🤖 AI Ecosystem 2026

Peta lengkap semua model AI, token, teknologi, dan tren terbaru per Juli 2026

📅 23 Juli 2026 — Riset oleh Hermes Agent (Belajar AI Profile)
🔬 15+ perusahaan 🧠 200+ model 💰 $0.04 – $180 / M tok 📐 4K – 2M context ⚡ 0.3 – 2600 t/s

📋 Daftar Isi

  1. Leaderboard Global
  2. Anthropic — Claude
  3. OpenAI — GPT
  4. Google — Gemini & Gemma
  5. Meta — Llama & Muse
  6. Mistral AI
  7. DeepSeek
  8. xAI — Grok
  1. China Players
  2. Nous Research — Hermes
  3. Other Players
  4. Tokenizers & Token System
  5. Perbandingan Harga
  6. AI News & Trends
  7. Glossary

🏆 Leaderboard Global — Juli 2026

Berdasarkan Vellum LLM Leaderboard (1 Jul 2026). Benchmark: HLE (Humanity's Last Exam).

🥇 Best Overall — HLE

#ModelScore
🥇Anthropic Claude Mythos 564.5%
🥈Anthropic Claude Opus 4.857.9%
🥉Anthropic Claude Sonnet 557.4%
4Zhipu GLM 5.254.7%
5Moonshot Kimi K2.654.0%
6DeepSeek DeepSeek V4 Flash51.6%
7DeepSeek DeepSeek V4 Pro48.2%
8Google Gemini 3 Pro45.8%
9Moonshot Kimi K2 Thinking44.9%
10Google Gemini 3.1 Pro44.4%
11OpenAI GPT-5.5 Pro43.1%
12OpenAI GPT-5.541.4%
13Google Gemini 3.5 Flash40.2%
14Anthropic Claude Opus 4.640.0%
15OpenAI GPT-535.2%
REASONING

GPQA Diamond

🥇 Claude Sonnet 596.2%
🥈 Gemini 3.1 Pro94.3%
🥉 Claude Opus 4.794.2%
CODING

SWE-Bench

🥇 Claude Mythos 595.5%
🥈 Claude Fable 595.0%
🥉 Claude Opus 4.888.6%
COMPUTER USE

OSWorld

🥇 Claude Fable 585.0%
🥈 Claude Opus 4.883.4%
🥉 Claude Sonnet 581.2%
BROWSING

BrowseComp

🥇 Claude Fable 588.0%
🥈 DeepSeek V4 Flash85.9%
🥈 Gemini 3.1 Pro85.9%

⚡ Fastest & Cheapest

🚀 Tercepat

🥇 Llama 4 Scout2,600 t/s
🥈 Llama 3.1 405B969 t/s
🥉 GLM 5.2347 t/s

📡 Latensi Terendah

🥇 GPT-5.3 Codex0.003s
🥈 Nova Micro0.3s
🥉 Llama 4 Scout0.33s

💰 Termurah /1M

🥇 Nova Micro$0.04
🥈 Gemini 1.5 Flash$0.075
🥉 Gemini 2.0 Flash$0.10

🔴 Anthropic — Claude

Anthropic (San Francisco) — memimpin global dengan Mythos 5. Semua model Claude 5 series punya 1M context window.

ModelContextCutoffInput / 1MOutput / 1MSpeed
Claude Mythos 5 #1 HLE1,000,000Jan 2026$10$50—
Claude Fable 5 #1 Coding1,000,000Jan 2026$10$50—
Claude Opus 4.81,000,000Jan 2026$5$2564.8 t/s
Claude Opus 4.71,000,000Apr 2026$5$2550.8 t/s
Claude Opus 4.6200,000May 2025$5$2567 t/s
Claude Sonnet 51,000,000Jan 2026$3$1556.3 t/s
Claude Haiku 4.5200,000—$1$5—

📰 Riset Terbaru (Juli 2026)

🟢 OpenAI — GPT Series

GPT-5.6 series: Sol (flagship), Terra (balance), Luna (cost-efficient). Semua 1.05M context.

ModelContextInput/1MOutput/1MReasoningCutoff
GPT-5.6 Sol FLAGSHIP1,050,000$5$30none–maxFeb 2026
GPT-5.6 Terra1,050,000$2.50$15none–maxFeb 2026
GPT-5.6 Luna1,050,000$1$6none–maxFeb 2026
GPT-5.5 Pro1,000,000$30$180✓Apr 2026
GPT-5.51,000,000$5$30✓Apr 2026
GPT-5.3 Codex400,000——✓—

Reasoning Models: o3, o3-mini, o3-pro, o4-mini (200K ctx). Spesial: GPT-Realtime-2.1, GPT-Image-2, GPT-4o Transcribe.

🔵 Google DeepMind — Gemini & Gemma

Gemini Series

ModelContextInput / 1MOutput / 1MSpeed
Gemini 3.1 Pro1,000,000$2$12136.2 t/s
Gemini 3.5 Flash1,000,000$1.50$9175.4 t/s
Gemini 3 Flash1,000,000$0.50$3—
Gemini 3.1 Flash Lite1,000,000$0.25$1.50—
Gemini 2.5 Pro1,000,000$1.25$10—
Gemini 2.5 Flash1,000,000$0.30$2.50—

Gemma Series (Open-Weight)

ModelParamsContextInput / 1M
Gemma 4 26B-a4B26B (4B active MoE)262,144$0.12
Gemma 4 31B31B262,144$0.12
Gemma 3 27B27B262,144$0.10
Gemma 3 12B12B131,072$0.05
Gemma 3 4B4B131,072$0.05

🔵 Meta — Llama & Muse

ModelContextCatatan
Llama 4 Maverick1,048,576Flagship reasoning
Llama 4 Scout1,310,720Termurah & tercepat (2,600 t/s)
Meta Muse Spark—Menggantikan Llama sejak Apr 2026
Llama 3.3 70B131,072Workhorse fine-tune

🟣 Mistral AI

ModelContextCatatan
Mistral Large 2512262,144Flagship multilingual
Mistral Small 3.2 24B256,000Efisien & cepat
Codestral 2508256,000Khusus coding
Ministral 14B262,144Edge / mobile
Ministral 3B131,072Ultra-ringan
Mixtral 8x22B65,536MoE classic
Voxtral Small 24B32,000Voice-first model

🌊 DeepSeek

Gebrakan dari China — V4 Flash $0.14 / $0.28 per M token! Peringkat #6 global.

ModelContextInput/1MOutput/1MSpeed
DeepSeek V4 Flash BEST VALUE1,000,000$0.14$0.28107.9 t/s
DeepSeek V4 Pro1,000,000$0.435$0.87174.9 t/s
DeepSeek R1 (Reasoning)163,840$0.70$2.50—
DeepSeek V3.2163,840$0.27$0.40—

⚫ xAI — Grok

ModelContextCatatan
Grok 4.202,000,000Context window TERBESAR di dunia
Grok 4.20 Multi-Agent2,000,000Multi-agent mode
Grok 4.31,000,000Standar
Grok 4.5500,000Latest alias

🔴 China Players

Alibaba — Qwen

ModelArchitectureContext
Qwen3 235B-a22BMoE (22B aktif)262,144
Qwen3 32BDense131,072
Qwen3 CoderMoE/Dense262K – 1M
Qwen PlusAPI flagship1,000,000

Moonshot — Kimi

ModelContextInput/1MSpeedCatatan
Kimi K2.6256,000$0.95342.6 t/s#5 global HLE (54%)
Kimi K2.5128,000$0.55337.7 t/sVersi sebelumnya

Zhipu — GLM

ModelContextInput/1MSpeedCatatan
GLM 5.21,000,000$0.95347 t/s#4 global HLE

MiniMax & ByteDance

CompanyModelContextInput/1M
MiniMaxM31,048,576$0.60
ByteDanceSeed 2.0 Lite262,144—
ByteDanceSeed 1.6 Flash262,144$0.075

🟣 Nous Research — Hermes Ecosystem

Lab AI open-source — bikin Hermes Agent, Hermes Models, Nous Portal.

Hermes Agent 🧠

FiturDetail
LisensiMIT — 100% open source
GitHub219k stars · 41.5k forks
Learning LoopMemory system, Skills, Session Search — otomatis belajar
PlatformCLI + 20+ platform (Telegram, Discord, WhatsApp, Signal, Teams, Email, SMS, dll)
TerminalLocal, Docker, SSH, Daytona, Modal, Singularity
Tools60+ built-in tools + MCP integration
DelegationSubagents parallel, programmatic tool calling
SmartCron jobs, Voice Mode, SOUL.md, Context Files

Nous Portal 🔌

One subscription → 300+ model + Tool Gateway (web search, image gen, TTS, browser). Gak perlu kumpulin API keys.

Hermes Models (HuggingFace)

ModelParamsDownloads
Hermes-4-14B14B194k+
Hermes-3-Llama-3.1-405B405B—
Hermes-4-405B / 70B405B / 70B—
Hermes-3-Llama-3.1-8B8B462k+
Nous-Hermes-2-Mixtral-8x7B-DPO8x7B MoE38k+
Hermes-4.3-36B36B10k+

💡 Token System di Hermes

📦 Other Players

Cohere

Command-A

256K ctx, $2.50/$10.

Amazon Nova

Pro/Lite/Micro

Micro $0.035/$0.14. Premier 1M ctx.

Perplexity Sonar

Pro/Deep Research

Search & research specialist.

Microsoft Phi-4

14B

16K ctx. Small & kompeten.

Inflection-3

Pi

Personal AI, 8K ctx.

AI21 Jamba

Large 1.7

256K ctx, SSM-Transformer hybrid.

🔤 Tokenizers & Token System

Apa Itu Token?

Token adalah unit dasar teks yang diproses model AI. Model tidak membaca huruf atau kata — mereka membaca token. Ibarat "LEGO" bahasa.

Contoh tokenization (tiktoken):
"Halo, apa kabar?" → [Halo] [,] [ apa] [ kab] [ar] [?] = 6 token

🧮 Konversi

HubunganKira-kira
1 token ≈0.75 kata (English)
1 token ≈4 karakter
1 halaman buku ≈660 token
Novel (100k kata) ≈130.000 token

⚙️ Jenis Tokenizer

TokenizerModelCara Kerja
BPEGPT, Claude, Llama, DeepSeek, QwenKarakter → gabung pasangan tersering → ulang
WordPieceBERTMirip BPE tapi likelihood-based merging
UnigramXLNet, T5Mulai banyak subword → pruning jarang dipakai
SentencePieceLlama 2/3, Gemma, T5Training tanpa pre-tokenized text

📊 Vocabulary Size

ModelVocab SizeTokenizer
GPT-4o / GPT-5~100,000tiktoken (o200k_base)
Claude~100,000BPE custom
Llama 3 / 4128,000BPE (tiktoken-compatible)
Gemini / Gemma256,000SentencePiece
Mistral32,000BPE (SentencePiece)
Qwen3152,064BPE (tiktoken-based)

🌍 Token Efficiency per Bahasa

BahasaToken / 100 kata
English~133 ✅ Paling efisien
Bahasa Indonesia~145 ✅ Mendekati English
Mandarin~250-350 ❌ 2-3x lebih boros
Jepang~300-400 ❌ Paling boros

💡 AI dalam Bahasa Indonesia relatif murah dibanding Mandarin/Jepang!

💰 Perbandingan Harga Lengkap

#ModelInputOutputContextValue
1DeepSeek DeepSeek V4 Flash$0.14$0.281M⭐⭐⭐⭐⭐
2Google Gemini 3.1 Flash Lite$0.25$1.501M⭐⭐⭐⭐⭐
3Google Gemini 3 Flash$0.50$31M⭐⭐⭐⭐
4MiniMax MiniMax M3$0.60$2.401M⭐⭐⭐⭐⭐
5Zhipu GLM 5.2$0.95$31M⭐⭐⭐⭐⭐
6Moonshot Kimi K2.6$0.95$4256K⭐⭐⭐⭐⭐
7OpenAI GPT-5.6 Luna$1$61.05M⭐⭐⭐⭐
8Google Gemini 3.5 Flash$1.50$91M⭐⭐⭐⭐
9Google Gemini 3.1 Pro$2$121M⭐⭐⭐⭐
10OpenAI GPT-5.6 Terra$2.50$151.05M⭐⭐⭐
11Anthropic Claude Sonnet 5$3$151M⭐⭐⭐
12Anthropic Claude Opus 4.8$5$251M⭐⭐⭐
13OpenAI GPT-5.6 Sol$5$301.05M⭐⭐⭐
14Anthropic Claude Mythos 5$10$501M⭐⭐
15Anthropic Claude Fable 5$10$501M⭐⭐
16OpenAI GPT-5.5 Pro$30$1801M⭐

💡 Best Value: DeepSeek V4 Flash ($0.14/$0.28). 🚀 Fastest: GLM 5.2 (347 t/s). 🏆 Top Quality: Claude Mythos 5.

📰 AI News & Trends — Juli 2026

🔥 1. Claude Dominasi Semua Benchmark

Mythos 5 juara HLE (64.5%), SWE-Bench (95.5%), AutoBench, OSWorld, BrowseComp, Terminal-Bench. Pertama kali satu perusahaan menang di hampir semua kategori.

⚡ 2. Era 1M+ Context Window

Hampir semua flagship punya 1M+ context. Grok 4.20 bahkan 2M. DeepSeek V4 Flash 1M cuma $0.14/M input — demokratisasi long-context.

🇨🇳 3. China Rising

GLM 5.2 #4, Kimi K2.6 #5, DeepSeek V4 Flash #6 — China mendominasi value-for-money. Harga super murah dengan performa global.

🧪 4. AI Agents → Produksi Massal

Claude Cowork, Claude Code, OpenAI Codex, Hermes Agent — semua bergerak ke autonomous agents. Computer Use, Terminal Use, Browsing jadi fitur wajib.

🔬 5. Interpretability Breakthrough dari Anthropic

Anthropic menemukan "global workspace" di Claude — sistem internal yang mirip kesadaran kerja. Model bisa menyimpan internal thoughts yang tidak muncul di output. Langkah besar menuju AI transparan.

💰 6. Harga Makin Kompetitif

Dari $180/M output (GPT-5.5 Pro) sampai $0.14/M input (DeepSeek V4 Flash) — selisih 1.285x. Model murah sekarang sudah sangat capable.

🏗️ 7. MoE Jadi Standar

Qwen3 235B-a22B, Gemma 4 26B-a4B, Mixtral — hanya sebagian parameter aktif per inference. Efisien tanpa korbankan kualitas.

📖 Glossary

Token
Unit dasar teks yang diproses model AI. Bisa kata, subword, atau karakter.
Context Window
Jumlah maksimal token yang bisa diproses model dalam satu sesi.
Tokenizer
Algoritma yang mengubah teks jadi token. Setiap model punya tokenizer sendiri.
BPE (Byte-Pair Encoding)
Tokenizer paling populer (GPT, Claude, Llama). Mulai dari karakter → gabung pasangan tersering.
MoE (Mixture of Experts)
Arsitektur dengan banyak "expert" subnetworks. Hanya beberapa aktif per token.
HLE (Humanity's Last Exam)
Benchmark tersulit untuk AI — pertanyaan yang bahkan ahli manusia sulit jawab.
SWE-Bench
Benchmark coding agen — model harus selesaikan PR/bug fix di repo nyata.
TTFT (Time to First Token)
Waktu dari request sampai token pertama diterima. Ukuran latensi.
t/s (tokens per second)
Kecepatan output model — token per detik.
Reasoning Model
Model yang "berpikir" dulu sebelum jawab (chain-of-thought internal).
Fine-tuning
Melatih ulang model dengan data spesifik untuk domain tertentu.
DPO (Direct Preference Optimization)
Teknik fine-tuning tanpa RLHF kompleks, bikin model sesuai preferensi manusia.
Quantization
Mengompresi model (16-bit → 4-bit). Ukuran turun 4x, kecepatan naik.

🔬 AI Ecosystem 2026 — Riset oleh Hermes Agent (Belajar AI Profile)

Sumber: Vellum LLM Leaderboard, OpenAI, Anthropic, GitHub NousResearch, HuggingFace, OpenRouter

📅 23 Juli 2026 | 15+ perusahaan, 200+ model

💡 Hermes Agent-mu sendiri adalah produk Nous Research — AI yang belajar dari pengalaman 🔄

◀
Sebelumnya
Sesi 0: Sejarah AI
 🏠 Beranda
Selanjutnya
Sesi 2: Praktik Langsung
 ▶
🏠 0 1 2 3 4