AI Models
Guide

The versatile all-rounder with dynamic internal routing.

• GPT-5.4 is the current flagship with a 1M token context window
• Uses an internal router to select the right sub-model per request in real time
• GPT-5.4 mini and nano offer GPT-5-class performance at lower cost; mini comes within 5% of flagship on coding benchmarks
• GPT-5.4 Thinking (Mar 5, 2026): reasoning-first variant integrated into ChatGPT — strongest at math, spreadsheets, research, and document tasks; comparable to a dedicated o-series model
• GPT-5.3-Codex: dedicated agentic coding variant combining Codex and GPT-5 training stacks
• Native computer use and tool calling for agentic automation
• Strong at documentation, unit tests, and complex SQL queries

The developer favorite for coding, reasoning, and safety.

• Opus 4.6 (flagship), Sonnet 4.6 (best value), Haiku 4.5 (fast/cheap)
• Leads human-preference leaderboards; strong ARC-AGI-2 scores
• Agentic capabilities: autonomous multi-step coding tasks
• Known for safety, steerability, and high-quality long-form writing
• Claude Mythos Preview (Apr 7, 2026): released as gated preview via Project Glasswing — limited to ~50 partner orgs; 93.9% SWE-bench Verified, 97.6% USAMO 2026; not publicly available due to cybersecurity risks

Multimodal powerhouse with top benchmark breadth.

• Gemini 3.1 Pro leads the Artificial Analysis Intelligence Index
• Native multimodal: processes text, images, audio, and video natively
• 1M token context window; deep Google Workspace integration
• Gemini 3 Flash: current default model in the Gemini app — strong balance of speed and capability
• Gemini 3.1 Flash-Lite (Mar 3): $0.25/1M input, 2.5× faster than prior Flash — cost-efficient tier for high-volume workloads
• Gemini 3.1 Flash Live: audio/voice model with 90+ language support and lower latency (March 26)

Real-time data meets raw reasoning power.

• Grok 4.20 uses native multi-agent architecture with four specialist agents that debate before responding
• Grok 4.20 Beta 2 (March 3): improved instruction following and hallucination reduction
• Real-time integration with X (Twitter) for current events
• Scored 100% on AIME 2025 math competition (Heavy variant)
• Grok 5 in training — reported ~6T parameter MoE model training on xAI's Colossus 2 supercluster; Q2 2026 target

Search-native AI built for grounded, cited answers.

• Built on Llama 3.3 70B, further trained for search-grounded factuality
• Runs at 1,200 tokens/sec on Cerebras inference hardware
• Model family: Sonar, Sonar Pro, Reasoning Pro, Deep Research
• Matches GPT-4o on user satisfaction benchmarks

Frontier-level coding model trained via RL on long-horizon software engineering tasks.

• 73.7 on SWE-bench Multilingual — top scores across CursorBench and Terminal-Bench 2.0
• Trained with RL to solve complex tasks requiring hundreds of sequential actions
• Fast variant is now default: frontier-level intelligence at lower cost than competing fast models
• Background agents run tasks autonomously while you work

Microsoft's own foundation model stack — independent of OpenAI, built for speech, voice, and vision.

• MAI-Transcribe-1: speech-to-text across 25 languages; outperforms Whisper-large-v3 on accuracy
• MAI-Voice-1: generates 60s of audio in 1s; supports voice cloning
• MAI-Image-2: high-quality image generation
• All available on Microsoft Foundry — signals Microsoft building foundational AI independent of OpenAI (Apr 2, 2026)

The leading open-source model family.

• Llama 4 Scout: industry-leading 10M token context window
• Llama 4 Maverick: 17B active / 128 experts — outperforms GPT-4o and Gemini 2.0 Flash on key benchmarks
• Fully open weights; can be self-hosted for complete data control
• Llama 4 Behemoth (288B active) still in training; widely covered as next milestone

Cost-redefining open-source model with 1T parameters and conditional memory.

• Launched early March 2026: 1T MoE parameters, 1M context window, $0.30/MTok input pricing
• Engram conditional memory and Manifold-Constrained Hyper-Connections for improved long-context performance
• 81% on SWE-bench Verified; native multimodal: text, image, and video
• DeepSeek V3.2 (open-source): bridges V3 and V4 — optimized for reasoning + agentic long-context and tool-use workloads
• Entire V3/V4 lineage trained for under $6M — redefining AI cost efficiency

The enterprise-safe European model family, now spanning text, edge, and speech.

• Mistral Large 3: EU AI Act-compliant flagship for regulated industries (finance, healthcare, gov)
• Mistral Small 4 (March 17): MoE architecture, 119B total / 6B active, 128 experts — fast and efficient
• Voxtral TTS (March 26): open-source 4B text-to-speech, 9 languages, runs on consumer hardware
• Strong European data sovereignty guarantees across the full model family

Open-source agentic powerhouse with Agent Swarm.

• 1T total parameters (32B active), MoE architecture with native vision-language
• Agent Swarm coordinates up to 100 specialized sub-agents in parallel
• Kimi Code CLI agent rivals Claude Code and Gemini CLI
• Backed by Alibaba and HongShan; strong global traction

Frontier-class model on a MIT license.

• 744B parameter MoE model (44B active) with 200K context window
• Released under MIT license; trained entirely on Huawei Ascend chips
• 77.8% on SWE-bench Verified; 50.4% on Humanity's Last Exam
• Priced roughly 6x cheaper than comparable proprietary models

Open-weight models from Gemini 3 research — optimized for on-device and frontier-class performance.

• Four Apache 2.0 models: E2B (2.3B), E4B (4.5B), 26B MoE (4B active), 31B dense (Apr 2, 2026)
• 31B ranks #3 on Arena AI leaderboard at 1452 Elo — outperforms models 20× its size
• E2B/E4B optimized for on-device Android: up to 4× faster and 60% less battery than prior Gemma
• All models natively multimodal; larger variants support 256K context

The multilingual giant — now split into multimodal and agentic branches.

• Qwen3.5-Omni (Feb 2026): native audio/video/text multimodal — Thinker architecture, 256K context, 113-language speech recognition
• Qwen3.6-Plus (Apr 2, 2026): 1M context; matches Claude Opus 4.5 on SWE-bench and Terminal-Bench 2.0
• 0.6B to 235B open-weight; Qwen3-Max (1T+) API-only; supports 119 languages
• Qwen3-Coder achieves 69.6% on SWE-Bench Verified, surpassing many frontier models