How much does AI consulting cost?

BlokAI's engagements range from $5,000 to $75,000 depending on scope. Simple automations and LLM integrations start around $5K–$15K. Custom AI agent systems and full data infrastructure projects run $30K–$75K. We scope every project before committing to a price.

What does BlokAI build?

BlokAI builds AI agents, LLM applications, RAG systems, automation pipelines, and data infrastructure for startups and SMBs. We take projects from architecture to production — not demos, not prototypes.

Is BlokAI based in Japan?

Yes. BlokAI is a solo studio run by Kent Tran, based in Tokyo, Japan. We work with clients globally in English.

What is RAG and why does my business need it?

RAG (Retrieval-Augmented Generation) lets AI systems search your own documents, databases, or knowledge bases before answering — instead of relying only on training data. It powers AI customer support, internal search tools, and sales enablement systems that are accurate and up-to-date.

Claude 3 and GPT-4 Turbo: Choosing the Right LLM for Your Business in 2024

"Claude 3 Opus not only delivers state-of-the-art performance on cognitive tasks, it is the first model to surpass human expert performance on GPQA, a challenging test of expert knowledge." — Anthropic Blog, March 4, 2024

March 2024 is a genuine inflection point in the large language model market. Anthropic released Claude 3 this month — three models across different capability/cost tiers (Haiku, Sonnet, Opus) — with benchmark scores that beat GPT-4 on multiple dimensions. Meanwhile, GPT-4 Turbo has matured significantly since its November launch. And there's a capable open-source ecosystem emerging for businesses that need to keep data in-house.

This is a good problem to have. The options are better than they've ever been. But "better options" means more decisions, and I've been getting a lot of questions from clients about how to navigate this.

Here's my practical take.

The Current Landscape (March 2024)

OpenAI:

GPT-4 Turbo (128k context, latest checkpoint) — their most capable model
GPT-3.5 Turbo — fast, cheap, capable for simpler tasks
Embeddings, vision, function calling all available

Anthropic:

Claude 3 Opus — matches or beats GPT-4 Turbo on most benchmarks, 200k context window
Claude 3 Sonnet — strong middle tier, good quality/cost balance
Claude 3 Haiku — fast, cheap, excellent for high-volume simpler tasks

Open Source:

Mistral 7B / Mixtral 8x7B — remarkable capability for the model size, commercially licensable
Llama 2 / Code Llama — strong foundation models, active fine-tuning ecosystem

Others:

Google Gemini (in preview) — entering the market with competitive specs
Cohere — strong for specific enterprise use cases, especially search

The Real-World Selection Criteria

Benchmarks are useful but not the whole story. Here's how I actually help clients choose:

1. What's the Primary Task?

Complex reasoning and analysis: Claude 3 Opus or GPT-4 Turbo. Both are excellent. Test on your actual task. The quality difference at this tier is real but often smaller than benchmarks suggest in practice. High-volume, cost-sensitive inference: GPT-3.5 Turbo or Claude 3 Haiku. At $0.25-0.50/million tokens, these are dramatically cheaper than the frontier models for tasks they handle well — classification, simple extraction, first-pass summarisation. Long document processing: Claude 3 (200k token context) edges out GPT-4 Turbo (128k) for very long documents. If you're processing whole books, lengthy legal contracts, or multi-document sets, the context advantage matters. Code generation: GPT-4 Turbo tends to perform slightly better on complex coding tasks in my experience. Useful for AI-assisted development tools. Instruction following and format adherence: Both frontier models are excellent. Claude 3 Opus has been notably strong in our testing on precise instruction following.

2. What Are Your Data Privacy Requirements?

Both OpenAI and Anthropic have enterprise agreements and data processing addendums (DPAs) that address data privacy. For most businesses, using either API with appropriate enterprise agreements is acceptable.

For businesses that cannot send data off-premise at all — certain healthcare, legal, or regulated finance contexts — open-source models (Mixtral, Llama 2) on your own infrastructure remain the answer.

3. What's Your Cost Model?

Quick comparison at current pricing (March 2024):

| Model | Input (per 1M tokens) | Output (per 1M tokens) |

|-------|----------------------|------------------------|

| GPT-4 Turbo | $10 | $30 |

| Claude 3 Opus | $15 | $75 |

| Claude 3 Sonnet | $3 | $15 |

| GPT-3.5 Turbo | $0.50 | $1.50 |

| Claude 3 Haiku | $0.25 | $1.25 |

For a system doing 100,000 queries/day with moderate context, the cost difference between GPT-4 Turbo and Claude 3 Haiku is substantial. Choose the most capable model that actually meets your quality bar — not always the most capable model available.

4. What's Your Ecosystem?

If you're deeply invested in the OpenAI ecosystem — using Assistants API, fine-tuning, their function calling patterns — switching to Anthropic involves some migration work. If you're building greenfield, start with whichever model performs better on your specific task.

One practical consideration: GPT-4 Turbo is available via Azure OpenAI Service, which matters significantly for enterprises with existing Microsoft agreements or compliance requirements that are already met by the Azure infrastructure.

What We're Currently Recommending

At BlokAI, our current defaults for new projects:

Reasoning-heavy applications (analysis, complex Q&A, multi-step reasoning): Test Claude 3 Opus and GPT-4 Turbo on your actual data. Use whichever wins.
RAG and knowledge retrieval: Claude 3 Sonnet or GPT-4 Turbo. The improved instruction following in both helps significantly with staying grounded in retrieved context.
High-volume classification or extraction: Claude 3 Haiku. Remarkable quality at the price point.
Privacy-sensitive on-premise: Mixtral 8x7B. Best open-source option for complex tasks.
Document embedding: OpenAI's text-embedding-3-large or text-embedding-3-small (just released, significant improvement over ada-002 at similar cost).

The Honest Answer

There is no universally right model. The answer is always "it depends on your task, your volume, your data requirements, and your quality bar — and you should test on your actual use case before committing."

Anyone who tells you there's one clearly correct LLM for all business applications in 2024 either hasn't done enough real-world building or has a financial incentive in the answer.

Test. Measure. Choose based on evidence, not benchmarks or hype.

Want to implement AI in your business? Book a 30-minute architecture call at blok-ai.com.