Platform availability gaps
Where you deploy changes what you can build, and it changes it in ways that are invisible until somebody tries. This is dry material, so here is the whole thing in three groups and then we are done.
Group one: first-party API only
Four capabilities exist on Anthropic’s own API and on no third-party platform at all.
Batches. The 50 percent discount. Gone.
Data residency. inference_geo and its 1.1 times multiplier. Gone.
Server-side fallback. Automatic failover to another model, handled for you. Gone, which means failover becomes code you write and test.
Advisor. Gone.
Worth separating those two examples, though, because they are not the same severity. Data residency is a requirement that cannot be met on that platform. Batches is a requirement that can be met at twice the price. One is a wall and the other is an invoice, and calling both of them blockers in a design review costs you credibility on the one that really is.
Group two: missing on Bedrock and Vertex, present on Foundry
Six things, and the pattern is worth remembering as a set:
Files API, Agent Skills, the MCP connector, code execution, web fetch, and programmatic tool calling.
Bedrock and Vertex do not have them. Foundry does. So among the three third-party platforms, Foundry is materially more capable, and a design that needs any of these has effectively picked its platform already.
Group three: everywhere except Bedrock
Two, and both are expensive to lose:
Web search and automatic prompt caching.
Losing automatic prompt caching on Bedrock deletes the 0.1 times read multiplier from your cost model. For a workload built around a big cached prefix, that is not a feature footnote, it is the difference between the design working economically and not.
How to actually use this
Do not memorise it as a matrix. Memorise it as three questions you ask in order:
Does the design need Batches, data residency, server-side fallback or advisor? Then it is first-party, decision closed.
Does it need Files, Skills, MCP connector, code execution, web fetch or programmatic tool calling? Then Bedrock and Vertex are out, and Foundry is the third-party option.
Is it going on Bedrock? Then check whether web search or caching were load-bearing, and rewrite the cost model if caching was.
Three questions, and they resolve almost every platform question you will be asked. The exam version of this usually presents a requirement plus a platform and asks what breaks, which is the same lookup running backwards.
Try it yourself
The four that are first-party only
Four features exist nowhere except the first-party API. Name them and say why each one hurts to lose.
Which four capabilities are not available on any third-party platform, and what does losing each one cost a design?
Reveal answer
Batches, data residency, server-side fallback, and advisor. Losing Batches costs you the 50 percent discount on all offline work. Losing data residency removes inference_geo and its 1.1 times multiplier, which may be a compliance blocker rather than a cost one. Losing server-side fallback means model failover becomes your code. Losing advisor removes that tool entirely. If any of these is load-bearing, the platform decision is already made.
Checking one feature against one platform
A design needs the Files API and the team has been told to deploy on Vertex. Before you argue with anyone, you want the fact. Where do you look?
Show answer
Correct answer: A — The platform comparison table listing feature availability across the first-party API, Bedrock, Vertex and Foundry
You need a per-platform answer, so you need the surface that is organised by platform, which is the comparison that puts the third-party options side by side and shows exactly where each feature drops out. The Files API page is the tempting answer because it is the page about the thing you care about, and a feature page will tell you what the feature does and which header it needs without necessarily enumerating which clouds carry it. The last option is the reasoning error the question is aimed at, since a beta header being defined says nothing about whether a given platform implements the feature behind it.
What Bedrock does not have
Your team must deploy on Bedrock. Which of these do you lose?
Show answer
Correct answer: D — Automatic prompt caching and web search
Web search and automatic prompt caching are available everywhere except Bedrock, which makes Bedrock the one platform where those two specifically drop out. Streaming, tool use, structured outputs, citations, vision and stop_sequences are all core Messages API surface and are not what gets carved out per platform. The reason this matters is that losing automatic prompt caching removes the 0.1 times read multiplier from your cost model entirely.
The Foundry exception
Files API, Agent Skills, the MCP connector, code execution, web fetch and programmatic tool calling share a platform pattern. What is it?
Show answer
Correct answer: B — They are unavailable on Bedrock and Vertex, but available on Foundry
That group is missing on Bedrock and Vertex while Foundry does carry them, which makes Foundry meaningfully more capable than the other two third-party options for feature-heavy designs. The first-party-only option is the tempting one because the list looks like the first-party-only set, but the genuinely first-party-only four are Batches, data residency, server-side fallback and advisor.
The requirement that ends the discussion
A design has three hard requirements, and a stakeholder wants it on Bedrock for procurement reasons. Which requirement, on its own, makes that impossible rather than merely awkward?
Show answer
Correct answer: C — All inference must happen on infrastructure located in the United States
Data residency is one of the four first-party-only capabilities, so a hard requirement for it closes the platform question outright with no workaround available. The Batches requirement is the genuinely tempting answer and it is a near miss on purpose, because Batches is also first-party only, but a budget assumption can be renegotiated by paying full price, which makes it expensive rather than impossible. Streaming and tool search are both available and are not constraints here at all. The habit worth taking away is separating a requirement that cannot be met from one that can be met at a cost.
Run a platform feasibility check
Take a feature list from any design and test it against a named platform.
Every feature in the design is marked available or unavailable on the chosen platform, using the three groupings (first-party only, missing on Bedrock and Vertex, missing on Bedrock only), and any unavailable feature has either a workaround or a platform change written next to it.