Overview: The 2026 Model Market

Key Points

• Frontier closed-source: OpenAI (GPT-5 / o-series), Anthropic (Claude 4 family), Google (Gemini 2.5).
• Strong open-weights: Meta Llama ¾, Mistral, DeepSeek (R1 reasoning), Qwen, Alibaba.
• Small + edge: Microsoft Phi-3/Phi-4, Llama small, Gemma.
• Specialty: embeddings (text-embedding-3, voyage-3, Cohere), rerankers, vision, audio.
• Multi-vendor strategy is the senior default — pin to provider only after measurement; abstract via IChatClient.

Frontier model families (2026)

OpenAI

• GPT-5 family: flagship reasoning + multimodal.
• GPT-4o / 4o-mini: cost-balanced; 128K context.
• o-series (o1, o3): reasoning models — explicit chain-of-thought; better math/code/logic.
• Embedding: text-embedding-3-small / -large.
• Whisper: audio transcription.

Anthropic

• Claude 4 Opus / Sonnet / Haiku: tier of capability/cost.
• 1M context in select models.
• Computer use (Claude operates a browser/computer).
• Long-context strengths — extended reasoning over big docs.

Google

• Gemini 2.5 Pro / Flash: multimodal native (text, image, video, audio).
• 2M context in Pro tier.
• Vertex AI for managed access; AI Studio for direct API.

Open weights

• Meta Llama 3.x / 4: state-of-art open; self-hosted or via Together/Groq/Replicate.
• Mistral: French; competitive small models (Mistral Large; Mixtral MoE).
• DeepSeek R1: reasoning-focused; open weights.
• Qwen (Alibaba): strong Chinese + multilingual; competitive English.

Microsoft Phi (small)

• Phi-3 / Phi-4: SLMs (Small Language Models). 3-14B params.
• Edge / on-device via ONNX Runtime GenAI.
• "Surprisingly good" for size.

What "best" means depends on workload

See Model Selection Decision Matrix.

Pricing (rough, 2026)

GPT-5:                $15-30 / M input  (most expensive)
GPT-4o:               $2.50 / M input
GPT-4o-mini:          $0.15 / M input  (very cheap)
Claude 4 Opus:        $15 / M input
Claude 4 Sonnet:      $3 / M input
Claude 4 Haiku:       $0.25 / M input
Gemini 2.5 Pro:       $3.50 / M input
Gemini 2.5 Flash:     $0.30 / M input
Llama 3 (Together):   $0.20-1 / M input
Self-hosted Llama:    fixed GPU $$/hr

Output ~3-5x input. Prompt caching (Anthropic, OpenAI) gives huge discount on repeated prefixes. See Pricing & Cost Engineering.

Capability tiers (informal)

Tier S (frontier flagships):   GPT-5, Claude 4 Opus, Gemini 2.5 Pro, o3
Tier A (production workhorses): GPT-4o, Claude 4 Sonnet, Gemini 2.5 Flash, Llama 3.3
Tier B (cost-sensitive smart): GPT-4o-mini, Claude Haiku, Gemini Flash 2.5, Llama 3.x small
Tier C (small/edge):           Phi-4, Llama 3 8B, Mistral 7B

Reasoning models

OpenAI o-series, DeepSeek R1, Claude with extended thinking. Trained for explicit chain-of-thought. Slower per token; better on hard reasoning. Different cost profile (more output tokens for thinking).

When to reach for reasoning: math, complex coding, multi-step planning. Don't use for simple Q&A — wasted cost.

Multimodal

For vision-heavy: GPT-4o or Gemini 2.5. For document images: Claude 4 (strong OCR-like ability).

Self-hosted vs API

For most teams: API. Self-host only with measured cost win + ops capacity.

Geographic / compliance

• Azure OpenAI: in your Azure region, data residency.
• AWS Bedrock: hosts Anthropic, Meta, Mistral, Amazon Nova.
• GCP Vertex AI: Gemini, Anthropic, others.
• Oracle: hosts some.
• Sovereignty regions: gov, EU, China.

Model cycle

Models age fast. Today's flagship → tomorrow's cheap commodity. Build vendor-portable code (IChatClient) so you can switch.

Senior strategy

1. Default to a frontier provider (Azure OpenAI typically).
2. Prototype with cheap tier (gpt-4o-mini).
3. Measure quality / latency / cost.
4. Promote to expensive only where measurably better.
5. Multi-provider fallback for resilience.
6. Keep abstraction layer (IChatClient).

Cross-references

• OpenAI Models
• Anthropic Claude
• Google Gemini
• Meta Llama & Open Source
• Microsoft Phi & Small Models
• Specialty Models
• Model Selection Decision Matrix
• Pricing & Cost Engineering
• Vendor Trade-offs & Lock-In