> For the complete documentation index, see [llms.txt](https://docs.blockbrain.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.blockbrain.ai/for-users/all-about-llms/how-to-choose-the-right-llm.md). # How to Choose the Right LLM A detailed LLM selection guide explaining how Blockbrain measures usage via Compute Blocks, with concrete model recommendations for company-wide deployment and specific use cases. When using Blockbrain, understanding how computational resources are measured and allocated is essential for getting the most out of the platform. At the heart of this system lies a transparent and usage-based metric: **Compute Blocks (CBs)**. Every action at Blockbrain - sending a message, uploading a file, or running an agent - consumes Compute Blocks (CBs). CBs reflect the actual computational cost of each operation and are composed primarily of tokens used by Large Language Models (LLMs). CB usage directly mirrors the input and output token pricing of each LLM. For example, if Opus 4.8 is 66.67% more expensive than Sonnet 4.6 in terms of input and output token prices, its CB consumption will also be approximately 66.67% higher. *** ### Default Recommendation for Company-Wide Use #### Primary Choice/s: Claude Sonnet 5 (AWS Bedrock) Blockbrain Metrics: * Quality: 4.7 | Speed: 4.4 | Cost Efficiency: 3.6 * Context: 1M tokens | Provider: AWS Bedrock (EU) **Pricing**: $1.87 input / $9.35 output per million tokens Why consider? Highest quality among budget-tier models. Excellent for teams that need better reasoning while maintaining cost efficiency. #### Primary Choice/s: Gemini 3.5 Flash (Google) Blockbrain Metrics: * Quality: 4.2 | Speed: 4.8 | Cost Efficiency: 3.7 * Context: 1M tokens | Provider: Google AI (EU) **Pricing**: $1.50 input / $9.00 output per million tokens Why consider? Competitive quality at low cost - excellent for high-volume deployment. #### Primary Choice/s: GPT 5.6 Terra (Azure) Blockbrain Metrics: * Quality: 4.4 | Speed: 3.4 | Cost Efficiency: 3.4 * Context: 1M tokens | Provider: Azure AI (EU) **Pricing**: $2.50 input / $15.00 output per million tokens Why this model? Excellent balance of quality, speed - ideal for diverse business use cases. *** ### Scenario-Based Recommendations #### Highest Quality (Premium Tasks) **Winner: Claude Opus 5** * Quality **4.9** | Speed 4.0 (Bedrock) | Cost Efficiency 3.0 * Context: 1M tokens | Provider: AWS Bedrock (EU) * Pricing: $4.67 input / $23.38 output per million tokens **When to use**: Reserve for mission-critical tasks, C-suite deliverables, complex strategic analysis, long-horizon agent runs, complex strategic analysis, anything where the model needs to catch its own mistakes. **Runner-up: GPT 5.6 Sol** * Quality **4.8** | Speed 3.5 | Cost Efficiency 2.9 * Context: 1M tokens | Provider: Azure AI (EU) * Pricing: $5.00 input / $30.00 output per million tokens **Budget Quality Option: Claude Sonnet 5** * Quality: **4.7** | Speed: 4.4 | Cost Efficiency: 3.6 * Context: 1M tokens | Provider: AWS Bedrock (EU) * Pricing: $1.87 input / $9.35 output per million tokens * Best premium option without extreme cost #### Maximum Speed **Winner: Gemini 3.5 Flash** * Speed **4.8** | Answer Quality **4.2** | Cost Efficiency 3.7 * Context: 1M tokens | Provider: Google AI (EU) * **Pricing**: $1.50 input / $9.00 output per million tokens **Why it wins**: Beats the old winner (Claude Sonnet 4.6 Fast) by 0.4 quality points at half the output cost. **Budget Quality Option: Gemini 3.1 Flash (Lite)** * Speed **4.9** | Quality 3.4 | Cost Efficiency **4.9** * Context: 1M tokens | Provider: Google AI (EU) * Pricing: $0.25 input / $1.50 output * **Better value** for most speed-critical applications #### Code Development Excellence **Winner: Claude Sonnet 5** Blockbrain Metrics: * Quality **4.7** | Speed **4.4** | Cost Efficiency 3.6 * Context: 1M tokens | Provider: AWS Bedrock (EU) * Pricing: $1.87 input / $9.35 output per million tokens **Why it wins**: Comes close to Opus 4.8 quality at a fifth of the output cost **Alternative: Gemini 3.5 Flash (High thinking)** Blockbrain Metrics: * Quality: **4.4** | Speed: 4.2 | Cost Efficiency: 3.5 * Context: 1M tokens | Provider: Google AI (EU) * **Pricing**: $1.50 input / $9.00 output per million tokens **Why it wins**: High quality for code development (4.4) with excellent speed. Purpose-built for agentic and coding work at Flash speed and Flash pricing. **Premium Option: Claude Opus 5** * Quality **4.9** | Speed 4.0 (Bedrock) | Cost Efficiency 3.0 * Context: 1M tokens | Provider: AWS Bedrock (EU) * Pricing: $4.67 input / $23.38 output per million tokens * Best for: Complex architectural decisions, critical code review #### Creative & Writing Tasks **Winner: Claude Sonnet 5** Blockbrain Metrics: * Quality **4.7** | Speed **4.4** | Cost Efficiency 3.6 * Context: 1M tokens | Provider: AWS Bedrock (EU) * Pricing: $1.87 input / $9.35 output per million tokens **Why it wins**: The latest version of Sonnet 4.6 which delivered flagship-quality writing (4.7) at mid-tier pricing—exceptional value for creative work. **Budget Alternative: Claude Haiku 4.5** * Quality **3.6** | Speed 3**.6** | Cost Efficiency 4 * Context: 200kk tokens | Provider: AWS Bedrock (EU) * Pricing: $0.94 input / $4.67 output per million tokens * Best for: Still the best cheap model *for tone* in the catalog **High Volume Alternative: GPT 5.6 Luna** * Quality **3.7** | Speed 4 | Cost Efficiency **5** * Context: 1M tokens | Provider: OpenAI (EU) * Pricing: $0.20 input / $1.20 output per million tokens * Best for: Use where volume beats polish #### Complex Reasoning Tasks **Winner: Claude Opus 5** * Quality **4.9** | Speed 4.0 (Bedrock) | Cost Efficiency 3.0 * Context: 1M tokens | Provider: AWS Bedrock (EU) * Pricing: $4.67 input / $23.38 output per million tokens * Performance: 43.3% on Frontier-Bench v0.1 (vs Opus 4.8's 21.1%); 1861 on GDPval-AA v2 for knowledge work * Best for: multi-step analysis, scientific reasoning, mathematical work, self-verifying agent loops **Runner-up: GPT 5.6 Sol** * Quality **4.8** | Speed 3.5 | Cost Efficiency 2.9 * Context: 1M tokens | Provider: Azure AI (EU) * Pricing: $5.00 input / $30.00 output per million tokens * Performance: 91.9% on Terminal-Bench 2.1. Strong on long-running autonomous tasks; uses fewer tokens per task than GPT 5.5 despite the deeper reasoning **Budget Alternative: Claude Sonnet 5** * Quality **4.7** | Speed **4.4** | Cost Efficiency 3.6 * Context: 1M tokens | Provider: AWS Bedrock (EU) * Pricing: $1.87 input / $9.35 output per million tokens *** ### Decision Matrix
PriorityPrimary RecommendationBudget AlternativePremium Option
Balanced everyday use Claude Sonnet 5 ($1.87/$9.35)Gemini 3.5 Flash ($1.50/$9)Claude Opus 5 ($4.67/$23.38)
Maximum cost savings GPT 5.6 Luna ($0.20/$1.20)Gemini 3.1 Flash Lite ($0.25/$1.50)Gemini 3.5 Flash ($1.50/$9)
Highest quality Claude Opus 5 ($4.67/$23.38)Claude Sonnet 5 ($1.87/$9.35)
Fastest response Gemini 3.5 Flash Low ($1.50/$9)Gemini 3.1 Flash Lite ($0.25/$1.50)GPT 5.4 Low Thinking ($2.50/$15)
Creative work Claude Sonnet 5 ($1.87/$9.35)Claude Haiku 4.5 ($.94/$4.67) Claude Opus 5 ($4.67/$23.38)
Code development Claude Sonnet 5 ($1.87/$9.35)Gemini 3.5 Flash ($1.50/$9)Claude Opus 5 ($4.67/$23.38)
Complex reasoning Claude Opus 5 ($4.67/$23.38)Claude Sonnet 5 ($1.87/$9.35)GPT 5.6 Sol ($5/$30)
*** ### Strategic Recommendations #### For Most Companies: Multi-Model Strategy We recommend a tiered approach: * **Tier 1** (80% of queries): Fast, cost-efficient models * Gemini 3.5 Flash (Low) or GPT 5.6 Luna * Use for: emails, summaries, Q\&A, basic analysis * **Tier 2** (15% of queries): Balanced premium models * Claude Sonnet 5 or Gemini 3.5 Flash * Use for: reports, complex content, strategic analysis * **Tier 3** (5% of queries): Flagship models * Claude Opus 5 * Use for: critical decisions, high-stakes content, C-suite materials Estimated Savings: 60–75% vs. using flagship models for everything *** ### Important Considerations #### Output Token Costs Matter Most For typical conversational AI: * **Input**: System prompt + user query = 500 tokens * **Output**: AI response = 200–500 tokens Example cost for 1,000 queries (500 input tokens, 300 output tokens): | Model | Input Cost | Output Cost | Total | | ----------------- | ---------- | ----------- | ------ | | GPT 4o Mini | $0.075 | $0.18 | $0.26 | | GPT 5.4 Mini | $0.20 | $0.48 | $0.68 | | Gemini 2.5 Flash | $0.25 | $0.90 | $1.15 | | Claude Haiku 4.5 | $0.50 | $1.50 | $2.00 | | Claude Sonnet 4.6 | $1.50 | $4.50 | $6.00 | | Claude Opus 4.8 | $2.50 | $7.50 | $10.00 | > Output-heavy use cases (reports, documentation, code generation) should prioritize low output-cost models. #### Context Window Value | Model | Context Window | | ---------------------- | ---------------- | | Gemini 2.5 Flash / Pro | 1M tokens | | Claude Sonnet 4.6 | 1M tokens | | Most others | 128k–400k tokens | | Mistral Codestral | 32k tokens | **When it matters**: Document analysis, long conversations, comprehensive research, multi-file code review. > Pro tip: A 1M context window can hold 750,000 words or 3,000 pages of text. #### Provider Considerations **All Blockbrain models are EU-hosted, ensuring:** * **GDPR compliance** – Data processed within EU boundaries * **Data residency** – Meets European regulatory requirements * **Lower latency** – For European customers *** ### Best Practices for Cost Optimization #### 1. Prompt Engineering **Reduce output tokens by 30–50%** Either add this in the initial instructions of the bot, or prompt it directly: * Request concise responses: "Answer in 2–3 sentences" or use in the sendbox \ Options → Length: Short / Very Short * Use structured outputs: "Respond in bullet points" * Avoid redundancy: "Don't repeat the question" > **Impact**: Can reduce costs by 40%+ for output-heavy models. #### 2. Smart Model Routing | Query Type | Recommended Model | | ----------------------------- | ---------------------------- | | Simple (FAQ, definitions) | GPT 5.4 Mini | | Standard (analysis, drafting) | Gemini 2.5 Flash | | Complex (strategic, critical) | Claude Sonnet 4.6 / Opus 4.8 | > **Impact**: 50–70% cost reduction vs. using premium models for everything. #### 3. Caching & Reuse * Cache common prompts (Prompt Library) * Reuse context where possible (e.g. via Insights) * Implement RAG (Retrieval-Augmented Generation) via the database to reduce context size **Impact**: 20–30% reduction in input token costs. *** ### Conclusion **The Blockbrain model portfolio offers excellent options for every use case and budget.** **For most companies, we recommend:** * Start with Gemini 3.5 Flash as your default model (e.g. in your Company GPT) * Add GPT 5.6 Luna for high-volume and cost-sensitive teams (e.g. classification, extraction, routing, metadata, bulk drafting) * Introduce specialists where they earn their place: Claude Sonnet 5 for agentic coding and quality writing; Gemini 3.5 Flash High for agentic work at Flash cost. * Reserve Claude Opus 5 for critical work only — and pin its effort setting, since adaptive thinking runs by default and thinking tokens bill at the output rate. **This approach typically delivers:** * 60–75% cost savings vs. premium-only deployment * 90%+ user satisfaction * Flexibility to scale and optimize over time --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://docs.blockbrain.ai/for-users/all-about-llms/how-to-choose-the-right-llm.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.