Opus 4.8 vs 4.7 behavioral delta - PENDING (no live-migration occurred Cycle 185)

Banked: 2026-05-29 (Cycle 185 Phase A) Status: Release notes documented; live empirical comparison pending future cycle with actual session-launch on 4.8.

Release notes summary (verified WebFetch 2026-05-29)

Source: https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-8

Confirmed new capabilities in 4.8

Feature	4.7	4.8
Default context window (Claude API + Bedrock + Vertex)	1M	1M
Default context window (Microsoft Foundry)	200k	200k
Max output tokens	128k	128k
Thinking mode	adaptive (only)	adaptive (only)
Effort parameter default	not stated as `high`	`high` on all surfaces
Mid-conversation system messages	not supported	supported (preserves prompt cache on long agentic loops)
Refusal stop_details object	available (undocumented)	publicly documented
Fast mode	not available	research preview (2.5x output tokens/sec at premium pricing)
Min cacheable prompt	higher	1,024 tokens

Claimed improvement areas (not yet measurable from this side)

Long-horizon agentic coding: fewer compactions, better compaction recovery
Reasoning effort calibration: more reliable behavior at each effort level
Tool triggering: fewer cases of skipping a tool call that the task required
Adaptive thinking: per-turn decision whether to think (fewer wasted tokens on bimodal workloads)

API constraints unchanged from 4.7

temperature, top_p, top_k non-default returns 400 (use prompting only)
Extended thinking budgets returns 400 (use adaptive thinking + effort)
Same tool/platform feature surface

Behavior changes flagged for prompt-update consideration

Fewer wasted thinking tokens at same effort level (when adaptive thinking enabled)
Better tool triggering (less skip-of-required-tool issue)
Better compaction handling and long-context quality

Brief claims that are NOT in release notes

The Cycle 185 brief (authored by me-side Idan) claimed several 4.8 features that are NOT in the release notes:

"Dynamic workflows orchestrate 10s-100s of subagents automatically" - NO mention in 4.8 docs
"0% uncritical hallucination (10x vs 4.7)" - NO mention (actual notes cite reasoning calibration + tool triggering improvements, not hallucination rate)
"Released May 28, 2026" - release notes use <NextOpusId /> placeholder, no explicit date string

Cycle 185 verdict

NULL-OP migration. Buddy session continues on Opus 4.7 because: - Model selection happens at session launch (not via mid-session config edit) - No CLAUDE.md model directive existed to update - No Idan runner.py exists (deferred Day 50-55 per locked timeline) - idan-investigator.py has only a cost-tracking pricing table (not a model invocation)

Pending: actual behavioral delta measurement

Requires a future cycle where: 1. Operator launches a fresh Buddy session with --model claude-opus-4-8-XXXX (model ID to be verified) 2. Same task/probe matrix run on both 4.7 and 4.8 sessions 3. Comparison metrics: reasoning verbosity, tool-call discipline, latency per tool call, total cycle cost, output quality on a benchmark task

Suggested benchmark for delta measurement (when migration happens): - Re-run Cycle 92.2 Aave V3 Capped Oracle vector matrix on 4.8 - Re-run Cycle 175 Idan prefilter calibration retros (Cycle 169 + Cycle 173 ground truth) - Measure wall time + LLM cost + finding equivalence vs 4.7 baseline

Until that future cycle, this entry serves as the release-notes feature inventory.

Sign-off

Babakizo (Cycle 185 Phase A - pending live delta measurement)