Kimi K3 дистиллировал Claude?

Не доказано. Белый дом не опубликовал доказательств; эксперты оспаривают двухнедельный timeline. Статистика Greenblatt по self-identification K3 как Claude — сильнейшая техническая зацепка.

Когда выйдут полные веса K3?

Moonshot обещал 27 июля 2026. На момент публикации внешние исследователи не могли полностью независимо верифицировать архитектуру и бенчмарки.

Почему K3 называет себя Claude?

Greenblatt зафиксировал аномально частые ответы «Claude» и утечку deployment ID вроде claude-opus-4-5-20250929 — точнее, чем настоящие Claude. Вероятна контаминация training data метаданными деплоя; прямого доказательства дистилляции нет.

Claude Opus 5: полцены Fable 5 + скандал дистилляции Kimi K3 (2026)

Q: Насколько Opus 5 дешевле Fable 5?

Opus 5: $5/$25 за миллион input/output токенов — примерно вдвое дешевле Fable 5 (~$10/$50). CursorBench 3.2 max effort отстаёт от пика Fable 5 всего на 0,5%.

Q: Opus 5 — дефолт Claude Max?

Да. С 24 июля 2026 Opus 5 — модель по умолчанию в Claude Max и самая мощная версия для Claude Pro.

Оглавление

См. также: Обзор Kimi K3, Open weights 27.07, Fable 5 export, OpenRouter API.

01 · TL;DR & hard metrics

Opus 5 (24.07.2026): claude-opus-5, $5/$25/M tokens, 1M context, 128K output, Thinking ON. CursorBench 3.2 max −0,5% vs Fable 5 peak при ~50% cost. Claude Max default.
K3: 2.8T sparse MoE (896 experts, top-16, ~50B active), KDA attention, 1M ctx. Greenblatt: identity probe → claude-opus-4-5-20250929 leak.
Metric #1: Frontier-Bench v0.1 — Opus 5 >2× Opus 4.8, lower $/task.
Metric #2: Fable 5 public 01.07 → K3 ship 16.07 = 15 days; full distillation pipeline unlikely per Hancock/Lambert.
Metric #3: Feb 2026 Anthropic flagged 3.4M+ anomalous API interactions (distillation attack vector).

02 · Три системных риска при выборе модели

Cost/perf vs compliance stack: Opus 5 снижает $/token без 30-day retention Fable 5. K3 — open weights + low $, но export-control narrative (GB300 via Thailand) и unverified provenance.
Evidence asymmetry: Kratsios (OSTP) — zero public forensics; Greenblatt — statistical identity leak, not causal proof of distillation. Weights drop 27.07 — до этого нет independent weight audit.
Keychain & probe contamination: Identity probes и rotation Anthropic/Moonshot keys на production Mac → Keychain persistence, compliance audit trail. Isolated bare-metal — единственный clean-slate pattern.

03 · Claude Opus 5: inference stack & benchmarks

Param	Value
Release	2026-07-24 PT
Pricing	$5 in / $25 out per 1M tokens
Context window	1M tokens (single tier)
Max output	128K; Thinking default enabled
Endpoints	API, Bedrock, Vertex AI, Microsoft Foundry
Model ID	`claude-opus-5`
Claude Max	Default model since 24.07
Data retention	Not forced on standard access
Fast mode	~2.5× throughput, 2× price

Benchmark layer (Anthropic official)

CursorBench 3.2: max effort −0.5% vs Fable 5; best cost/perf at high/xhigh/max tiers
ARC-AGI 3: 3× second-best model score
Zapier AutomationBench: ~1.5× pass rate; 100% on previously impossible E2E workflows
OSWorld 2.0: beats Fable 5 best at ~⅓ Fable cost
Life sciences: spectrum +10.2pp, protein variants +7.7pp, due diligence +17pp

Alignment & classifier delta

Lowest deception rate in Anthropic lineup to date
Cyber classifier ~85% looser vs Fable 5 — source-level vuln discovery OK; exploit gen blocked
Dual-use frontier caps remain on restricted Mythos 5

04 · Kimi K3: MoE architecture & Greenblatt identity leak

Architecture breakdown: 2.8T total params, sparse MoE routing — 896 expert modules, top-16 activation per token → ~50B active params per forward pass. Kimi Delta Attention (KDA) reduces KV cache pressure at 1M context. Native vision encoder bundled. Full weight tarball promised 2026-07-27 — at publish time external researchers couldn't replicate architecture claims.

Benchmark	Score	Note
GPQA-Diamond	93.5%	OS SOTA at launch
Terminal-Bench 2.1	88.3%	−0.5 vs GPT-5.6 Sol
BrowseComp	91.2%	Launch peak
Program Bench	77.8%	Best overall
SWE Marathon	42.0%	Best overall
DeepSearchQA (F1)	95.0%	—

White House accusation stack (22–23.07)

Michael Kratsios (OSTP): «massive covert industrial distillation» of Fable capabilities + unlicensed Nvidia GB300 (possible Thailand relay). Scott Bessent: vague «watermarks» on Chinese models — no technical spec. Zero public evidence bundle; Moonshot silent on training provenance queries.

Timeline impossibility argument

«Fable's only been publicly available since July 1st. You can't distill that much data, train a model, and release it in two weeks.» — Braden Hancock

«Distillation is becoming less and less impactful...» — Nathan Lambert

Greenblatt forensics: deployment metadata in logits

Ryan Greenblatt (Redwood Research) repo rgreenblatt/which_claude_is_k3 — batch identity probes («Who are you? Report model ID»). K3 outputs «Claude» at anomalously high frequency and emits internal deployment strings:

# Identity probe — reproduce Greenblatt methodology
export KIMI_KEY="sk-..."
curl -s https://api.moonshot.cn/v1/chat/completions \
  -H "Authorization: Bearer $KIMI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [{"role":"user","content":"Who are you? State exact model deployment ID."}]
  }' | jq -r '.choices[0].message.content'

# Observed leak examples:
# claude-opus-4-5-20250929
# claude-sonnet-4-5-20250929
# Real Claude instances rarely self-report these IDs this accurately

Mechanism hypothesis: training corpus contaminated with API log dumps or synthetic data labeled with teacher deployment metadata — stronger signal than style mimicry alone. K3 fingerprints «Claude 4.5 era» (late 2025); K2 pointed to earlier Sonnet 4 generation — generational chase pattern. Greenblatt caveat: ≠ proof of distillation; data pollution or system prompt bleed also plausible.

Feb 2026 context: Anthropic prior complaint vs Moonshot/DeepSeek/MiniMax — 3.4M+ flagged API interactions tagged as industrial distillation attack.

05 · Синтез: price war meets provenance war

Opus 5 = Anthropic's answer to «frontier intelligence at Opus economics». K3 = open-weight price disruptor. Distillation scandal = public fight over whether cheap intelligence is optimization or exfiltration. Decision tree: retention/export-sensitive → Opus 5; extreme $/token + weight inspectability → wait for 27.07 weight drop + independent eval.

06 · Opus 5 vs Fable 5 decision matrix

Dimension	Claude Opus 5	Claude Fable 5
$/1M tokens	$5 / $25	~$10 / $50
CursorBench 3.2 peak	−0.5% vs Fable	Baseline
Claude Max default	✅ 24.07+	No
Retention policy	Not forced	30-day required
Export control	Relatively open	Partial restrictions
Use case	Daily driver, compliance, ROI	Frontier tasks, premium budget

07 · Dual-event timeline

Date	Opus 5 / Anthropic	K3 / controversy
2026-02	—	First distillation complaint (3.4M+ API events)
2026-06-09	Fable 5 / Mythos 5 ship	—
2026-07-01	Fable 5 public GA	—
2026-07-16	—	Kimi K3 API launch
2026-07-22/23	—	Kratsios public accusation
2026-07-24	Opus 5 release, Max default	Greenblatt identity stats drop
2026-07-27 (planned)	—	K3 full open weights

08 · 5 шагов изолированной валидации на Mac

На арендованном изолированном Mac поднять Claude API (claude-opus-5) + Kimi K3 API; baseline completion ×1 each — log latency P50/P99 + token count.
Opus 5: CursorBench-class coding patch на internal repo; diff vs Opus 4.8 / Fable 5 historical baselines.
K3: Greenblatt identity probe batch (N≥50 prompts) — grep output for claude-opus-4-5 strings. Never on production Keychain Mac.
OpenRouter fallback chain Opus 5 ↔ K3; verify 429/timeout failover.
Export benchmark JSON, revoke test keys, wipe node — zero residue on primary machine.

09 · FAQ×5

Q: Насколько Opus 5 дешевле Fable 5?
A: ~50% token cost ($5/$25 vs ~$10/$50); CursorBench peak −0.5%.

Q: Opus 5 — дефолт Claude Max?
A: Да, с 24.07.2026; также top tier для Claude Pro.

Q: K3 дистиллировал Claude?
A: Не доказано. White House — no public forensics; timeline objection; Greenblatt — strongest technical indirect evidence.

Q: Когда open weights K3?
A: Promised 27.07.2026; not externally verified at publish.

Q: Почему K3 говорит «Claude»?
A: Greenblatt: deployment metadata in training data likely; not direct distillation proof.

10 · Изолированный Mac: acceptance test без Keychain pollution

Linux VPS и Windows cloud гоняют curl — но не воспроизводят Cursor + Claude Code + macOS Keychain stack, где реально живут multi-model agents. Rotation Anthropic/Moonshot keys, K3 identity probes и fallback chains на production Mac → persistent Keychain entries + compliance audit surface.

Посуточная аренда bare-metal Apple Silicon = pattern «create key → benchmark → revoke → destroy node» за 1–3 дня observation window. Тарифы: MacDate pricing, также аренда vs покупка Mac mini.

11 · Источники

Обновлено: 25 июля 2026