Proprietary Models — Catalog

Comparative reference. Not available for local fine-tuning, but important as a baseline and for use via API.


GPT (OpenAI)

Version Context Highlights
GPT-3 (2020) 2K 175B; in-context learning; revolutionized NLP
GPT-3.5 Turbo 16K Base of ChatGPT; speed and cost
GPT-4 (Mar/2023) 128K Multimodal; advanced reasoning
GPT-4o (May/2024) 128K Native multimodal (textimageaudio); faster
GPT-5 (Aug/2025) 512K+ General SOTA; multiple sizes (55-mini5-nano)
GPT-5.3 Codex Specialized in code; 85% SWE-bench
GPT-5.5 (Apr/2026) Latest release; reasoning + computer use

SWE-bench Verified: GPT-5 74.9% Docs: platform.openai.com/docs


Claude (Anthropic)

Version Context Highlights
Claude 3 HaikuSonnetOpus (Mar/2024) 200K Capability tiering; vision 3.5+
Claude 3.5 Sonnet (Jun-Oct/2024) 200K Computer use; general SOTA in 2024
Claude 3.7 Sonnet (Feb/2025) 200K Extended thinking; reasoning model
Claude Sonnet 4 / Opus 4 (May/2025) 200K Hybrid reasoning
Claude Sonnet 4.5 / Opus 4.5 (Sep-Nov/2025) 200K (1M preview) Advanced multimodal
Claude Opus 4.7 (Apr/2026) 1M 87.6% SWE-bench; vision 3.75MP; self-verification
Claude Mythos Preview (2026) 1M 93.9% SWE-bench; not public via API

Constitutional AI: the alignment basis of all Anthropic models Docs: docs.anthropic.com


Gemini (Google DeepMind)

Version Context Highlights
Gemini 1.0 (Dec/2023) 32K UltraProNano; first native multimodal model
Gemini 1.5 Pro (Feb/2024) 1M Breakthrough in long context; MoE
Gemini 1.5 Flash 1M Fast, economical variant
Gemini 2.0 Flash (Jan/2025) 1M New speed standard
Gemini 2.5 Pro (mid/2025) 1M "Thinking model"; step-by-step reasoning
Gemini 3 Pro (2025) 1M Led the LMArena Leaderboard
Gemini 3 Deep Think 1M Reasoning variant; iterative multi-path
Gemini 3.1 Pro (Feb/2026) 1M+ GPQA Diamond 94.3%; ARC-AGI-2 77.1%; native MCP

Gemini 3.1 Pro — Benchmarks (February 2026):

Benchmark Score
SWE-bench Verified 78.80%
GPQA Diamond 94.3%
ARC-AGI-2 77.1% (2× Gemini 3 Pro)
AIME 2026 89.2% (Gemma 4 31B)

Others:

  • Deep Research with native MCP for long-horizon research
  • Gemini 3.2 announced for Google Cloud Next 2026 (context >1M tokens, lower latency)

Video: Gemini 3.0 processes video at 60fps in real time Docs: ai.google.dev


Grok (xAI)

Version Context Highlights
Grok-1 (open) 8K 314B MoE; Apache 2.0; JAX
Grok-3 131K Advanced reasoning
Grok-4 256K Frontier; $3/1M input
Grok-4.1 2M (4.1 Fast) Very low hallucination rate; speed
Grok-4.3 Beta (Apr/2026) 2M Native video; slide creation; 16-agent Heavy

Grok-4.3: Released on 04172026 without a formal announcement; activates via SuperGrok Heavy ($300/month). New features: native video processing, slide generation, keeps the 16-agent system and the 2M-token window from 4.1.

Grok 5: In training on Colossus 2 (1.5 gigawatts of compute); release expected for Q2 2026.

Docs: docs.x.ai


MAI (Microsoft AI)

Microsoft's first family of own foundation models, announced at the Microsoft AI keynote at Build 2026 (06022026). It marks the strategic turn from "consuming third-party models (OpenAIAnthropic)" to "building, hosting and controlling the economics of intelligence" — all trained from scratch with commercially licensed data, without distillation of third-party frontier models (a deliberate choice to avoid legalcommercial risk and dependence).

Model Type Highlights
MAI-Thinking-1 Reasoning (MoE) 35B active params, 256K context; multi-step reasoning, long-context, code generation; private preview in Microsoft Foundry
MAI-Code-1-Flash Agentic coding ~5B params; text → source code (apps/sites); inference-efficient; integrated into GitHub Copilot + VS Code
MAI-Image-2.5 (+ Flash) Image Text-to-image and image-to-image (description or sketch); embedded in PowerPoint
MAI-Transcribe-1.5 ASR High-precision transcription, 43 languages; real-time streaming "soon"
MAI-Voice-2 (+ Flash) TTS 15+ additional languages; new voices

MAI-Thinking-1 performance claims (Microsoft, Build 2026 — not independently verified):

  • In a blind evaluation by Surge (independent human evaluators), preferred over Claude Sonnet 4.6.
  • Matches Claude Opus 4.6 in coding on SWE-Bench Pro.
  • Mustafa Suleiman (CEO Microsoft AI): after fine-tuning for McKinsey, it surpassed GPT-5.5 in quality with ~10× more cost efficiency — ⚠️ estimate derived from public pricing across model scales, not a production measurement.

Why it matters for Kode: direct external evidence for the thesis of policies/self-hosted-first.kmd / Flipping Point — a hyperscaler that was 100% dependent on rented models (invested US\( 13 bn in OpenAI, US\) 5 bn in Anthropic) publicly declaring that it will build its own models for costmargincontrol. The "clean licensed data without distillation" argument also reinforces 04-training/data-pipeline.md (provenance) as a competitive differentiator, not just compliance. MAI-Code-1-Flash is Microsoft's analog to the positioning that Kode occupies (a coding model integrated into the editor).

Sources: MAI Build 2026 Keynote (transcript) · CNBC, 06022026 · TechTimes — MAI-Thinking-1


Other Relevant Proprietary Models

Model Company Highlights
Command R / R+ Cohere Specialized RAG; 128K; structured outputs
Orca 2 Microsoft Reasoning in small models (predecessor of the pre-MAI era)
Grok-1 (open) xAI 314B MoE; the only large-scale open proprietary one

SWE-bench Verified Comparison (April 2026)

Model Score
Claude Mythos Preview 93.9%
Claude Opus 4.7 87.6%
GPT-5.3 Codex 85.0%
Gemini 3.1 Pro 78.80%
Claude 4 77.2%
GPT-5 74.9%

Note: Kimi K2.6 (open-source) scores 58.6% on SWE-Bench Pro (the harder variant).


Arena Elo Leaderboard (March 2026)

Lab Elo
Anthropic 1,503
xAI 1,495
Google 1,494
OpenAI 1,481
Alibaba 1,449
DeepSeek 1,424

  1. Context windows: 1M tokens standard at the frontier; Claude/Gemini lead
  2. Reasoning modes: All top-tier models have a "thinking" / extended reasoning mode
  3. Multimodality: Text + image + audio + video becoming standard
  4. Computer use: Claude 3.5+ and GPT-5.4+ with GUI automation
  5. SWE-bench: The 90% barrier approaching with Claude Mythos