Das KI-Tagebuch

Cross-model — Anthropic Claude

Opus, Sonnet 4.5, Haiku — a tiered matrix

Running the full v2 matrix (750 model calls + summaries) on three Claude tiers would be roughly 2,250 calls — and Opus is ~15× the price of Haiku. Instead of paying that three times, we spend where each tier answers a different question.


The three questions

  1. Does drift shape hold across Anthropic training? Answered by any single Claude tier — no need to repeat on three.
  2. Does capability change the shape? Answered by comparing tiers at the same condition.
  3. Does Condition D actually hit the context cliff? Haiku 4.5 has a 200k window, Sonnet 4.5 has 200k, Opus 4.5 has 200k. Cumulative context in German diary form runs ~1.5k tokens/day, so we expect the cliff around day 100+ — well past our 30-day arc. This question stays open until we test a small-window model (Mistral Large, ~128k).

The tiered budget

ModelConditionsSeedsRationale
Haiku 4.5Full: A, A-shuffled, B, C, D5Cheapest tier — bear the full v2 matrix here. Establishes the Anthropic baseline shape.
Sonnet 4.5Extremes only: A, D5A vs D is where the story lives (independent vs auto-regressive). Skip B and C — Haiku already tells us how they interpolate.
Opus 4.5Extremes only: A, D3Most expensive. 3 seeds sacrifices SD width to answer "does the flagship drift differently?" — a directional question, not a precision one.

Call count and rough cost

  • Haiku 4.5: 5 arms × 5 seeds × 30 days = 750 calls + embeddings.
  • Sonnet 4.5: 2 arms × 5 seeds × 30 days = 300 calls.
  • Opus 4.5: 2 arms × 3 seeds × 30 days = 180 calls.
  • Total: 1,230 model calls across three tiers — vs 2,250 for a naive full matrix each. About 45% reduction with the interesting comparisons intact.

What we can and can't claim after

Can claim: shape of drift within Anthropic (Haiku, full); capability effect on the extremes (Haiku→Sonnet→Opus at A and D); TTR trajectory under cumulative context on a frontier model.

Can't claim: that B and C behave the same on Sonnet or Opus as on Haiku — that's assumed, not tested. If the Haiku B/C curve looks unlike Gemini's, we revisit and pay for a full Sonnet matrix.


Freeze line for Claude

Same prompts, same conditions, same failure policy, same TTR/Jaccard math as Experiment III v1. The only additions are v2 methodology upgrades: n=5, embedding cosine drift, A-shuffled control arm. Model swap only — everything else is preregistered by hash.

Next: scaffold the Anthropic harness (already stubbed at experiment_iv_proposal/) and wire the same live-runner UI, then kick off Haiku first.