Setting the file. One moment.
Subchapter 27.5
references/design-refs/ai-gemini-to-bedrock.mdMarkdown20 KBView on GitHub
Applies to: Vertex AI Generative AI (Gemini models) → Amazon Bedrock
This file is loaded by design-ai.md when ai-workload-profile.json has = or . It provides model mapping tables with pricing and honest competitive analysis for Gemini → Bedrock migration decisions.
summary.ai_source"gemini""both"Verify all pricing against references/shared/pricing-cache.md.
Model lifecycle: Before recommending any Bedrock model, check references/vendored/ai/ai-model-lifecycle.md. Do not recommend Legacy models as primary selections for new migrations. Legacy models are annotated below where they appear.
Recommend defaults (Sep 2026): Claude Sonnet 5 (anthropic.claude-sonnet-5) for balanced/flagship; Claude Opus 4.8 for hardest reasoning; Claude Haiku 4.5 for cost/speed. Sonnet 5 is $2/$10 — the launch rate became the standard price on Sep 1, 2026 (the scheduled increase to $3/$15 was cancelled); comparison tables below use $2/$10. Do not default to any Claude Fable / Mythos frontier model.
Gemini 3.5 Flash is now GA (May 2026) — the current flagship Flash model. Gemini 3.1 Pro is the current Pro tier. Be honest with users:
Where Bedrock still wins:
Migration case by tier:
| Model | Best For | Complexity | Speed | Context |
|---|---|---|---|---|
| Claude Sonnet 5 | Agentic tasks, tool use | High | High | 1M |
| Claude Opus 4.6 | Maximum reasoning | High | Medium | 200K |
| Claude Haiku 4.5 | Simple + fast | Medium | High | 200K |
| Llama 4 Maverick | Cost-effective + multimodal | Medium | High | 1M |
| Llama 4 Scout | Ultra-long context, cheapest | Medium | Medium | 10M |
| Nova 2 Pro | AWS flagship, multimodal | High | High | 1M |
| Nova 2 Lite | AWS mid-tier, long context | Medium | High | 1M |
| Nova Pro | AWS balanced | Medium | High | 300K |
| Nova Lite | AWS fast + cheapest | Medium | High | 300K |
| Nova Micro | AWS fastest, text-only | Low | High | 128K |
| Nova Premier | Complex reasoning — Legacy (EOL Sep 14, 2026) prefer Nova 2 Pro | High | Medium | 1M |
| DeepSeek-R1 | Chain-of-thought reasoning | High | Medium | 128K |
| Mistral Large 3 | EU/Multilingual | High | Medium | 256K |
| Gemini Model | Price (in/out per 1M) | Best Bedrock Match | Bedrock Price | Winner |
|---|---|---|---|---|
| Gemini 3.1 Pro | $2.00 / $12.00 | Claude Sonnet 5 | $2.00 / $10.00 | Bedrock 13% cheaper |
| Gemini 3.1 Pro | $2.00 / $12.00 | Claude Opus 4.8 | $5.00 / $25.00 | Gemini 54% cheaper |
| Gemini 3.1 Pro | $2.00 / $12.00 | Nova 2 Pro | $1.38 / $11.00 | Bedrock 14% cheaper |
| Gemini 3 Pro | $0.50 / $3.00 | Llama 4 Maverick | $0.24 / $0.97 | Bedrock 64% cheaper |
| Gemini 3 Pro | $0.50 / $3.00 | Llama 4 Scout | $0.17 / $0.66 | Bedrock 75% cheaper |
| Gemini 3 Pro | $0.50 / $3.00 | Nova Pro | $0.80 / $3.20 | Gemini 17% cheaper |
| Gemini 2.5 Pro | $1.25 / $10.00 | Claude Sonnet 5 | $2.00 / $10.00 | Gemini 11% cheaper |
| Gemini 2.5 Pro | $1.25 / $10.00 | Nova Pro | $0.80 / $3.20 | Bedrock 62% cheaper |
| Gemini 2.5 Pro | $1.25 / $10.00 | Nova 2 Pro | $1.38 / $11.00 | Gemini 9% cheaper |
Gemini 3.1 Pro breakpoint: $4.00/$18.00 per 1M for prompts >200k tokens. Table above uses ≤200k rates.
| Gemini Model | Price (in/out per 1M) | Best Bedrock Match | Bedrock Price | Winner |
|---|---|---|---|---|
| Gemini 3.5 Flash (GA) | $1.50 / $9.00 | Nova Lite | $0.06 / $0.24 | Bedrock 94% cheaper — strong migration case; 3.5 Flash is 5x more expensive than old 2.5 Flash |
| Gemini 3.5 Flash (GA) | $1.50 / $9.00 | Claude Sonnet 5 | $2.00 / $10.00 | Gemini 14% cheaper — but Sonnet leads on agentic reliability |
| Gemini 3.1 Flash-Lite | $0.25 / $1.50 | Nova Lite | $0.06 / $0.24 | Bedrock 76% cheaper |
| Gemini 3.1 Flash-Lite | $0.25 / $1.50 | Nova Micro | $0.035 / $0.14 | Bedrock 88% cheaper |
| Gemini 2.5 Flash | $0.30 / $2.50 | Nova Lite | $0.06 / $0.24 | Bedrock 88% cheaper |
| Gemini 2.5 Flash Thinking | $0.30 / $0.60–$3.50 (varies by thinking budget) | Claude Sonnet 5 with extended thinking | $2.00 / $10.00 | Gemini is cheaper at the listed rates: even at $3.50/M output, its 2:1 blended rate is ~$1.37/M vs ~$4.67/M for Sonnet 5 (~71% lower). Consider Sonnet 5 for capability requirements, not expected savings; profile each model’s billed thinking tokens. |
| Gemini 2.0 Flash | $0.10 / $0.40 | Nova Micro | $0.035 / $0.14 | Bedrock 65% cheaper |
| Gemini Flash 1.5 | Legacy — EOL Sep 24, 2025. Migrate to Gemini 3.5 Flash. | Nova Lite | $0.06 / $0.24 | If still in use, migrate source model first; strong Bedrock cost case once on 3.5 Flash |
| Gemini Model | Price (in/out per 1M) | Best Bedrock Match | Bedrock Price | Winner |
|---|---|---|---|---|
| Gemini 1.5 Pro | Legacy — EOL Sep 24, 2025. Migrate to Gemini 2.5 Pro or 3.x Pro. | Claude Sonnet 5 | $2.00 / $10.00 | If still in use, migrate source model first |
| text-bison / chat-bison | Legacy | Llama 4 Scout | $0.17 / $0.66 | Bedrock (better quality + cheaper) |
| text-embedding-004 | $0.025 / N/A | Titan Embeddings V2 | $0.02 / N/A | Bedrock 20% cheaper |
| Gemini Embedding 2 (multimodal — text/image/video/audio/PDF) | text $0.20/1M, image $0.45/1M ($0.00012/img), audio $6.50/1M, video $12.00/1M | Amazon Nova Multimodal Embeddings | text $0.135/1M, image $0.00006/unit (see shared/pricing-cache.md § Multimodal embeddings) | Bedrock ~32% cheaper on text, ~50% cheaper on image |
| imagen-* | Varies | Stable Image Core | $0.04/img | Nova Canvas excluded (EOL Sep 30, 2026); Ultra $0.08/img if quality-first |
Percentages are blended savings using a 2:1 input-to-output token ratio. Actual savings depend on your input/output ratio.
Gemini 3.1 Pro Preview matches or beats Opus 4.6 on most reasoning benchmarks at less than half the cost. Be transparent:
Gemini Flash → Nova Micro (<200ms, text-only, cheapest), Haiku 4.5 (<400ms, vision), or Llama 4 Scout (<300ms, cheapest capable)
Low (<1M tokens/day): Use best model for quality. Cost difference minimal at this volume.
Medium (1-10M tokens/day): Present cost comparison at volume. At 5M input + 2.5M output/day:
| Model | Monthly Cost |
|---|---|
| Gemini 3 Pro | $300 |
| Llama 4 Maverick | $109 (-64%) |
| Llama 4 Scout | $75 (-75%) |
| Nova Pro | $360 (+20%) |
| Claude Sonnet 5 | $1,050 (+250%) |
High (10-100M tokens/day): Cost optimization critical. Recommend multi-model tiered approach. Llama 4 Maverick/Scout or Nova for output-heavy workloads.
Very high (>100M tokens/day): Mandatory multi-model tiered strategy:
| Gemini Model | Monthly | Best Bedrock Match | Monthly | Difference |
|---|---|---|---|---|
| Gemini 3.1 Pro Preview ($2/$12) | $1,200 | Claude Sonnet 5 ($2/$10) | $1,050 | -13% |
| Gemini 3.1 Pro Preview ($2/$12) | $1,200 | Claude Opus 4.8 ($5/$25) | $2,625 | +54% |
| Gemini 3.1 Pro Preview ($2/$12) | $1,200 | Nova 2 Pro ($1.38/$11.00) | $1,032 | -14% |
| Gemini 3 Pro ($0.50/$3.00) | $300 | Llama 4 Maverick ($0.24/$0.97) | $109 | -64% |
| Gemini 3 Pro ($0.50/$3.00) | $300 | Llama 4 Scout ($0.17/$0.66) | $75 | -75% |
| Gemini 2.5 Pro ($1.25/$10) | $938 | Nova 2 Pro ($1.38/$11.00) | $1,032 | +9% |
| Gemini 2.5 Pro ($1.25/$10) | $938 | Nova Pro ($0.80/$3.20) | $360 | -62% |
| Gemini 2.5 Flash ($0.30/$2.50) | $233 | Nova Lite ($0.06/$0.24) | $27 | -88% |
| Gemini 2.0 Flash ($0.10/$0.40) | $45 | Nova Micro ($0.035/$0.14) | $16 | -64% |
Difference column shows blended savings at a 2:1 input/output token ratio. Positive = Bedrock costs more (Gemini cheaper), negative = Bedrock cheaper.
Cache frequently-used system prompts for 90% cost reduction on cached portions. Example: 10K token system prompt repeated 1000x → $30 without caching, $3 with caching.
Not available on other Bedrock models. This is a significant Claude advantage for applications with heavy system prompt repetition.
| Gemini Feature | Bedrock Equivalent | Notes |
|---|---|---|
| Function calling | Claude tools (excellent), Mistral (good) | Minimal changes |
| Structured output/JSON | Claude (excellent), Nova Pro (good) | Most models via prompt |
| Streaming | All major models | Same SSE pattern |
| Vision | Claude Sonnet/Haiku, Llama 4 Maverick | Multimodal parity |
| Context caching | Claude prompt caching | Different mechanics, not a drop-in: Gemini uses explicit TTL-based cachedContent objects you create/reference by name; Claude uses inline cache-control breakpoints on the request itself. Re-implement the caching call sites, don’t just swap endpoints. 90% savings on cached portions once re-implemented. |
| Audio/video input | Nova 2 Sonic (speech), Transcribe/Rekognition (preprocessing) | Nova Sonic v1 is Legacy; use Nova 2 Sonic |
| Embeddings (text-only) | Amazon Titan Embeddings ($0.02/1M, 1536 dims) | Must re-embed all docs |
| Embeddings (multimodal — text+image/video/audio) | Amazon Nova Multimodal Embeddings (see pricing table above) | Do NOT route to the text-only Titan Embeddings row above — Gemini Embedding 2 unifies text+image+video+audio into one model, so detect by call arguments (image/blob content present), not by model name alone. Must re-embed all docs into the new vector space regardless of target. |