Pricing Cache

Scope: AI/Bedrock only. This file prices Bedrock model inference and the source-provider baselines used for migration ROI. It does NOT price infrastructure — azure-to-aws prices EC2/RDS/S3/ElastiCache/etc. from references/vendored/pricing/aws-infra-pricing.json (the file the pricing-coverage gate enforces). Two infra pricing sources would be a drift trap, so this cache deliberately carries no infra rows. estimate-ai.md reads this file for AI cost; estimate-infra.md reads the infra JSON.

Prices vary by region and change over time; use for estimation only. There is no live pricing lookup — on a cache miss apply pricing-fallback.md (estimated or unavailable). Staleness warning: if today is more than 30 days after Last updated, treat AI model prices as potentially stale; set pricing_source: "cached_stale" in estimation-ai.json and note it. Verify against aws.amazon.com/bedrock/pricing (opens in a new tab).

Bedrock Models (On-Demand)

Anthropic Claude (Standard on-demand) figures match US East (N. Virginia) on Amazon Bedrock pricing (opens in a new tab) as of cache refresh. Recommend defaults (new migrations): Claude Sonnet 5 (flagship), Claude Opus 4.8 (hardest reasoning), Claude Haiku 4.5 (cost/speed). Do not default to Claude Fable 5 (frontier). Long-context SKUs do not all use the same multiplier; confirm batch/cache and cross-region rows per model on that page. See references/vendored/ai/ai-model-lifecycle.md for lifecycle detail — do not recommend Legacy/excluded models for new migrations.

Multi-provider quick reference (per 1M tokens)

Quick-reference rows use — for model ID and context; resolve in the Bedrock console or AWS model documentation. This is a curated subset for migration selection — see the Bedrock pricing page for the full catalog.

Embeddings — Bedrock (per 1M input tokens, US East)

Embedding models are input-only — priced per 1M input tokens, no output charge. Use for embedding capability workloads (RAG corpora, semantic search, FAQ retrieval). A migration from an OpenAI/Azure embedding deployment lands on one of these; note the dimension must match (or the corpus must be re-embedded and any similarity threshold recalibrated).

Default target for a migrating OpenAI/Azure embedding workload: Titan Text Embeddings v2 (amazon.titan-embed-text-v2:0) — cheapest, configurable dimensions (1024 default; 512/256 for cost/latency). Choose Cohere Embed v4 when the source used a large-dimension model and matrix compatibility or multilingual quality matters. Titan v2 output dimension is configurable, so map text-embedding-3-large (3072-dim) or -small (1536-dim) to Titan 1024 and re-embed — there is no dimension-preserving swap. See references/vendored/ai/ai-openai-to-bedrock.md.

Stability AI — Image Generation (per image, US East)

Priced per image (not per token). Use for image_generation capability workloads.

Cost comparison note: DALL-E 3 (OpenAI / Azure OpenAI) charges $0.04–$0.12/image by resolution. When comparing, use per-image cost directly — do not convert Stability AI prices to per-token equivalents.

Anthropic Claude — batch & prompt cache (Standard, US East N. Virginia)

Per 1M tokens unless noted.

OpenAI on Bedrock — the same-model path

The proprietary GPT-5.x models on Bedrock are priced at OpenAI’s data-residency tier ≈ 1.10x the OpenAI standard rates below, and run bedrock-mantle only, in-region only (no bedrock-runtime, no cross-region inference profile for GPT-5.5/5.4). gpt-oss DOES support bedrock-runtime. For endpoint paths, region matrix, quotas, and prompt-caching detail see references/shared/openai-on-bedrock.md (the fact base the mapping guide reads).

Source Provider Pricing (for Migration Comparison)

Use alongside the Bedrock rates to compute migration ROI. This is the source-side baseline — what the customer pays today on their current provider. Per estimate-ai.md, the authoritative “today” figure is MEASURED/STATED spend (from discovery or clarify); these list rates are the fallback when no measured spend exists, and the source side of the per-model “vs source” column.

Azure OpenAI uses the OpenAI rows below. Azure OpenAI serves the same GPT models at list prices that track OpenAI’s, so ai_source: azure_openai reads the OpenAI table here — there is no separate Azure-OpenAI table (plan §19.9a). Azure OpenAI’s enterprise/PTU discounts vary per contract; when the customer’s actual spend is known, that overrides these list rates.

OpenAI / Azure OpenAI (Standard Tier)

Prices per 1M tokens.

Tier note — these are OpenAI STANDARD-tier rates. Bedrock in-region for the same models is the OpenAI data-residency tier, exactly 1.10x these figures (see the OpenAI-on-Bedrock note above). So a same-model move for GPT-5.6 Sol/Terra/Luna, GPT-5.5, GPT-5.4 is a ~10% increase, not parity — compute the target from the Bedrock table, not by carrying these over. These rows are the right source-side baseline, and the only figures available for models with no Bedrock equivalent (GPT-5.x Pro, GPT-5.2/5.1, GPT-4.x, o-series).

Azure OpenAI note. Azure lists the same models under Azure-specific deployment names (e.g. gpt-4o, gpt-4.1) and bills per 1M tokens at rates that track the table above. Azure adds Provisioned Throughput Units (PTU) as an alternative to pay-as-you-go; a customer on PTU has a committed monthly cost that their stated spend captures directly — prefer stated spend over these list rates when available.

Embeddings — OpenAI / Azure OpenAI source (per 1M input tokens)

The source-side baseline for a migrating embedding workload. Input-only. Map the “$X today” from measured/stated spend when available; these list rates are the fallback.

Azure OpenAI bills the same models under deployment names at rates that track this table (PTU caveat above applies). A text-embedding-3-large → Titan v2 move is not a dimension-preserving swap (3072 → 1024) — it requires re-embedding the corpus and recalibrating similarity thresholds; surface that as a migration task, not a silent cost line.

Gemini source rows from gcp-to-aws are intentionally omitted: azure-to-aws has no gemini ai_source (plan §19.9b).