Claude Opus 5 за полцены Fable 5 — Kimi K3 и distillation scandal

Шесть pain points: как не облажаться на этой неделе

1. «Half price» ≠ full Fable 5 replacement: Opus 5 близок на agent coding/automation, но Anthropic intentionally не пушит dual-use frontier (cyber, bio) — там Mythos 5.
2. White House statement ≠ proven fact: Kratsios/Bessent без publicly auditable evidence; Moonshot молчит про training — silence ≠ guilt.
3. Distillation timeline: Fable 5 public с 01.07, K3 launch 16.07 — 15 days для deep RL distillation unrealistic по мнению нескольких researchers.
4. Identity signal underrated: Greenblatt cross-entropy analysis ценнее политики — указывает на Claude samples с deployment metadata, но не доказывает distillation standalone.
5. Data retention blind spot: Fable/Mythos — 30-day retention opt-in; Opus 5 — no forced retention by default. Для enterprise compliance это часто tie-breaker.
6. Multi-model agent на ноутбуке: Opus 5 + K3 routing через Gateway требует 7×24 node; sleep mode = broken long-context sessions.

Claude Opus 5: release 24 июля — specs и benchmarks

24 июля 2026 (US Pacific) Anthropic выпустила Claude Opus 5 (model ID: claude-opus-5) как Claude Max default и strongest Opus tier для Pro. Positioning: daily driver с intelligence near Fable 5 при ~half token price.

Price flat, performance jump

Input $5 / 1M tokens, output $25 / 1M tokens — same as Opus 4.8. 1M context default, max output 128K; Thinking on by default, controlled via Effort param. Fast mode: ~2,5× speed, 2× price.

Anthropic highlights (vendor-reported):

• Frontier-Bench v0.1: >2× Opus 4.8, lower cost per task.
• CursorBench 3.2 max: −0,5% vs Fable 5 peak, ~50% token cost.
• ARC-AGI 3: 3× second-best model score.
• OSWorld 2.0: beats Fable 5 best at <1/3 cost.
• Zapier AutomationBench: ~1,5× pass rate; Zapier claims 100% on tasks prior models failed.

Cursor team quote: «Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost.»

Alignment и retention policy

Opus 5 — most aligned Claude generation to date; dual-use frontier остаётся у restricted Mythos 5. Compliance note: Opus 5 does not force data retention by default; Fable/Mythos require 30-day retention.

Kimi K3 distillation scandal: White House, timeline, Greenblatt evidence

Moonshot launched K3 16 июля (2,8T params, 1M context, sparse MoE). Full weights promised 27 июля — на момент скандала external verification impossible.

White House + Anthropic backstory

22–23 июля OSTP director Michael Kratsios accused Moonshot of «massive, covert industrial distillation» of Fable capabilities + alleged unlicensed Nvidia GB300 chips (possibly via Thailand routing). Treasury Secretary Scott Bessent mentioned «watermarks» in Chinese models without specifics. February 2026: Anthropic already accused Moonshot, DeepSeek, MiniMax of 3,4M+ anomalous API interactions — Moonshot neither confirmed nor denied.

Researchers: timeline doesn't add up

TechCrunch (23.07): Snorkel AI co-founder Braden Hancock — can't distill, train, and ship at this scale in two weeks after Fable 5 public release. Allen AI's Nathan Lambert: marginal gain of simple SFT distillation shrinks near frontier; RL differentiation takes longer.

Greenblatt: K3 says it's Claude

Ryan Greenblatt (Redwood Research, GitHub: rgreenblatt/which_claude_is_k3) ran cross-entropy comparison on identity responses:

• K3 disproportionately self-identifies as Claude and emits internal deployment IDs (claude-opus-4-5-20250929, claude-sonnet-4-5-20250929).
• Real Claude Sonnet 4.5 says «Claude Sonnet 4.5»; Opus 4.5 often omits version — K3 is more accurate than the teacher model.
• Version strings locked to Claude 4.5 generation (late 2025), not Fable/Mythos; K2 showed similar pattern with older Sonnet 4 signals.

Interpretation: likely Claude training data with deployment metadata — not standalone distillation proof; data contamination or prompt leak also possible. Greenblatt explicitly: findings do not directly prove distillation occurred.

Decision matrix: Claude Opus 5 vs Kimi K3

6-step runbook: model selection + agent deployment

1. Prioritize scenario: agent coding, automation, compliance → evaluate Opus 5 (see coding assistant matrix); extreme cost + open source → retest K3 post-27.07.
2. Audit retention/compliance: Opus 5 (no forced retention) vs Fable/Mythos (30 days); cross-border projects need separate legal review.
3. Calibrate Effort tiers: Opus 5 Thinking on by default — high/max for coding quality, lower tiers for automation benchmarks to avoid token waste.
4. Layer K3 evidence: politics (White House) → timeline (researchers) → tech (Greenblatt); after 27.07 reproduce identity prompts + benchmarks.
5. Deploy multi-model Gateway: OpenClaw or Kimi Code on dedicated node; isolate API keys and provider fallback per environment.
6. Post-27.07 routing update: GPQA-Diamond, BrowseComp, identity tests — update internal model routing table and archive decision record.

Три hard numbers для tech review

• Opus/Fable price-performance gap: CursorBench 3.2 max −0,5% at ~50% token cost ($5/$25 vs ~$10/$50).
• ARC-AGI 3: Opus 5 score 3× second-best; OSWorld 2.0 above Fable at <1/3 cost.
• K3 identity anomaly: Greenblatt — disproportionate Claude deployment ID output, locked to Claude 4.5 gen not Fable/Mythos.

Bottom line: price/performance war + model provenance

Opus 5 ships near-flagship intelligence at half price; K3 attacks with open source + low cost. Distillation controversy = public fight over whether cheap frontier-near performance comes from own optimization or borrowed model knowledge.

Most teams: compliance/retention → Opus 5; extreme cost/open weights → verify K3 after 27.07.

Multi-provider agent stacks on laptop fail on sleep, network, key sprawl — dedicated 7×24 Mac node is the pragmatic fix. MACCOME: real macOS, SSH handoff, isolation for Claude Code, OpenClaw, multi-model routing. Pricing: тарифы аренды Mac mini.

Sources: Anthropic (anthropic.com/news/claude-opus-5), Moonshot/Kimi, TechCrunch, Ryan Greenblatt (rgreenblatt/which_claude_is_k3), CNBC, The Verge. Data as of 25 July 2026.