AI Platforms and Model Comparison

Last reviewed: September 2026 | This is a fast-moving area subject to quarterly review.

Overview

Note

New to AI? Read Getting Started with AI first. This document focuses on service comparison.

From Traditional ML to Generative AI

Generation	Core Tech	Characteristics	Cloud Examples
Traditional ML	Regression, classification, clustering	Structured data, feature engineering required	SageMaker, Azure ML, Vertex AI
Deep Learning	CNN, RNN, Transformer	Unstructured data (image, text, speech). GPU required	GPU instances, managed training platforms
Generative AI	Foundation models (LLM, multimodal)	Text/image/code generation. Used via API	Bedrock, Microsoft Foundry, Gemini
Agentic AI	LLM + tool use + autonomous execution	Plans, executes, verifies given a goal	AgentCore, Foundry Agents, details→

Note

The table above is a deep-dive from a model-selection and technical-characteristics angle. For an introductory overview of what each generation is and which way of using AI it leads to, see Getting Started with AI — Background: The AI Technology Lineage.

When This Helps

• Chatbot/support automation — Responding to customer inquiries around the clock
• Document summarization/classification — Rapidly analyzing large volumes of reports, emails, and contracts
• Translation/content generation — Multilingual support, marketing copy generation, automated product descriptions
• Code writing/review — Improving developer productivity, detecting security vulnerabilities
• Data analysis — Gaining data insights through natural language questions

On-premises AI/ML requires purchasing GPU servers, installing frameworks, and building training infrastructure yourself. In the cloud, you can rent GPUs by the hour and train and deploy models on managed platforms.

Generative AI Model Types

Type	Input → Output	Representative Services	Use Cases
Text (LLM)	Text → Text	GPT-6 class, Claude Opus 5.5, Gemini 3.x class	Chatbots, summarization, code generation
Image Generation	Text → Image	DALL-E, MAI-Image, Imagen, Titan Image	Marketing, design
Speech (TTS/STT)	Text ↔ Speech	Polly, Nova Sonic, MAI-Voice, Azure Speech, Cloud TTS	Transcription, IVR, accessibility, voice agents
Video Generation	Text → Video	Nova Reel, Veo 3.1, Gemini Omni	Ads, short-form content
Video Understanding	Video → Text	Bedrock (TwelveLabs Pegasus), Vertex AI (Gemini), Azure AI Content Understanding, OCI Generative AI (Video Understanding)	Video search, scene analysis, captioning/summarization
Multimodal	Text+Image+Speech → Text	GPT-6 class, Gemini 3.x class, Claude Opus 5.5	Document understanding, image analysis
Embeddings	Text/Image → Vector	Titan Embeddings, Gemini Embedding, Cohere Embed	RAG, similarity search

Note

Video understanding describes, searches, and summarizes video as text — a different role from embeddings that turn video into vectors (e.g., TwelveLabs Marengo). Availability and status vary by cloud: regions or eligible customers may be limited (e.g., Bedrock offers it via cross-region inference only), and Preview/GA stages differ, so check each vendor’s official docs for current status before adoption.

Decision Models (System One, Non-Generative)

Most of the types above generate content, but since September 2026 a class of models that does not generate text — instead returning structured decisions to predefined questions — has been forming rapidly. Vendors call these “decision models” or “System One” models. They take a state and typed questions and return choices, scores, or probabilities with calibrated confidence, so software can branch on the result directly without parsing free text. In exchange for giving up flexibility, they offer low latency, high efficiency, and answers always drawn from the enumerated options; they are unsuited to complex reasoning or generation (coding, summarization, chatbots).

Model	Provider	Form	Notes
Jev	TypeSafe AI	Hosted	The model TypeSafe named “System One” and released in mid-September 2026. Typed questions (Noul, Choice, Score) returning probability/confidence decisions. 32,000-token context, $0.042 per 1M input tokens / no output charge, zero data retention. Offered via Cloudflare Workers AI
GLiNER2.5-Decide	Fastino Labs	Open-weight (Apache 2.0)	Released late September 2026. Per the model card, 340M parameters that run on a CPU without a GPU. DeBERTa-v3 encoder-based, specialized for operational decisions such as routing, sentiment, yes/no gates, and ordinal scores. Returns probability distributions, confidence, and constraint-feasibility metadata. A decision derivative of the open-source GLiNER2 family
Strands Decider 2B	Strands Labs	Open-weight (Apache 2.0)	Released early October 2026. Removes the generation head from a Qwen3.5-2B decoder and adds a pointer head (~1M parameters) that selects one of the enumerated options. Runs locally (CPU, CUDA, MPS, MLX) and returns confidence from the option distribution. The release ships the adapter and head; the Qwen base weights are downloaded separately. Latency varies widely by task and hardware per the vendor’s own measurements (e.g., ~115ms overall median on an RTX 3090). Blog · GitHub

The models above share a similar input/output contract but differ in deployment form (hosted vs. open-weight) and runtime (managed API vs. local CPU/GPU). The selection axes are “where will it run (data-residency constraints, edge or not)” and “required accuracy vs. latency.”

Note

Decision models suit tasks where the decision space is clearly enumerated — classification, routing, policy adjudication. For free-form narration, summarization, or generation, a generative LLM is still the right fit; the two are complementary, not substitutes. Since Jev’s September 2026 debut this category has grown quickly, including open-source reimplementations, so validate the supported question types, evaluation accuracy, and calibration quality against your own data before adoption. Many published benchmark figures are vendor self-measured, so don’t take ranking or superiority claims at face value without independent verification.

Generative AI Services

Foundation Model APIs

Provider	Key Models	1P (Direct)	3P (Cloud-hosted)
OpenAI	GPT-6 family (Astra flagship / Sol mid / Luna low-cost), GPT-5.6, o-series	api.openai.com	Azure Foundry, Bedrock
Anthropic	Claude Fable 5.1, Opus 5.5, Sonnet 5, Haiku 5.5	api.anthropic.com	Bedrock, Vertex AI
Google	Gemini 3.x generation (Pro/Flash — check official docs for generation/variant and Preview/GA status), Gemini Omni	Gemini API	Vertex AI (native)
SpaceXAI (formerly xAI)	Grok 4.7, Grok Imagine	x.ai/api	OCI, Vertex AI, Bedrock, Azure
Moonshot AI	Kimi K3 (open-weight, multimodal)	platform.moonshot.ai	Bedrock, OCI (import), self-host
Meta	Llama 4 (open-weight)	llama.com	Bedrock, Vertex, Azure, OCI
Amazon	Nova 1 (Premier/Pro/Lite/Micro/Sonic, etc.) + Nova 2 (Lite GA, Pro Preview, Nova 2.5 Sonic speech GA)	— (Bedrock only)	Bedrock
Microsoft	MAI (Image/Voice/Transcribe)	— (Foundry only)	Azure Foundry
Mistral	Large, Small, Codestral	api.mistral.ai	Bedrock, Azure, Vertex
Upstage	Solar Pro 4/3, Solar Mini 4, Solar Open 2	console.upstage.ai	AWS/Azure Marketplace
LG AI Research	EXAONE 4.x	Direct contract	Marketplace, self-host

Note

1P vs 3P difference — The same model may differ in feature scope, quotas, and billing depending on channel. See LLM Channel Selection Guide.

Note

Anthropic tiers and version numbers — Claude Fable 5.1 is a higher (harder) tier than Opus 5.5. Tiers revise on different cadences, so a higher tier’s minor version can lead a lower one (e.g., Fable 5.1 > Opus 5.5). Don’t infer recency from the number alone.

Cloud Platform Strengths

Platform	Strengths
Amazon Bedrock	Multi-model single API, AgentCore, AWS IAM/VPC integration, EDP consumption
Microsoft Foundry	Primary OpenAI channel, M365/GitHub integration, Foundry Local (air-gapped), PTU
Vertex AI / Gemini Platform	Native Gemini, 2M token context, Model Garden 200+ models, ADK
OCI Enterprise AI	Oracle DB integration, dedicated GPU clusters (RDMA), 10TB egress free

AI Agents / RAG

Vendor	Agent Platform	RAG
AWS	Bedrock AgentCore	Bedrock Knowledge Bases
Azure	Microsoft Foundry Agents	Azure AI Search
Google Cloud	Gemini Enterprise Agent Platform	Vertex AI RAG Engine
OCI	OCI Enterprise AI Agents	OCI Search integration

Code Assistants / AI Agents

Coding agents (Kiro, Claude Code, Codex, Copilot, etc.) and agent platforms (AgentCore, Foundry Agents, etc.) are covered in AI Agents.

ML Platforms

For organizations that need to train and deploy their own models.

Vendor	Product	Notes
AWS	SageMaker AI	Training, tuning, deployment, MLOps
Azure	Azure Machine Learning	Notebooks, AutoML, pipelines, model registry
Google Cloud	Vertex AI	Training, deployment, pipelines, Feature Store
OCI	OCI Data Science	Notebooks, training/deployment, pipelines

GPU / AI Accelerators

Vendor GPU instances for model training/inference (Blackwell B200/GB200, Hopper H100/H200, A100), custom AI chips (AWS Trainium/Inferentia, Azure Maia, Google TPU), and reserved/spot options are consolidated — together with per-generation spec/performance and regional availability — in Multicloud AI — GPU Availability.

Once you have chosen GPU instances, how to combine them into a cluster for training/inference (network, placement, distributed training, Kubernetes, capacity, cost) is covered in the GPU Infrastructure document set.

Key Differences

Amazon Bedrock — Provides access to a wide range of provider models — including its own Amazon Nova models (1st-generation Premier/Pro/Lite/Micro/Sonic, etc., and Nova 2 Lite/Pro, etc. — check the official model list for generation and availability status) as well as Anthropic Claude, the OpenAI GPT series, and more — through a single API. It offers a broad model selection, and its strength lies in operational capabilities such as AgentCore for building AI agents.

Microsoft Foundry — Formerly Azure AI Foundry, now the unified upper-level platform that consolidated the brand. It is the primary channel for using OpenAI models (the GPT-6 family and the GPT-5.6 family) in enterprise environments, and it also offers a wide range of third-party models such as Anthropic and Meta. It has added its own MAI model family (Image-2.5, Voice-1, Transcribe-1) and Foundry Local (for local/air-gapped execution). Its greatest strength is deep integration with the existing Microsoft ecosystem, including Microsoft 365, GitHub, and Power Platform.

Gemini Enterprise Agent Platform — A full agent-centric overhaul of the former Vertex AI. Its strengths are the native multimodal capabilities of Google’s own Gemini 3.x generation series (Pro/Flash — check the official documentation for each variant’s generation and Preview/GA status and limits) and its TPU infrastructure. Its differentiators are long context, reasoning mode, Gemini Omni (multimodal), low-code agent development through Agent Studio, and integration with Google Search/BigQuery.

OCI Enterprise AI — An expanded platform that evolved from the former OCI Generative AI. It hosts models such as Cohere, Meta Llama, SpaceXAI Grok, and Google Gemini on OCI infrastructure, and supports high-performance workloads with Dedicated AI Clusters and RDMA-based Bare Metal GPUs. It has added AI Guardrails (content moderation, PII detection, prompt injection defense) and Enterprise AI Agents (GA). Through its partnership with OpenAI, OpenAI models (API) and Codex are available on the OCI Marketplace via Oracle Universal Credits, and its strength is native integration with Oracle Database and applications.

AI System Lifecycle & Enterprise Adoption Matrix

The 7-stage end-to-end AI System Life Cycle, 4-Tier adoption matrix with control boundaries, technical task selection guide, and MLOps vs LLMOps operational frameworks are detailed comprehensively in AI System Lifecycle and Engineering.

Where AI Adoption Is Expanding

Applied AI (Industry-Specific Turnkey Services)

Area	Vendor Services	Trend (2025–2026)
Contact center	Amazon Connect, Azure Contact Center, Google CCAI	Copilot → autonomous agent transition
Document processing	Textract, Document Intelligence, Document AI	Combined with LLM/multimodal reasoning
BI	Amazon Quick, Copilot in Power BI, Gemini in Looker	Agentic analytics, dashboard agents
Healthcare	Amazon Connect Health, Azure Health Bot	HIPAA-compliant agents

Physical AI (Connecting to the Physical World)

Physical AI — connecting AI to the physical world of sensors, robots, and equipment (edge inference, digital twins and simulation, robotics foundation models) — is covered in detail from a vendor-neutral perspective in the dedicated Physical AI document.

Multicloud Model Access (2025–2026)

Event	Impact
OpenAI-Microsoft exclusivity ended (2026.04)	OpenAI models available on non-Azure platforms
OpenAI → Bedrock (2026.04)	GPT-5.x available on Bedrock post-exclusivity
SpaceXAI (formerly xAI) Grok multicloud expansion	Available on Azure, Vertex AI, OCI, Bedrock (Grok 4.7, including GovCloud in 2026.10)
Anthropic Claude channel expansion	Beyond Bedrock/Vertex to additional channels
Flagship generation refresh (2026.09)	OpenAI GPT-6 family (Astra/Sol/Luna) and Anthropic Opus 5.5 rolling out across API, Bedrock, Vertex, and Foundry

Inference Cost Optimization

Strategy	Description	Vendor Support
Flex/Batch Inference	Process latency-tolerant workloads at lower priority	Bedrock Flex, Azure Batch API, Vertex Batch Predictions
Model Routing	Route simple queries to lightweight models, complex ones to frontier models	Bedrock IntelligentPromptRouter, custom
Prompt Caching	Cache repeated system prompts/context to reduce token costs	Anthropic Prompt Caching, OpenAI Cached Tokens, Gemini Context Caching
Long Context vs RAG	Extended context windows (1–2M+ tokens) may eliminate RAG need	Gemini 3 Pro class, Claude Opus
GPU Price Competition	Hyperscaler GPU instance pricing trending downward	AWS, Azure, GCP competitive pricing

Note

Inference pricing changes frequently. Check each vendor’s official pricing page for current rates. Cost tracking and budget management are detailed in LLMOps.

Common Mistakes

• Starting with fine-tuning — RAG often suffices; fine-tuning wastes cost/time on problems RAG can solve.
• Not pinning model versions — Vendor model updates can silently degrade production prompt quality.
• Single model for all workloads — Using frontier models for simple classification tasks inflates costs. Route by task complexity.

Checklist

• Selected models matching workload characteristics and compared cost/quality
• Pinned model ID/version in code; upgrades go through evaluation before rollout
• Following staged approach: RAG → Fine-tuning → Train from scratch

References

AWS

• Amazon Bedrock
• Amazon SageMaker AI

Azure

• Microsoft Foundry
• Azure Machine Learning

Google Cloud

• Vertex AI
• Gemini API

OCI

• OCI AI Services
• OCI Enterprise AI