📋 Оглавление

01 · Все метрики OpenRouter июнь

OpenRouter — единственный публичный источник с агрегированным production token billing миллионов разработчиков. Июньский рейтинг = фактические вызовы API, не PR-бенчмарки.

Источники: OpenRouter live traffic, Artificial Analysis Intelligence Index, SWE-bench Pro (июнь 2026). Контекст: Claude Fable 5 снят с экспорта (июнь), IPO-сигналы OpenAI/Anthropic, китайские модели >60 % всех token на платформе.

Vendor rank (weekly token)

#	Vendor	Origin	Weekly token	Share
1	DeepSeek	🇨🇳 China	5.13T	17.6%
2	Anthropic	🇺🇸 USA	4.34T	14.8%
3	Google	🇺🇸 USA	3.66T	12.5%
4	OpenAI	🇺🇸 USA	2.46T	8.4%
5	Xiaomi	🇨🇳 China	2.42T	8.3%
6	MiniMax	🇨🇳 China	2.37T	8.1%
7	Tencent	🇨🇳 China	2.36T	8.1%
8	Alibaba Qwen	🇨🇳 China	1.26T	4.3%

China aggregate (top-8): ~46 %

Model rank (daily token top-10)

#	Model	Vendor	Daily token
1	DeepSeek V4 Flash	DeepSeek	619B
2	Hy3 Preview	Tencent	451B
3	MiniMax M3	MiniMax	447B
4	MiMo-V2.5	Xiaomi	327B
5	DeepSeek V4 Pro	DeepSeek	300B
6	Claude Opus 4.7	Anthropic	263B
7	Claude Opus 4.8	Anthropic	~200B
8	Claude Sonnet 4.6	Anthropic	178B
9	Gemini 3 Flash Preview	Google	156B
10	Kimi K2.6	Moonshot AI	~150B

Hard numbers: DeepSeek V4 Flash 619B daily (#1); 5/10 top models — China; Claude tier держит high-value traffic.

02 · Три инженерные ошибки

1. Путать usage rank с quality rank. OpenRouter считает paid invocations, не SWE-bench. DeepSeek V4 Flash #1 из-за $/token, не из-за agent ceiling.

2. Игнорировать token economics. Dev из San Diego: Claude coding ~$10/hr, DeepSeek <$0.50/hr. Без routing policy — либо burn budget, либо under-provision на critical path.

3. Hardcode single provider перед Q3 release storm. GPT-6, Opus 5, Gemini 4, DeepSeek V5 — integration rewrite за 90 дней. См. тренды OpenRouter начала июня.

03 · США 70→30 %: чистая экономика

Bloomberg / Exponential View / OpenRouter:

Июнь 2025: US models ~70 % token share
Июнь 2026: ~30 %

40 pp ушли к китайским вендорам — преимущественно dev из US/EU/India. Драйвер: cheap + fast + good enough для 80–90 % daily tasks.

04 · Volume ≠ quality: два слоя

Quality ceiling: Claude Opus 4.8

Artificial Analysis Intelligence Index (конец мая 2026):

Model	Quality index	SWE-bench Pro	Note
Claude Opus 4.8	61.4 (#1)	69.2%	Long context + agent lead
GPT-5.5	59–60	63.1%	Ecosystem + tool calls
Gemini 3.1 Pro	57	—	Hard reasoning
Qwen 3.7 Max	57	—	China closed flagship
Claude Sonnet 4.6	—	80.8% (Verified)	Writing + IF

20-task shootout: Opus 4.8 wins 16, GPT-5.5 5, Gemini 3.1 Pro 4 — long-context gap максимален.

Claude Fable 5: quality score 100/100, global takedown mid-June (export control). Fable 5 ban & alternatives.

Volume kings: China via unit economics

Price: MiniMax M3 $0.60/M input vs Opus 4.8 $5.00/M — 8× delta
Good enough: daily coding/translation ~80–90 % frontier quality
Open weights: DeepSeek V4, MiniMax M3 — self-host eliminates cloud prompt leakage

Dallas stack: complex $500/mo Claude+ChatGPT; 90 % coding $200/mo MiniMax+Kimi+MiMo — complexity-based routing.

05 · Scenario matrix (июнь 2026)

Scenario	Model	Rationale
Complex code / agent	Claude Opus 4.8	Best composite, long context
Daily coding assist	DeepSeek V4 Flash / MiMo-V2.5	$/token + latency
Max API efficiency	MiniMax M3	$0.60/M, open weights
Long context	Kimi K2.6 (1M)	1M window, fair price
Google stack	Gemini 3.5 Flash	Workspace native
Realtime web	Grok 4.3	X/Twitter live
Self-hosted	GLM 5.2 / Kimi K2.6	Top open weights
Image gen	ChatGPT Images 2.0	Text rendering
Daily chat	GPT-5.5	−52.5 % hallucination vs GPT-5.3

06 · H2 forecast: Q3 release density

Model	Vendor	ETA	Key spec
GPT-6	OpenAI	Aug–Sep 2026	1.5M context (rumor), agent upgrade
Claude Opus 5	Anthropic	~Sep 2026	Opus 4.8 successor, long agent
Gemini 4	Google	Q3 2026	Multimodal video/audio
DeepSeek V5	DeepSeek	Q3 2026	Open weights, >1T params
GLM 5.2	Z.ai	Shipped	Top open weight, coding
Grok 4.3+	xAI	Q3 2026	1M context, live web

Три flagship в 6-week window mid-Aug – late Sep — benchmark leadership rotate быстрее media cycle.

07 · Пять макротрендов

1. Competition axis → scenario fit — hardest 5 % closed frontier, 95 % China open weights.

2. China share ↑, enterprise compliance cap — OpenRouter retail possibly 70 %+; Fortune 500 procurement China API possibly <30 %.

3. Agent production = battlefield — Anthropic Agent Report 2026: 44 % Claude API = math/CS tasks.

4. OpenAI + Anthropic IPO — June 2026 signals; pricing transparency vs price war.

5. Local inference breakthrough — by 2027, 32GB consumer GPU >80 % SWE-bench; см. локальный DeepSeek на Mac.

08 · Margin compression & provider-agnostic arch

Model layer commoditizing: DeepSeek 2025 доказал efficiency без peak CAPEX; Xiaomi/Tencent/MiniMax/Moonshot скопировали floor pricing. US split: OpenAI ecosystem, Anthropic quality moat, Google speed/multimodal.

High-value skill: не pick #1 model, а build switchable routing — июньский лидер через квартал может быть #4.

09 · 5 шагов: routing validation pipeline

Аренда clean macOS: Mac mini M4, SSH; local user без prod Apple ID.
OpenRouter + complexity routing: sandbox .env; peak anthropic/claude-opus-4.8, daily deepseek/deepseek-v4-flash or minimax/minimax-m3.
20-task benchmark: log USD cost, p95 latency, tool-call success — replicate Opus-16/GPT-5 methodology.
Wire Cursor or OpenClaw Gateway: model ID swap без business code diff; check 1M context gateway timeout.
Export CSV + release: revoke test keys, disk wipe, commit routing rules to team wiki.

                        # Complexity routing (OpenRouter)

                        export OPENROUTER_API_KEY="sk-or-..."

                        # Daily: DeepSeek V4 Flash (~$0.10/M in)

                        # Peak: Claude Opus 4.8 ($5.00/M in)

Multi-key + CLI + OpenClaw Gateway + Xcode signing на одном MacBook = quota burn + Keychain pollution risk. Изолированная посуточная аренда для A/B routing — тарифы bare-metal, FAQ SSH/VNC.