.5 (none)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Latency (mean)</td>\n<td>237s</td>\n<td>81s (2.9x faster)</td>\n</tr>\n<tr>\n<td>↳ gain from websocket</td>\n<td>—</td>\n<td>~1.3x</td>\n</tr>\n<tr>\n<td>↳ gain from model</td>\n<td>—</td>\n<td>~2.2x</td>\n</tr>\n<tr>\n<td>Quality (F1, mean)</td>\n<td>0.47</td>\n<td>0.48</td>\n</tr>\n<tr>\n<td>Average cost</td>\n<td>$1.21</td>\n<td>$0.69</td>\n</tr>\n</tbody>\n</table>\n<p>Here's a comparison:</p>\n<blockquote>\n<p>How does Kubernetes' HorizontalPodAutoscaler handle missing pod metrics when scaling down — does it assume missing pods are at 100% of their resource requests, or 100% of the target utilization? Cite the function and logic in the source.</p>\n</blockquote>\n<p>Sonnet 4.6 (left) took 2 minutes and cost $1.08, while GPT-5.5 (right) took 40 seconds and cost just $0.47.</p>\n",
  "content_markdown": "The [Librarian](https://ampcode.com/news/librarian) is now ~3x faster and 43% cheaper, with the same quality.\n\nIt now runs on GPT-5.5 (no reasoning) with [websocket mode](https://ampcode.com/news/faster-deep-rush) and an updated system prompt that encourages more parallel exploration. The Librarian fires ~8 tool calls in parallel per turn, up from ~3 with Sonnet, and wraps up a search in ~5 turns instead of ~15.\n\nIn our internal eval, about a quarter of that speedup comes from OpenAI's websocket mode and the rest from switching to GPT-5.5 with no reasoning:\n\n|                       | Sonnet-4.6 (medium) | GPT-5.5 (none)    |\n| --------------------- | ------------------- | ----------------- |\n| Latency (mean)        | 237s                | 81s (2.9x faster) |\n| ↳ gain from websocket | —                   | \\~1.3x            |\n| ↳ gain from model     | —                   | \\~2.2x            |\n| Quality (F1, mean)    | 0.47                | 0.48              |\n| Average cost          | $1.21               | $0.69             |\n\nHere's a comparison:\n\n> How does Kubernetes' HorizontalPodAutoscaler handle missing pod metrics when scaling down — does it assume missing pods are at 100% of their resource requests, or 100% of the target utilization? Cite the function and logic in the source.\n\nSonnet 4.6 (left) took 2 minutes and cost $1.08, while GPT-5.5 (right) took 40 seconds and cost just $0.47.\n",
  "dateCreated": "2026-10-03T23:09:50.514863041+00:00",
  "date_created": "2026-06-18T00:00:00+00:00",
  "date_modified": "2026-06-18T00:00:00+00:00",
  "date_published": "2026-06-18T00:00:00+00:00",
  "description": "The Librarian is now ~3x faster and 43% cheaper, with the same quality. It now runs on GPT-5.5 (no reasoning) with websocket mode and an updated system prompt that encourages more parallel exploration. The Librarian fires ~8 tool calls in…",
  "external_url": "https://ampcode.com/news/a-faster-librarian",
  "git": {
    "commit": "ebb0941c5c06a2ab2377342d607535262a3e969e",
    "path": "items/ampcode-com/2026/06/2026-06-18-a-faster-librarian.md",
    "permalink": "https://github.com/aymericbeaumet/aggr-instance/blob/ebb0941c5c06a2ab2377342d607535262a3e969e/items/ampcode-com/2026/06/2026-06-18-a-faster-librarian.md"
  },
  "id": "urn:aggr:item:37c653e61402cbc2720ce10af32b70fa536595dc",
  "isBasedOn": {
    "@id": "https://ampcode.com/news/a-faster-librarian"
  },
  "mainEntity": {
    "@id": "https://ampcode.com/news/a-faster-librarian",
    "@type": "CreativeWork",
    "archivedAt": {
      "@id": "https://aggr.aymericbeaumet.com/items/ampcode-com/2026-06-18-a-faster-librarian/#webpage"
    },
    "dateModified": "2026-06-18T00:00:00+00:00",
    "datePublished": "2026-06-18T00:00:00+00:00",
    "headline": "A Faster Librarian",
    "timeRequired": "PT1M",
    "url": "https://ampcode.com/news/a-faster-librarian",
    "wordCount": 175
  },
  "name": "Archived snapshot: A Faster Librarian",
  "preview": null,
  "reading_minutes": 1,
  "replicated_at": null,
  "source": "ampcode-com",
  "source_name": "Amp News",
  "summary": "The Librarian is now ~3x faster and 43% cheaper, with the same quality. It now runs on GPT-5.5 (no reasoning) with websocket mode and an updated system prompt that encourages more parallel exploration. The Librarian fires ~8 tool calls in parallel per turn, up from ~3 with Sonnet, and wraps up a search in ~5 turns instead of ~15. In our internal eval, about a quarter of that speedup comes from OpenAI's websocket mode and the rest from switching to GPT-5.5 with no reasoning: Sonnet-4.6 (medium) GPT-5.5 (none) Latency (mean) 237s 81s (2.9x faster) ↳ gain from websocket — ~1.3x ↳ gain from model — ~2.2x Quality (F1, mean) 0.47 0.48 Average cost $1.21 $0.69 Here's a comparison: How does Kubernetes' HorizontalPodAutoscaler handle missing pod metrics when scaling down — does it assume missing pods are at 100% of their resource requests, or 100% of the target utilization? Cite the function and logic in the source. Sonnet 4.6 (left) took 2 minutes and cost $1.08, while GPT-5.5 (right) took 40 seconds and cost just $0.47.",
  "tags": [],
  "temporalCoverage": "2026-10-03T23:09:50.514863041+00:00",
  "title": "A Faster Librarian",
  "url": "https://aggr.aymericbeaumet.com/items/ampcode-com/2026-06-18-a-faster-librarian/",
  "word_count": 175
}