Skip to content

Model-catalog knowledge update log

2026-09-07

  • Update: Live refresh via process:update-model-catalog from arena.ai/leaderboard/text (Sep 2, 2026; 400 models; 8.0M votes).
  • Attested — existing catalog scores moved to the live board where the name matched exactly, via punctuation / :free aliases, or as a same-model alias that already shared a score. Typical drift is 1–3 Elo. Named examples: muse-spark-1.1 1490 → 1492, glm-5.3-flash 1469 → 1474, glm-5.3-max 1484 → 1482, gemini-3.7-flash-high 1490 → 1491, nemotron-3.5-lightning 1348 → 1355.
  • New catalog entriesclaude-fable-5.1-max / claude-fable-5-1-max (shipping claude-fable-5.1 / claude-fable-5-1 estimated from max), gemini-3.8-flash-high (shipping gemini-3.8-flash estimated from high), granite-4.2-30b / granite-4.2-8b / granite-4.2-3b.
  • Not copied — dated snapshots onto a different dated/generic id; thinking / xHigh / max tiers onto the base id (muse-spark-1.2 (xHigh)muse-spark-1.2-xhigh only; claude-fable-5.1-max is not the shipping base id). deepseek-v3.1-terminus-thinking (1418) was split off the deepseek-v3.1 line (1417).
  • Estimates — left unchanged except the new shipping siblings (claude-fable-5.1, gemini-3.8-flash) and moving muse-spark-1.2 into auto-fill.
  • Update: Targeted addition of the shipping GPT-5.6 ids — no board fetch (the 2026-08-29 snapshot is still inside stale_after), so the attested list is untouched.
  • New catalog entriesgpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, and the gpt-5.6 alias. GA 2026-07-09; 1.05M context / 128K max output per the OpenAI models reference. Added because Kai now routes this family to the Responses API (issue #469) and the ids were falling back to the 100K default context window, which trimmed history far earlier than the model needs.
  • Estimates — each base id inherits its attested -xhigh sibling: Sol / the gpt-5.6 alias 1482, Terra 1466, Luna 1452. Recorded in the auto-fill block, not the attested snapshot: the board lists only the -xhigh tiers, and the quality-tier rule forbids attesting those onto a base id.
  • Correctiongpt-5.6-*-xhigh context windows moved 1_100_000 → 1_050_000 to match the documented figure and the rest of the family.

2026-08-29

  • Update: Live refresh via process:update-model-catalog from arena.ai/leaderboard/text (Aug 27, 2026; 395 models; 7.9M votes).
  • Attested — existing catalog scores moved to the live board where the name matched exactly, via punctuation / :free aliases, or as a same-model alias that already shared a score. Typical drift is 1–3 Elo. Named examples: qwen3.8-max 1491 → 1479, grok-4.6-high 1464 → 1461, claude-opus-5-max 1491 → 1488, muse-spark-1.2-xhigh 1499 → 1498, gemini-3.5-flash-high 1477 → 1479.
  • New catalog entriesclaude-opus-4-6-high / 4-7-high / 4-8-high (and 4.5/Sonnet 4.5 high-32k rows), gemini-3.7-flash-high (1490), gemini-3.6-flash-high (1481), glm-5.3-max (1484) / glm-5.3-flash (1469), deepseek-v4-pro-high-20260813 (1462), qwen3.8-27b (1436), inkling-small (1407), grok-3-mini-high (1364).
  • Not copied — dated snapshots onto a different dated/generic id; thinking / xHigh / high tiers onto the base id (muse-spark-1.2 (xHigh)muse-spark-1.2-xhigh only; Claude -high is not -thinking); unrelated ids that only shared a source line (gpt-4o-mini vs gpt-oss-20b).
  • Estimates — left unchanged except where an id was newly attested (grok-3-mini-high, qwen3.6-plus-free, minimax-m2.5-free, nvidia-llama-3.3-nemotron-super-49b-v1.5).

2026-08-12

  • Update: Live refresh via process:update-model-catalog from arena.ai/leaderboard/text (Aug 12, 2026; 390 models; 7.8M votes).
  • Attested — existing catalog scores moved to the live board where the name matched exactly, via punctuation / :free aliases, or as a same-model alias that already shared a score. Typical drift is 1–3 Elo. Named examples: qwen3.8-max 1497 → 1491, kimi-k3-max 1485 → 1489, claude-opus-5-max 1488 → 1491, muse-spark-1.2-xhigh 1498 → 1499.
  • New catalog entriesgrok-4.6 / grok-4.6-high (1464), muse-glimmer (1426), solar-pro4 (1378), nemotron-3.5-lightning (1350).
  • Not copied — dated snapshots onto a different dated/generic id; thinking / xHigh tiers onto the base id; unrelated ids that only shared a source line (gpt-4o-mini vs gpt-oss-20b, command-r vs command-r-08-2024, glm-4-plus vs glm-4-plus-0111).
  • Estimates — left unchanged except where an id was newly attested.
  • Initialization: Created OKF model-catalog bundle (arena scores + matching policy). Runtime source of truth remains ModelCatalog.kt.