01 Model Landscape 13 min read 2,753 words

Model Taxonomy

This chapter provides a comprehensive guide to the model landscape as of March 2026, covering model families, capabilities, and selection criteria for production systems.

model-selectionfundamentalsreasoningintro

01categories

Model Categories

1.1

By Capability Level (March 2026 Reality)

Tier	Characteristics	Examples	Use Case
Frontier	State-of-the-art reasoning, agentic mastery	Claude 3.7 Sonnet, GPT-4.5, o3, Grok 3	Complex reasoning, coding, production agents
Fast/Efficient	Sub-200ms, cost-optimized	Gemini 2.0 Flash, Claude 3.5 Haiku, o3-mini	High-volume streaming, UI, real-time
Battle-Tested	Mature, widely-deployed, stable	GPT-4o, Claude 3.5 Sonnet	Enterprise production workloads
Small/Edge	Private, edge, specialized	Llama 3.3 8B, Phi-4, Gemma 3	Local privacy, on-device
Reasoning-Heavy	Extended internal CoT	o3, DeepSeek-R1, Claude 3.7 (thinking)	Math, code debug, multi-step logic

1.2

By Reasoning Mode (2025–2026)

Mode	Capability	Models	Use Case
Standard	Fast, intuitive response	GPT-4o, Claude 3.5 Sonnet	Chat, simple extraction
Extended Thinking	Internal scratchpad CoT before output	Claude 3.7 Sonnet, o3, DeepSeek-R1	Math, code debugging, planning
Hybrid	User-controllable reasoning depth	Claude 3.7 Sonnet	Variable complexity tasks

02models march

Frontier Models (March 2026)

2.1

Claude 3.7 Sonnet (Anthropic)

Attribute	Value
Context Window	200K tokens
Input Cost	$3.00 / 1M tokens
Output Cost	$15.00 / 1M tokens
Extended Thinking	Native (configurable budget_tokens)
Multimodal	Text + Vision
Highlights	Top SWE-bench Verified; best production coding model
Released	February 2025

Best for: Autonomous software engineering, Claude Code agent, complex reasoning. Considerations: Enable Extended Thinking for hard tasks; use standard mode for high-volume.

2.2

GPT-4.5 (OpenAI)

Attribute	Value
Context Window	128K tokens
Input Cost	$75.00 / 1M tokens
Output Cost	$150.00 / 1M tokens
Multimodal	Text, Vision
Highlights	Highest EQ/creativity scores; strong instruction following
Released	February 2025

Best for: Creative tasks, customer-facing chat, nuanced role-following. Considerations: Expensive; not the top choice for heavy reasoning or volume.

2.3

o3 (OpenAI)

Attribute	Value
Context Window	200K tokens
Input Cost	$10.00 / 1M tokens
Output Cost	$40.00 / 1M tokens
Reasoning	High-compute internal CoT (configurable effort: low/medium/high)
Highlights	#1 on most reasoning benchmarks (ARC-AGI, SWE-bench, AIME)
Released	January 2025

Best for: Autonomous tool use, complex math/science, agentic tasks requiring deep reasoning. Considerations: Cost scales dramatically on "high" effort; use o3-mini for volume.

2.4

o3-mini (OpenAI)

Attribute	Value
Context Window	200K tokens
Input Cost	$1.10 / 1M tokens
Output Cost	$4.40 / 1M tokens
Reasoning	Lightweight reasoning (effort: low/medium/high)
Highlights	Best cost/reasoning tradeoff in its tier
Released	January 2025

Best for: High-volume reasoning tasks, coding assistance, STEM Q&A.

2.5

Gemini 2.0 Flash (Google)

Attribute	Value
Context Window	1M tokens
Input Cost	$0.10 / 1M tokens
Output Cost	$0.40 / 1M tokens
Multimodal	Native: Text, Vision, Audio, Video
Highlights	Fastest frontier model; Live API for real-time multimodal
Released	December 2024

Best for: Real-time multimodal apps, high-volume pipelines, long-context RAG.

2.6

Gemini 2.0 Pro (Google)

Attribute	Value
Context Window	1M tokens
Input Cost	$3.50 / 1M tokens
Output Cost	$10.50 / 1M tokens
Highlights	Best Google model for complex tasks; leads on MMLU
Released	January 2025

2.7

Grok 3 (xAI)

Attribute	Value
Context Window	131K tokens
Input Cost	$3.00 / 1M tokens (API preview)
Highlights	Competitive with claude/o3 on reasoning; DeepSearch for real-time web
Released	February 2026

Best for: Live web research, reasoning-heavy tasks, frontier alternative.

2.8

Model Comparison: Frontier Tier (March 2026)

Model	Reasoning	Coding	Context	Agentic	Cost
o3	★★★★★	★★★★★	★★★★	★★★★★	$$$$
Claude 3.7 Sonnet	★★★★★	★★★★★	★★★★	★★★★★	$$$
GPT-4.5	★★★★	★★★★	★★★	★★★★	$$$$$
Gemini 2.0 Flash	★★★	★★★	★★★★★	★★★	$
Grok 3	★★★★	★★★★	★★★	★★★★	$$$

2.9

Production Heritage & Maturity

While frontier models lead on benchmarks, many enterprise systems rely on battle-tested models:

Model Family	Production Since	Maturity Note
GPT-4o	May 2024	Most mature ecosystem; lowest latency variance; highest rate limits.
Claude 3.5 Sonnet	June 2024	Gold standard for tool-use reliability and structured output.
Gemini 1.5 Pro	May 2024	Pioneer of mass-scale context; highly stable for long-document analysis.
o1	Sept 2024	First production reasoning model; well-understood failure modes.

Why stay on "older" frontier models?

Consistency: New models have "release-window" latency spikes and behavior shifts.
Cost Efficiency: Previous generation is often 50-80% cheaper after a new release.
Guardrail Tuning: Security and moderation layers are more refined.

03source models

Open Source Models

3.1

Llama 3.3 Family (Meta)

Model	Parameters	Context	License	Notes
Llama 3.3 8B	8B	128K	Llama 3.3	Best-in-class small model (March 2026)
Llama 3.3 70B	70B	128K	Llama 3.3	Strongest open-weight general model
Llama 3.1 405B	405B	128K	Llama 3.1	Largest Meta model; frontier-competitive

Strengths:

Native multimodality across sizes (Llama 3.2 vision variants)
Excellent tool-use and JSON following
Largest open community and tooling ecosystem

3.2

DeepSeek Family

Model	Parameters	Status	Notes
DeepSeek-V3	671B (MoE)	Frontier	GPT-4o level at a fraction of training cost; open weights
DeepSeek-R1	671B (MoE)	Reasoning	Matches o1 on math/code; first open-source reasoning model
DeepSeek-R1-Distill	7B–70B	Reasoning	Distilled to smaller models; cost-efficient reasoning

Key 2026 context: DeepSeek shocked the industry by demonstrating frontier-level performance at dramatically lower training costs. Open weights available on Hugging Face.

3.3

Qwen 2.5 Family (Alibaba)

Model	Parameters	Notes
Qwen2.5-Coder-32B	32B	Top open coding model; rivals GPT-4o on HumanEval
Qwen2.5-72B	72B	Best multilingual open model; strong CJK support
Qwen2.5-7B	7B	Efficient self-hosted option

3.4

Mistral Family

Model	Parameters	Notes
Mistral Large 2	123B	Strong reasoning; permissive license
Mistral Small 3	24B	Ultra-efficient; matches 7B quality at less compute
Mixtral 8x22B	141B (MoE)	Best open MoE for throughput

04models

Specialized Models

4.1

Coding Mastery (March 2026)

Model	Specialization	Why it wins
Claude 3.7 Sonnet	Software Engineering	Highest SWE-bench Verified score; powers Claude Code
o3	Algorithmic reasoning	Best at complex multi-step code logic; USACO problems
Qwen2.5-Coder-32B	Open-source coding	Best price-to-performance for self-hosted IDEs
DeepSeek-R1-Distill-70B	Open reasoning+code	Best open reasoning model for coding at 70B

4.2

Reasoning & Math

Model	Approach	Best For
o3	High-compute internal CoT	ARC-AGI, competition math, autonomous agents
Claude 3.7 Sonnet (thinking)	Extended Thinking mode	Software planning, complex logic
DeepSeek-R1	RL-based thinking	Open-source logical inference, competitive math
Grok 3 (DeepSearch)	Web-grounded reasoning	Research tasks needing live information

4.3

Long Context (1M+)

Model	Window	Recall Performance
Gemini 2.0 Flash	1M	99%+ Needle-in-a-Haystack (verified)
Gemini 2.0 Pro	1M	Best quality at 1M context
Claude 3.7 Sonnet	200K	Reliable, high-quality long context

05models

Embedding Models

5.1

API Embedding Models (March 2026)

Model	Dimensions	Max Tokens	MTEB Score	Cost/1M
OpenAI text-embedding-3-large	3072	8191	64.6	$0.13
OpenAI text-embedding-3-small	1536	8191	62.3	$0.02
Voyage-3	1024	32000	67.8	$0.06
Cohere embed-v3	1024	512	66.4	$0.10
Google text-embedding-004	768	2048	66.1	$0.025

5.2

Open Source Embedding Models

Model	Dimensions	Max Tokens	MTEB	Notes
BGE-large-en-v1.5	1024	512	63.9	Instruction-tuned
E5-mistral-7b-instruct	4096	32768	66.6	Strong with instructions
Nomic-embed-text-v1.5	768	8192	62.3	Long context, open
GTE-Qwen2-7B	3584	32K	72.1	State-of-the-art open embedding

5.3

Embedding Selection Guide

Requirement	Recommended	Why
Best quality	Voyage-3 or text-embedding-3-large	Highest MTEB
Cost-efficient	text-embedding-3-small	$0.02/1M
Self-hosted	GTE-Qwen2-7B	Best open MTEB
Long documents	Nomic or Voyage-3	8K+ context
Multilingual	Cohere embed-v3	Built for multilingual

06selection framework

Model Selection Framework

6.1

Decision Tree

Figure 1

Model choice is settled by naming the binding constraint first, not by comparing benchmarks

One column per constraint; the teal line in each cell is the threshold that picks the row.

Text version of this diagram

Texttext · 21 lines

123456789101112131415161718192021

What is your primary constraint?

├── Cost → Use smaller model, consider open source
│   ├── Very cost sensitive → o3-mini, Claude 3.5 Haiku, Gemini 2.0 Flash
│   └── Moderate budget → Claude 3.7 Sonnet, o3-mini
│
├── Quality + Reasoning → Use frontier models
│   ├── Highest reasoning → o3 (high effort)
│   └── Coding + reasoning → Claude 3.7 Sonnet (Extended Thinking)
│
├── Latency → Use fast models
│   ├── <100ms response → Gemini 2.0 Flash, GPT-4o-mini
│   └── <500ms response → Claude 3.5 Haiku, o3-mini (low)
│
├── Self-hosting → Use open models
│   ├── Maximum capability → Llama 3.1 405B, DeepSeek-V3
│   ├── Good balance → Llama 3.3 70B, Qwen2.5-72B
│   └── Edge/mobile → Llama 3.2 3B, Phi-4
│
└── Privacy → Self-host or use on-prem
    └── Choose open models with appropriate license

6.2

Semantic Routing

In 2025-26, static decision trees are being replaced by Semantic Routers:

How it works: A small, fast embedding model vectorises the query. If it matches a "known easy" cluster → cheap model (e.g., Gemini 2.0 Flash). If it hits an "agentic/logic" cluster → o3 or Claude 3.7.
Benefit: Automates cost-optimization without hardcoded rules.
Implementation: Tools like semantic-router (Python) or custom Weaviate/Pinecone classifiers.

07data residency

Sovereign AI and Data Residency

The 2026 Regulatory Reality: Enterprises must comply with GDPR (EU), DPDPA (India), Saudi Arabia PDPL, and sectoral rules. "Sovereign AI" is now a product category.

Solution	Provider	Use Case
Azure Government/Sovereign	Microsoft	Dedicated infra in 40+ regions; approved for US Gov/EU NIS2
AWS Sovereign Cloud	Amazon	Physically isolated VPCs; GDPR-safe EU regions
Google Distributed Cloud	Google	Air-gapped on-prem Gemini deployment
Private Llama 3.3	Meta (self-host)	Maximum data sovereignty; open weights
DeepSeek (self-host)	DeepSeek (open)	Open weights; no data leaves your infra

Tradeoff: Sovereign clouds carry a 20-30% premium over standard global regions but are mandatory for finance and government.

7.1

Cost Comparison at Scale (March 2026)

Assume 1M requests/day, 1K input + 500 output tokens:

Model	Input Cost/Day	Output Cost/Day	Total/Month
GPT-4o	$2,500	$5,000	$225,000
Claude 3.7 Sonnet	$3,000	$7,500	$315,000
o3-mini	$1,100	$2,200	$99,000
Gemini 2.0 Flash	$100	$200	$9,000
Self-hosted Llama 3.3 70B*	—	—	~$50,000

*Self-hosted assumes 4× H100 GPUs

08comparison

Capability Comparison

8.1

Benchmark Performance (March 2026, Verified)

Model	MMLU	HumanEval	SWE-bench Verified	AIME 2025
o3 (high)	91.6	96.7	71.7%	96.7%
Claude 3.7 Sonnet	90.5	93.6	70.3%	80.0%
GPT-4.5	89.4	86.8	38.0%	36.7%
Grok 3	90.2	88.9	—	93.3%
DeepSeek-R1	90.8	92.6	49.2%	79.8%
Llama 3.1 405B	88.6	89.0	—	—
Gemini 2.0 Flash	85.5	83.0	—	—

Source: Respective technical reports and LMSYS Chatbot Arena, March 2026. Always verify with current leaderboards.

8.2

Task-Specific Recommendations (March 2026)

Task	Recommended Models	Why
Autonomous Coding Agent	Claude 3.7 Sonnet	Powers Claude Code; best SWE-bench + tool reliability
Complex Reasoning	o3, DeepSeek-R1, Claude 3.7 (thinking)	Extended Thinking / internal CoT
High-Volume API	Gemini 2.0 Flash, o3-mini	Lowest cost per token in class
Long Context RAG	Gemini 2.0 Flash (1M), Claude 3.7 (200K)	Verified long-range recall
Multimodal Real-time	Gemini 2.0 Flash Live API	Real-time audio/video/text native
Private Production	Llama 3.3 70B, Qwen2.5-72B	High capability with local control
Open-source Coding	Qwen2.5-Coder-32B, DeepSeek-R1-Distill	Self-hosted, top HumanEval
Creative/Chat	GPT-4.5	Top EQ scores; best conversation quality

09questions

Interview Questions

Q: How would you select a model for a production RAG system?

Strong answer: I evaluate across these dimensions:

1. Quality requirements:

Test on representative queries from the actual domain
Measure answer correctness, hallucination rate, citation accuracy

2. Cost analysis:

Texttext · 1 line

Monthly cost = requests/day × 30 × avg_tokens × rate

Always calculate for top 2-3 candidates.

3. Latency requirements:

If <200ms TTFT needed: Gemini 2.0 Flash, Claude 3.5 Haiku, o3-mini (low)
If quality is paramount: Accept 2-3s with Claude 3.7 or o3

4. Operational requirements:

Self-hosting: Llama 3.3, DeepSeek-V3
Compliance / data residency: Azure Sovereign or self-hosted

5. Practical selection:

Start with Claude 3.7 Sonnet or GPT-4o for prototyping
A/B test Gemini 2.0 Flash for 80% of queries (cost)
Keep frontier on hard queries via semantic routing

Q: Explain the tradeoffs between proprietary and open source models.

Strong answer:

Factor	Proprietary (OpenAI, Anthropic)	Open Source (Llama, DeepSeek)
Quality	Generally higher (slightly)	Catching up rapidly
Cost	Per-token pricing	Compute + ops
Control	Limited	Full
Privacy	Data goes to provider	Stays on-prem
Updates	Automatic	Manual
Customization	Limited fine-tuning	Full fine-tuning
Ops overhead	None	Significant

Key insight (2026): DeepSeek-V3 and R1 changed this conversation — open models now match GPT-4o on benchmarks. The gap is real but narrower than ever.

Q: What is the difference between o3 and Claude 3.7's Extended Thinking?

Strong answer: Both use internal chain-of-thought, but the mechanics differ:

o3: OpenAI's compute-scaling approach. Allocates variable compute effort (low/medium/high). Internal thoughts are never exposed. Single API parameter: reasoning_effort.
Claude 3.7 Extended Thinking: Returns thinking tokens in a separate <thinking> block. Configurable budget_tokens (1024–100K). You can inspect the reasoning chain for debugging, though it's not shown to end users.

Production choice: For debugging and trust-building, Claude's visible thinking is more transparent. For simple high-accuracy tasks, o3-mini on medium effort is most cost-effective.

10references

References

Anthropic: https://anthropic.com/claude/claude-3-7-sonnet anthropic.com
OpenAI Platform: https://platform.openai.com/docs/models platform.openai.com
Google AI: https://ai.google.dev/ ai.google.dev
Meta Llama: https://llama.meta.com/ llama.meta.com
DeepSeek: https://api-docs.deepseek.com/ api-docs.deepseek.com
LMSYS Chatbot Arena: https://chat.lmsys.org/ chat.lmsys.org
Hugging Face Open LLM Leaderboard: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard huggingface.co

Next: Capability Assessment

• summary · added by this rebuild

Key takeaways

Five tiers, not one leaderboard

The page sorts models into Frontier, Fast/Efficient, Battle-Tested, Small/Edge and Reasoning-Heavy; a model can top a benchmark and still be the wrong tier for your latency or privacy constraint.

The cost spread across tiers is 25x

At 1M requests a day with 1K input and 500 output tokens, the page puts Gemini 2.0 Flash at $9,000 a month against $225,000 for GPT-4o and $315,000 for Claude 3.7 Sonnet.

Semantic routing replaces the decision tree

A small embedding model clusters the incoming query and sends easy ones to a cheap model and agentic ones to a frontier model, so cost control stops being hardcoded rules.

Extended thinking differs by vendor

o3 exposes only a reasoning_effort dial and hides its thoughts; Claude 3.7 returns a separate thinking block with a budget_tokens setting from 1024 to 100K that you can inspect while debugging.

Sovereign deployment carries a premium

Azure Sovereign, AWS Sovereign Cloud and Google Distributed Cloud run 20-30% above standard global regions, which the page treats as mandatory rather than optional for finance and government.