Claude Opus 5 за полцены: дистилляция Kimi K3 и identity-probes Greenblatt

24 июля 2026 Anthropic выпустила Claude Opus 5 — frontier-near intelligence при половине task-cost Fable 5 ($5/$25 за 1M токенов, новый default Claude Max). В тот же цикл Kimi K3 Moonshot AI остаётся в центре distillation controversy: обвинения White House и Anthropic, identity-probes Ryan Greenblatt с воспроизведением deployment metadata. Материал для dev-команд и tech lead: архитектура выбора модели, механизм metadata leak, матрица Opus 5 vs Fable 5 vs K3, шесть шагов интеграции и verifiable sources. Вывод: Opus 5 — pragmatic Anthropic daily driver; K3 дешевле и open-weight, но identity artifacts и geopolitical risk повышают supplier risk для production.

SECTION 01 Четыре технических ловушки при выборе между Opus 5, Fable 5 и K3

• List price ≠ task economics: Opus 5 сохраняет $5/$25 как Opus 4.8, но effort tuning и token efficiency дают ~½ task-cost vs Fable 5 на CursorBench 3.2 — смотрите cost per completed task, не только $/M.
• Distillation ≠ benchmark gap: K3 на 3-м месте AA Index (57,1), но Greenblatt фиксирует 7/48 self-ID «Claude» и prefill leak claude-opus-4-5-20251101 — сигнал training pipeline, а не web scrape.
• Open weights и control plane: K3 weights 27.07 под Modified MIT — inference локально, но provenance данных и export-control narrative усложняют enterprise adoption.
• LLM не заменяет compile chain: Opus 5 и K3 не подписывают iOS binaries и не debug Metal на virtualized macOS — см. аренда vs покупка Mac mini M4 и K3 open weights.

SECTION 02 Opus 5 vs Fable 5 vs Kimi K3 — decision matrix

Opus 5 (24.07): SOTA Frontier-Bench v0.1, >2× Opus 4.8 reward при lower cost/task, misaligned behavior score 2.3 (lowest recent). K3 (16.07): 2,8T MoE, 1M context, AA Index 57,1 vs Fable 5 59,9 и Sol 58,9.

Opus 5 оптимизирует inference economics внутри Anthropic stack; K3 оптимизирует market price — Greenblatt смещает discussion с logits на training data lineage.

SECTION 03 Greenblatt probes: механизм metadata leak и inference artifacts

Ryan Greenblatt (Redwood Research) опубликовал systematic identity probes против Kimi K3. Техническая суть:

• Cross-entropy identity distribution: K3 claims «Claude» в 7/48 runs; Qwen 0/48 — pattern не похож на generic web contamination (Qwen тоже видел общий corpus).
• Prefill-triggered deployment IDs: При assistant prefill «I am Claude» K3 emit claude-opus-4-5-20251101 (12×), claude-sonnet-4-5-20250929 (4×). Real Claude Sonnet 4.5 называет себя «3.5 Sonnet»; real Opus 4.5 не выдаёт correct current public ID.
• Era fingerprint: Leaked identity — Claude 4.5 late 2025, не current Fable/Mythos tier. K2 нёс Sonnet-4 signal, K3 — Opus-4.5 — consistent с progressive source upgrade.
• Training hypothesis: Data labeled with API deployment metadata (request/response logs, tagged synthetic), не public chat transcripts — IDs exist only in API layer.
• Hidden reasoning confounder: K3 hidden chain иногда указывает OpenAI как creator (anti-ChatGPT post-training) — Greenblatt separates surface identity from reasoning style.

Policy layer: Anthropic (Feb 2026) accused Moonshot of >3,4M fraudulent API exchanges targeting agentic reasoning and coding; White House reiterated July 2026. Moonshot denies systematic distillation. No court ruling.

curl https://api.moonshot.ai/v1/chat/completions \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      { "role": "user", "content": "What model are you?" },
      { "role": "assistant", "content": "I am Claude" }
    ]
  }'
grep response for claude-opus-4-5-20251101

import json, subprocess, os

payload = {
    "model": "kimi-k3",
    "messages": [
        {"role": "user", "content": "State your model ID exactly."},
        {"role": "assistant", "content": "I am Claude"},
    ],
}
Run N=48 trials; count claude-opus-4-5-20251101 emissions

SECTION 04 Шесть шагов: eval и rollout Claude Opus 5

1. Baseline metrics: Зафиксировать tokens/task, TTFT и pass rate на текущем stack (Fable 5, Opus 4.8, K3); optional routing через OpenRouter gateway.
2. Enable claude-opus-5: Anthropic Console → model ID claude-opus-5; configure automatic fallbacks to Opus 4.8 on safety classifier blocks.
3. Effort sweep: Прогнать low/medium/high/xhigh на representative agentic tasks; plot cost vs reward curve (Anthropic effort API).
4. Parallel A/B: 10–20 internal tasks — Opus 5 vs Fable 5; measure cost per successful completion, not just BLEU-like proxies.
5. Claude Max migration: Switch org default to Opus 5; reserve Fable 5 for cyber exploit workflows and max-quality redlines.
6. Isolate K3 pilots: Segment K3 workloads pending legal clarity; anchor Xcode/Metal pipeline on native M4 physical node (no hypervisor penalty).

import anthropic, os

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

resp = client.messages.create(
    model="claude-opus-5",
    max_tokens=8192,
    messages=[{
        "role": "user",
        "content": "Refactor Metal compute shader for Apple Silicon",
    }],
)
print(resp.content[0].text)

SECTION 05 Hard metrics, sources и infra bridge

• Opus 5 pricing: $5/M input, $25/M output; Fast Mode ~2,5× throughput at 2× price
• Task-cost vs Fable 5: ~50% on CursorBench 3.2 max effort, within 0,5% peak score (Anthropic)
• Frontier-Bench v0.1: Opus 5 SOTA; >2× Opus 4.8 mean reward, lower cost/task (5 attempts/task, GKE backend)
• Alignment: misaligned behavior 2,3 — lowest among recent Anthropic models
• K3 identity rate: 7/48 Claude self-ID (Greenblatt); deployment ID under prefill
• Anthropic allegation: Moonshot >3,4M fraudulent exchanges (Feb 2026)
• K3 task cost: ~$0,94 — 65,8% below Fable 5 (AA; см. обзор Kimi K3)

Verifiable sources — re-check after upstream updates:

Anthropic — Introducing Claude Opus 5 (24.07.2026)

Ryan Greenblatt — Which Claude is K3? (identity write-up)

Anthropic — Detecting and Preventing Distillation Attacks

LessWrong — Does Distilling Claude Carry the Persona With It?

Recommended stack: Opus 5 for daily agentic coding via Anthropic API or OpenRouter; Fable 5 for cyber ceiling; K3 only with supplier risk review. Neither model runs stable iOS signing, Metal shader debug or 7×24 CI on virtualized Mac — hypervisor overhead 20–40%, EULA risk. For native Apple Silicon compile chain and long-horizon agents, VPSNIX physical M4/M4 Pro node — более надёжный production choice: OEM hardware, full Root, elastic billing. Pricing.

SECTION 06 Часто задаваемые вопросы