{
  "asOf": "2026-09-06",
  "discovery": "Exa API search followed by direct primary-page content retrieval",
  "scope": "Sources added during the final edit: the Fable 5 redeployment, the GLM history and architecture, and GPT-6 Astra pricing. Chart snapshots and earlier pricing observations are unchanged.",
  "sources": [
    {
      "id": "fable-5-redeployment",
      "url": "https://www.anthropic.com/news/redeploying-fable-5",
      "title": "Redeploying Claude Fable 5 \\ Anthropic",
      "published": null,
      "kind": "vendor announcement",
      "observation": "Fable 5 launched on 9 June 2026 with the strongest safeguards Anthropic had applied to a model. US export controls suspended access from 12 June and were lifted on 30 June. Anthropic then deployed a stricter classifier, which it states flags benign coding and debugging requests more often.",
      "conditions": [
        "The essay cites this page as an example of safeguards being a lever labs adjust. The page documents safeguards being tightened and access being restored, not safeguards being loosened."
      ],
      "retrievedAt": "2026-09-06",
      "retrieval": "Exa content retrieval",
      "extractedTextSha256": "bfd07bcbe27fc8260e34147ade27d2fe6495dbe61fba30c5d6df3ee4feb87e6e"
    },
    {
      "id": "glm-history",
      "url": "https://www.interconnects.ai/p/glm-53-how-chinese-labs-keep-stride",
      "title": "GLM-5.3: How Chinese labs keep stride with the frontier",
      "published": "2026-08-14T21:23:35.000Z",
      "kind": "analyst commentary",
      "observation": "Nathan Lambert dates Zhipu AI's founding to 2019, the first GLM release to March 2021 from Tsinghua's THUDM group, and ChatGLM to March 2023. He reads GLM-5.3 as extended post-training on the GLM-5.2 base rather than distillation, and notes Z.ai's ties to Tsinghua University.",
      "conditions": [
        "The article concerns GLM-5.3, not the Flash variant plotted in the essay.",
        "Founding and release dates are the author's compilation, not a primary corporate record."
      ],
      "retrievedAt": "2026-09-06",
      "retrieval": "Exa content retrieval",
      "extractedTextSha256": "f53405d450e7e3b461390c60510717400c8c9cd1ed706538f05b6f919160b57d"
    },
    {
      "id": "glm-5-3-flash-architecture",
      "url": "https://docs.z.ai/guides/vlm/glm-5.3-flash",
      "title": "GLM-5.3-Flash",
      "published": null,
      "kind": "vendor documentation",
      "observation": "Z.ai documents GLM-5.3-Flash at 320B total parameters with 18B activated, combining sparse and linear attention, with video, image, text and file input, a 1M context window and 128K maximum output.",
      "conditions": [
        "Parameter counts and the attention claim are the vendor's own statements."
      ],
      "retrievedAt": "2026-09-06",
      "retrieval": "Exa content retrieval",
      "extractedTextSha256": "74f7b435f2b8c31a74e96311b9c6485f7e80c879f08a8b9cb5d1f334f3c6d42c"
    },
    {
      "id": "astra-pricing",
      "url": "https://developers.openai.com/api/docs/models/gpt-6-astra",
      "title": "GPT-6 Astra Model | OpenAI API",
      "published": null,
      "kind": "published model prices",
      "observation": "GPT-6 Astra lists USD 10 per million input tokens and USD 50 per million output tokens. Prompts above 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request. Batch and Flex are 50% of standard rates. Fast mode is 2x the applicable rates.",
      "standardUsdPerMillionTokens": {
        "input": 10,
        "output": 50
      },
      "longContextUsdPerMillionTokens": {
        "input": 20,
        "output": 75
      },
      "conditions": [
        "Cached input and cache write rates differ and are not quoted in the essay.",
        "The pricing page row for gpt-6-astra corroborates the same figures."
      ],
      "retrievedAt": "2026-09-06",
      "retrieval": "Exa content retrieval",
      "extractedTextSha256": "12dc91589cecbb8c5a55d7cc6f8945ce204de431ad39c6497c874dada4be7c96"
    }
  ]
}