Some sandbox hosts cannot reach api.together.xyz but can reach
openrouter.ai, which serves moonshotai/kimi-k3 through Together among other
hosts; a run can pin that host so the serving is unchanged. Row per
openrouter.ai/api/v1/models (read 2026-09-09); effort table gains the
lowercase slug.
The GLM-5.2 rows (Z.ai, Together, Fireworks, OpenRouter) still carried
128,000 although the model shipped with a 1M window in June; DeepSeek V4 Pro
on Together is 512,000, not 128,000; Kimi K3 on Together is 1,048,576, not
the rounded 1,000,000; Kimi K2.6 / K2.7 Code on Together are 262,144.
Sources, all read 2026-09-08: Together's /v1/models `context_length` for
every Together row (exact, per model), docs.z.ai for the native GLM-5.2 row,
openrouter.ai/z-ai/glm-5.2, fireworks.ai/models/fireworks/glm-5p2. Rows
without a source consulted today (DeepSeek native, Fireworks/OpenRouter
DeepSeek and Kimi K2.6) are left as they were.
Effect: with the 250,000 compaction cap these rows all now compact at
250,000 instead of 102,400. Tests updated with the new values.
`anthropic:claude-sonnet-5` and `anthropic:claude-fable-5-1` (the models in
active use) had no matrix row, so their context window fell back to the
128,000 guess and auto-compaction fired at 102,400 tokens — a tenth of their
real window — rebuilding the prompt cache each time (observed twice in one
33-minute Sonnet 5 run).
- matrix.py: rows for claude-fable-5-1, claude-opus-5 and claude-sonnet-5
(1M context, vision + PDF), per platform.claude.com/docs/en/models/overview
read 2026-09-08.
- engine.py: `_compaction_config` logs one warning per engine when a model
has no matrix window and the fallback applies, naming the model, the
assumed window and the resulting trigger.
Tests: tests/test_matrix_current_models.py — the current rows and their
windows; every provider's recommended model is a matrix row (Ollama
excepted) and the ones without a verified window are exactly the known
set (MiniMax M2.5, Muse Spark 1.1); every model in reports/reviewer-eval-*
has a window; trigger arithmetic; the fallback warning fires once for an
unlisted model, not for a listed one, and not when compaction settings
supply a window.
Devika Verma