Fable 5.1 on Bedrock vs Vertex vs Foundry: Which Wins?
The Short Answer
Claude Fable 5.1 went generally available on September 1, 2026 at $10 per million input tokens and $50 per million output tokens — the same rate on Anthropic’s API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Cache reads dropped to $0.25 per million tokens, a 75% cut versus Fable 5.
Because the token price is identical everywhere, the per-token comparison is a dead end. The decision is made by four things the price sheet does not show: routing multipliers, adjacent-service cost, compliance boundaries, and how fast each platform gets new capabilities.
| Anthropic API | Amazon Bedrock | Google Cloud | Microsoft Foundry | |
|---|---|---|---|---|
| Input / output | $10 / $50 per MTok | $10 / $50 per MTok | $10 / $50 per MTok | $10 / $50 per MTok |
| Cache reads | $0.25 / MTok | $0.25 / MTok | $0.25 / MTok | $0.25 / MTok |
| Routing surcharge | — | ~1.1x for US-only vs global profile | Region-dependent | Region-dependent |
| Gov / sovereign | — | AWS GovCloud (US) | — | — |
| Buy with committed spend | No | Yes | Yes | Yes |
| New-feature lag | None — first | Days to weeks | Days to weeks | Days to weeks |
The Model Itself
Fable 5.1 is Anthropic’s frontier tier for long-horizon agentic work: ambitious coding, sustained research, enterprise knowledge work. It carries a 1 million token input context with up to 128,000 output tokens.
The self-reported jump that got attention at launch was Terminal-Bench-Science at 52.6 versus Fable 5’s 24.7 — roughly a doubling on a long-running scientific-agent benchmark. Treat that as vendor-reported: it is a self-published figure on a benchmark the vendor selected, and independent replication is the thing to wait for. Our own guidance on this pattern is in reported vs audited AI benchmark scores — the gap between the two is routinely large.
The cache-read cut is the underrated change. Dropping cache reads from $1.00 to $0.25 per million is a 75% reduction on the single line item that dominates agentic workloads. An agent that re-reads a large system prompt, a codebase index or a document set on every turn spends most of its input budget on cache reads. For that shape of workload the effective cost fell far more than the headline $10/$50 suggests.
1. Amazon Bedrock — Best If You Have Committed AWS Spend or GovCloud Needs
Fable 5.1 is available through US Geo cross-region inference (us.) and Global cross-region inference (global.) profiles, and in AWS GovCloud (US) on both the bedrock-runtime and bedrock-mantle endpoints.
The 1.1x US-only premium is the one number to internalise. Pinning inference to US regions instead of the global profile costs roughly 10% more on both input and output. That is a data-residency purchase, not a latency or quality purchase. If your compliance position does not require US-only inference, the global profile is a straight 10% saving.
⚠️ Where Bedrock costs bite is not the model. Bedrock’s adjacent services carry standalone economics. Knowledge Bases have no separate fee but sit on a vector store — OpenSearch Serverless has a baseline cost in the region of a few hundred dollars a month even at zero query volume. Budget the platform, not the tokens.
Choose Bedrock when: you have an AWS Enterprise Discount Program commitment to burn down, you need GovCloud, or your data cannot leave a VPC.
2. Google Cloud — Best If Your Data Gravity Is BigQuery
Fable 5.1 is GA through Google’s Agent Platform (the Vertex AI lineage), positioned for autonomous knowledge work and long-running asynchronous tasks. Google publishes a retirement date no earlier than March 1, 2027, which is the sort of commitment that matters for anything you have to certify.
Published retirement dates are a genuine differentiator. If your compliance process requires a documented support window for every dependency, an explicit “not before” date is worth more than an equal token price.
Choose Google Cloud when: your data already lives in BigQuery or GCS, you are building on Google’s agent tooling, or you need a documented model lifecycle date on paper.
3. Microsoft Foundry — Best If You Are an Azure Shop
Fable 5.1 is in the Microsoft Foundry model catalogue. Microsoft’s own launch guidance is unusually pointed: evaluate it with sustained reasoning tasks — building a feature across a whole codebase, running a multi-stage research workflow — rather than side-by-side single-prompt comparisons.
That advice is correct and generalises beyond Azure. Single-prompt bake-offs systematically undersell long-horizon models. Fable 5.1’s advantage shows up in task completion over hours, not in one-shot answer quality. If your evaluation harness is a spreadsheet of 50 prompts, you will conclude the cheaper model is fine, and you will be wrong about the workload you actually run.
Choose Foundry when: your identity, governance and procurement already run through Entra and Azure, or you want Claude and Azure OpenAI models behind one catalogue and one bill.
4. Anthropic’s Own API — Best for Speed of Access
No marketplace layer, no routing multiplier, and new capabilities land here first. Features that depend on Anthropic-side infrastructure — new tool types, beta headers, agent SDK integrations — typically appear on the direct API days to weeks before the clouds.
Choose direct when: you are building a product on the model rather than embedding it in an enterprise estate, and you value being current over being procured.
The Decision, Compressed
| If this is true… | Use |
|---|---|
| You have unspent AWS commitment | Bedrock (global profile) |
| You need US-only data residency | Bedrock us. profile, accept ~1.1x |
| You are a US public-sector body | Bedrock GovCloud |
| Your data lives in BigQuery | Google Cloud |
| You need a documented retirement date | Google Cloud |
| Entra owns your identity model | Microsoft Foundry |
| You want features the week they ship | Anthropic API |
| You are optimising per-token price alone | It’s a tie — stop optimising this |
The Mistake to Avoid
Teams routinely spend a week comparing per-token rates across four platforms that charge the same rate, then deploy a caching strategy that leaves 60% of savings on the table. With cache reads at $0.25 per MTok versus $10 for uncached input — a 40x difference — prompt-cache design is worth roughly two orders of magnitude more attention than platform choice.
Get the caching right first. Then pick the platform your procurement team was going to force you into anyway.
Last verified: September 8, 2026.