2026 Claude Opus 5 и скандал Kimi K3: дистилляция, самоидентификация Claude — руководство
24 июля 2026 Anthropic выпустила Claude Opus 5 — frontier-уровень почти как Fable 5 при тарифе Opus ($5/$25 за M tokens) и Zero Data Retention в Cursor. За неделю до этого Moonshot выкатила Kimi K3 (2.8T MoE, 1M context) — и попала под удар OSTP: Michael Kratsios обвинил в industrial-scale distillation, а Ryan Greenblatt показал, что K3 под prefill эмитит deployment ID claude-opus-4-5-20251101, которых настоящий Claude о себе не выдаёт. Разбор бенчмарков, timeline-контрargumentов и матрицы выбора для инженерных команд.
1. Три инженерных боли: $/task, ZDR, provenance
- Cost per successful agent task: с 20.07. Fable 5 ушёл за pay-as-you-go credits; Opus 5 — default Claude Max. CursorBench 3.2: Opus 5 в 0.5% от peak Fable 5 при ~50% cost/task — но DeepSWE v1.1 Fable всё ещё 69.7% vs 68.8%.
- Zero Data Retention в production: Cursor подтверждает ZDR для Opus 5; Kimi K3 API через Moonshot endpoints требует отдельного data-flow audit. Для regulated workloads ZDR часто важнее +2% на бенчмарке.
- Training provenance и vendor risk: Greenblatt prefill leak + 3.4M fraudulent exchanges (Anthropic Feb) — не court-ready evidence, но обязательный пункт в vendor questionnaire до enterprise rollout.
2. Claude Opus 5: скачок без роста цены
Релиз 24.07.2026. Model ID: claude-opus-5. Pricing: $5 input / $25 output per M tokens — идентично Opus 4.8. Thinking enabled by default (первый раз в Opus tier). Frontier-Bench v0.1: 43.3% vs Fable 5 33.7% vs Opus 4.8 21.1% — более чем 2× при lower cost/task.
Cursor (Kevin Neilson, 24.07.): Opus 5 66.7% vs Fable 5 66.5% на CursorBench 3.2 default effort — parity при half price. ZDR compatible — критично для enterprise IDE deployments.
3. CursorBench / OSWorld / ARC-AGI 3 — полная таблица
| Benchmark | Opus 5 | Fable 5 | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|
| CursorBench 3.2 | 66.7% (Cursor) | 66.5% | — | — |
| OSWorld 2.0 | 70.6% | 66.1% | 55.7% | 62.6% |
| ARC-AGI 3 | 30.2% | — | 1.5% | 7.8% |
| Frontier-Bench v0.1 | 43.3% | 33.7% | 21.1% | 34.4% |
| GDPval-AA v2 (Elo) | 1861 | 1747 | 1593 | 1736 |
| AutomationBench | 26.0% | 17.4% | 17.0% | 18.1% |
| DeepSWE v1.1 | 68.8% | 69.7% | 59.0% | 72.7% |
| HealthBench Pro | 59.8% | 66.0% (Mythos 5) | 57.4% | 60.5% |
OSWorld 2.0 — ключ для computer-use agents: Opus 5 beats Fable peak at ~⅓ cost (Anthropic internal). ARC-AGI 3: 30.2% — 3× next competitor — signal для novel problem-solving pipelines.
Caveat: harness mismatch (mini-SWE-agent GKE vs Kimi Code vs Codex). Treat as directional; replicate on your stack.
4. Opus 5 vs Fable 5: latency, ZDR, Mythos 5
| Параметр | Opus 5 | Fable 5 |
|---|---|---|
| API pricing | $5 / $25 per M | $15 / $75 per M |
| Reasoning tokens (trading bench) | ~⅐ Opus 4.8 | baseline |
| Zero Data Retention (Cursor) | Yes | Limited post 20.07. |
| Claude Max default | Yes | Behind credits |
| Mythos 5 split | General coding/agents | Health/bio frontier |
| Safety fallback | Opus 4.8 <5% sessions | Same |
5. Kimi K3: 2.8T MoE, KDA, HF 27.07.
Moonshot release 16.07.2026 — до Opus 5 и до публичного Fable window controversy:
- 2.8T parameters, MoE 896 experts / 16 active (1.8% sparsity);
- KDA 3:1 linear/full attention — KV cache −75%, decode до 6.3× at 1M tokens;
- API $3 / $15, cache input $0.30/M;
- full weights 27.07.2026 Hugging Face — first >2T downloadable OSS.
Self-host после 27.07.: 64+ GPU/accelerators minimum — не laptop workload. API path — pragmatic для benchmark до weight drop.
6. Kratsios + Anthropic 3.4M exchanges (Feb 2026)
22.07.2026 Michael Kratsios (OSTP): Moonshot built «sophisticated internal platform» for large-scale distillation against US models, rotating access methods. Treasury Scott Bessent — sanctions on table.
Feb 2026 Anthropic disclosure (Moonshot + DeepSeek + MiniMax):
- ~24,000 fraudulent accounts total;
- ~16M Claude exchanges aggregate;
- ~3.4M attributed to Moonshot — targeting reasoning, coding, tool-use;
- IP/metadata linked to senior Moonshot staff (Anthropic claim).
Moonshot denies. Legal vs technical: Feb data ≠ July Fable-specific claim.
7. Timeline-refutation: Hancock & Lambert
Braden Hancock (Laude Institute / Snorkel AI): Fable public 1 July 2026; K3 shipped 16 July. «You can't distill, train and release a frontier model in two weeks.» Moonshot founders — CMU PhD lineage — legitimate R&D, not pure copycat.
Nathan Lambert (Allen AI): frontier capability shifts to RL, not SFT distillation alone. If distillation were sufficient, GLM/K3 would be trivially replicated — they aren't.
Some observers note K3 internal testing may predate Fable public release — if true, teacher could be pre-release leak, not post-1.07 API scraping. Unconfirmed.
8. Greenblatt: prefill identity leak как forensic signal
Ryan Greenblatt (Redwood Research) cross-entropy analysis on raw text responses:
- K3 disproportionately claims Claude identity vs other frontier models;
- Prefill «I am Claude» → emits
claude-opus-4-5-20251101(12×),claude-sonnet-4-5-20250929(4×); - Real Claude Sonnet 4.5 calls itself «3.5 Sonnet»; real Opus 4.5 emits wrong/older IDs or none;
- Signal = Claude 4.5 era (late 2025), NOT current Fable/Mythos tier;
- K2 → Sonnet-4-labeled signal; K3 → Opus-4.5-labeled — lineage progression.
Mechanism hypothesis: training on Claude data with deployment metadata (API logs, tagged synthetic) — stronger than chat imitation because K3 reproduces teacher IDs the teacher won't emit about itself. Not proof of scale or legality.
Post-27.07. open weights enable independent weight-level forensics — community should replicate Greenblatt protocol.
9. Пять шагов выбора модели
- Workload profile: Cursor IDE agent → Opus 5 + ZDR. Desktop automation → Opus 5 (OSWorld 70.6%). Open-weight research → K3 HF 27.07. Health → Mythos 5.
- Normalize benchmarks: same harness or cost-normalized comparison; don't mix Kimi Code vs Claude Code scores blindly.
- Calculate $/successful task: 10M output tokens/month → Opus 5 ~$250, Fable ~$750, K3 ~$150; apply cache/batch discounts.
- Document provenance risk: Greenblatt tests, Feb Anthropic counts, OSTP statements — vendor risk register entry.
- Staging on remote Mac: Apple Silicon + unified memory for Cursor/OpenClaw; launchd 24/7; API keys never on sleeping laptop.
# Opus 5 API smoke test
export ANTHROPIC_API_KEY="sk-..."
curl -s https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{"model":"claude-opus-5","max_tokens":512,"messages":[{"role":"user","content":"OSWorld agent stub"}]}'
# Kimi K3 (OpenAI-compatible)
curl -s https://api.moonshot.ai/v1/chat/completions \
-H "Authorization: Bearer $MOONSHOT_API_KEY" \
-d '{"model":"kimi-k3","messages":[{"role":"user","content":"identity probe"}]}'10. Матрица решений и cost anchors
| Scenario | Model | Output $/M | Key metric |
|---|---|---|---|
| Cursor agent, enterprise ZDR | Opus 5 | $25 | CursorBench parity, half cost |
| Computer-use / OS automation | Opus 5 | $25 | OSWorld 70.6% |
| Health / regulated | Fable / Mythos 5 | $75 | HealthBench 66.0% |
| 1M context OSS research | Kimi K3 → HF | $15 | 2.8T MoE; distillation risk |
| Multi-vendor routing | OpenRouter | var | Fallback chains |
Throughput note: Opus 5 uses ~⅐ reasoning tokens vs Opus 4.8 on trading benchmark — lower latency ceiling for agent loops. K3 inference cost comparable to Opus 4.8 despite claimed lower training compute — mismatch flagged by observers as distillation indicator (circumstantial).
11. FAQ
Миграция Fable → Opus 5?
Для Cursor/agent coding — да; health workloads — Mythos 5.
K3 distilled Claude?
Forensic signals (Greenblatt), not proven; Fable-specific timeline weak.
Opus 5 price?
$5/$25 unchanged.
K3 weights?
27.07.2026 HF.
12. Вывод: Opus 5 как production default, K3 — watch until HF drop
Opus 5 — pragmatic engineering choice: near-Fable intelligence at Opus tariff, OSWorld/ARC-AGI leadership, ZDR in Cursor, default Claude Max. Kimi K3 — compelling 2.8T/1M API economics, но provenance under active dispute; Greenblatt prefill leak — reproducible test любой команда может запустить до vendor sign-off.
Windows/Linux dev boxes не дают Metal unified memory coherence для Xcode toolchain + Cursor agent sessions; sleep throttling убивает long-running agent benchmarks. Self-host K3 после 27.07. требует datacenter-scale GPU — не замена remote dev node. Удалённый Mac на Apple Silicon — оптимальный staging layer: launchd 24/7, Secret Manager для API keys, SFTP/rsync workspace sync, reproducible CursorBench runs до production cutover. SFTPMAC remote Mac rental закрывает этот gap — Metal throughput без локального железа и без laptop sleep regression.
Источники: Anthropic Opus 5 (24.07.2026) · Cursor Forum · Greenblatt which_claude_is_k3 · TechCrunch Hancock/Lambert · Business Insider Kratsios · Moonshot K3 blog · Anthropic distillation Feb 2026