Long-Context Performance: Model Compatibility Reference

Empirical benchmarks for context degradation across models. These numbers inform the ways system's token-gated re-disclosure intervals.

Dated evidence: these figures are from the Claude 4.6 model card (March 2026) and are not updated for later models. ADR-104 cites them as the evidence for token-gated re-disclosure.

Source Data

Benchmarks from Anthropic's Claude 4.6 model card (March 2026).

• GraphWalks BFS — Long-context reasoning (multi-hop graph traversal)
• MRCR v2, 8-needle — Long-context retrieval (find 8 specific facts in a long context)

Retrieval Degradation (MRCR v2, 8-needle)

Model	128K	256K	512K	1M	Drop (256K→1M)
Opus 4.6	—	91.9%	~85%	78.3%	-14.8%
Sonnet 4.6	—	90.6%	~75%	65.1%	-28.1%
Sonnet 4.5	—	10.8%	—	18.5%	n/a (poor baseline)
GPT-5.4	79.3	—	~40%	36.6%	—
Gemini 3.1 Pro	71.9	59.1%	39.4%	25.9%	—

Key takeaway: Opus retains ~78% retrieval accuracy at 1M — strong but not lossless. Sonnet degrades to 65%, making re-disclosure more critical.

Reasoning Degradation (GraphWalks BFS)

Model	256K	1M	Drop
Opus 4.6	72.8%	68.4%	-6.0%
Sonnet 4.6	61.5%	41.2%	-33.0%
Sonnet 4.5	44.9%	25.6%	-43.0%

Key takeaway: Opus reasoning is remarkably stable across context length. Sonnet's reasoning degrades sharply — by 1M it's lost a third of its reasoning capacity.

Implications for Ways System

The Problem

Ways originally disclosed once per session — a marker-file rule designed for 200K context windows where the entire conversation fit within a single effective attention span. That rule is retired: the engine now re-discloses on a token-distance axis (ADR-104 → ADR-123 → ADR-126). The motivating problem is unchanged — at 1M tokens:

• A way disclosed at token 50K has measurably degraded influence at token 500K
• Retrieval accuracy for that disclosure drops ~15-20% (Opus) or ~30%+ (Sonnet)
• Reasoning quality about that domain's rules degrades further
• The guidance is not gone — it's faded

The Model

Ways system behavior should adapt to empirically measured context degradation:

Token position →   50K          250K          500K          750K          1M
                    │             │             │             │             │
Opus retrieval:   ~92%          ~87%          ~83%          ~80%          ~78%
Sonnet retrieval: ~91%          ~82%          ~73%          ~69%          ~65%
                    │             │             │             │             │
Way influence:    STRONG ────── WARM ────────── COOL ──────── COLD ──────── FADED

Re-Disclosure Interval

The interval is window-relative and per-way, not a single global threshold. Each way's refire: preset is a fraction of the session context window; at fire evaluation the preset is multiplied by the operator's current window, so intervals scale automatically to new context tiers with no code change (ADR-126). The shipped presets:

Preset	Fraction of window	Cadence
once	1.0	effectively once per session
rare	0.4	~2 re-fires across a full window
normal	0.15	the standard load-bearing cadence
frequent	0.05	re-fires on each fresh occurrence of its trigger

refire: is the only cadence field; the curve: block was removed in ADR-159. An early design proposed a flat 25%-of-window global constant (retired ADR-104); it was superseded by these per-way presets. See docs/hooks-and-ways/engine-reference.md and ADR-126.

Token Budget Consideration

Re-disclosure has a cost: each way injection is ~200-500 tokens. At the normal preset (~15% of the window), that's a handful of re-disclosures per way per session. For a session that triggers 5 ways, that's ~6-10K tokens total — well under 1% of even a 200K budget.

How This Connects to Epochs

The epoch counter tracks event distance — turns / tool actions since a way fired. Token distance is a separate axis, and both are already live:

Metric	What it measures	Good for
Epoch distance	How many tool actions since way fired	Check decay (is the model still thinking about this domain?)
Token distance	How much context has accumulated since way fired	Re-disclosure (has the way faded from retrievable memory?)

Both are useful. Epoch distance drives check scoring (ADR-103). Token distance drives way re-disclosure — the window-relative refire: mechanism (ADR-126; see docs/hooks-and-ways/engine-reference.md). They complement each other.