Most teams reach Model Garden looking for Gemini and leave surprised by how much of the catalog is not Google’s. More than half the models you can call on Vertex AI come from other builders, and the way you run one of those models decides your bill, your data path, and who you have to sign a contract with. This part is about that side of the catalog.
Third-party models on Vertex AI run in three ways: as a managed API (serverless, per token), as a proprietary partner model you license and self-deploy into your own network, or as an open model you host yourself. Claude, Mistral, Grok, DeepSeek, and Qwen are the headline managed APIs. The managed path is the one to start on. You reach for the licensed VPC path only when a named rule, not a preference, forces the traffic to stay inside your perimeter.
Where third-party models live in Model Garden
Model Garden is the single catalog inside Vertex AI where you discover, test, and deploy models. Vertex AI is Google Cloud’s managed platform for building and serving models. Part 2 walked through the catalog as a whole. Here I am narrowing to one shelf: the models Google did not build.
Google says Model Garden holds more than 200 foundation models drawn from three sources. Its own first-party models sit on one shelf: Gemini, Imagen for images, Veo for video, and the open Gemma family. Third-party commercial partners sit on a second shelf: Anthropic’s Claude, Mistral AI, xAI’s Grok, DeepSeek, and more. Open-source projects sit on a third. This post is the second and third shelves.
The trap is thinking the shelf a model sits on tells you how to run it. It does not. What matters is the consumption mode, and a single partner like Mistral can show up in more than one mode at once. There are three modes, and you should learn them before you learn any model name.
What runs as a managed API
Model as a Service, or MaaS, means Google runs the model and you call an endpoint. Serverless. No GPU to size, no replica count to set, no scaling policy to babysit. You send a request to a Vertex AI URL and pay per token, the same billing shape you saw for Gemini in Part 3. The difference is the model on the other end was built by someone else, and your prompts and responses are not shared with that third party when you go through the Vertex AI API.
The managed catalog is deeper than most people expect. Here is the shape of it as of this run.
| Provider | Example managed models | Publisher id | Good first pick for |
|---|---|---|---|
| Anthropic | Claude Opus 4.7, Sonnet 4.6, Haiku 4.5 | anthropic | agents, coding, long tool chains |
| Mistral AI | Mistral Medium 3, Codestral 2, Mistral OCR | mistralai | code, document parsing, cost |
| xAI | Grok 4.20, Grok 4.1 Fast | xai | document understanding, agents |
| DeepSeek | DeepSeek-V3.2, R1-0528, DeepSeek-OCR | deepseek-ai | reasoning on a budget |
| Qwen | Qwen 3 235B, Qwen 3 Coder | qwen | multilingual, open weights option |
| OpenAI (open weights) | gpt-oss-120b, gpt-oss-20b | openai | open model with a managed option |
Model names and publisher ids verified against Google’s partner-models documentation for this run. Model IDs carry version suffixes; confirm the exact string on the model card before you wire it into code.
Anthropic runs the widest bench on the managed side, with roughly a dozen Claude versions live at once. That matters for a boring reason: version pinning. When Opus 4.7 lands, 4.5 does not vanish, so a workload you validated last quarter keeps answering the same way. Here is the count by provider.
Compare before you commit
Before you move a workload off Gemini onto Claude or Codestral, measure it on your own data. Vertex AI has a Gen AI evaluation service that scores model output, and it works against partner models, including Anthropic and Llama, not just Google’s own. One requirement trips people up: you have to enable the partner model in Model Garden first, because the evaluation runs against the live model. Send the same prompt set through your current model and the candidate, score both on the metrics that matter for your job, and let the numbers decide. A public leaderboard is not your workload.
I treat this as a gate, not a nicety. A model that tops a benchmark can still lose on your retrieval-heavy prompts, your formatting rules, or your language mix, and the only way to find out is to run your own traffic through it. Budget a day. The cost of the evaluation is trivial next to the cost of cutting a production path over to a model that turns out worse on the exact work you do all day. For steady, high-value traffic there is also a capacity-assurance option: Google sells provisioned throughput on some partner models, reserving served capacity for a fixed fee so your requests stay prioritized when the system is busy. Prove the quality first, then decide whether to reserve capacity.
How do you call a partner model?
You still send the request to a Vertex AI endpoint. The URL names the publisher and the model, and you pick where the request is served. There are three endpoint styles, and the choice is about data residency, not features.
A regional endpoint serves the request from one region you name, so the traffic stays there. The global endpoint lets Google serve from any region the model supports, which raises availability and can lower error rates, at the cost of not pinning a location. A multi-region endpoint, still in preview, keeps data inside a broader geography such as the United States or the European Union. One practical note: the global endpoint supports prompt caching but not provisioned throughput, so if you have bought reserved capacity you must call a regional endpoint to use it.
Before any of this works you have to enable the model once in Model Garden and accept its terms. Claude carries one extra rule worth knowing up front: if your billing account is managed by a reseller Anthropic prohibits, you cannot accept the terms or enable Claude, and you have to sort that out with the reseller. Here is a live call against Claude on the global endpoint.
# Call Claude on Vertex AI, global endpoint, streamed PROJECT_ID=my-genai-project MODEL=claude-sonnet-4-5 # confirm exact versioned id on the model card curl -X POST -H "Authorization: Bearer $(gcloud auth print-access-token)" -H "Content-Type: application/json" "https://aiplatform.googleapis.com/v1/projects/${PROJECT_ID}/locations/global/publishers/anthropic/models/${MODEL}:streamRawPredict" -d '{ "anthropic_version": "vertex-2023-10-16", "max_tokens": 512, "messages": [ {"role": "user", "content": "Summarize our refund policy in two sentences."} ] }'
Expected output: a stream of server-sent events, each a JSON chunk with a content delta, ending on a message_stop event. Failure modes: a 403 means the model is not enabled in this project or your billing account hit the Anthropic reseller rule; a 404 usually means the publisher path or region is wrong, for example calling a model on global that only serves from a specific region. Swap streamRawPredict for rawPredict if you want a single non-streamed JSON body back.
In practice
Do not hardcode the global endpoint everywhere just because it is the most available. The first time an auditor asks where a European customer’s prompt was processed, a global answer of anywhere the model supports is not an answer. Default to a regional or EU multi-region endpoint for regulated traffic and keep global for the internal tooling that does not carry personal data.
Self-deploy a proprietary partner model into your VPC
In October 2025 Google added a mode that closes a real gap. You can now take certain closed-source partner models, license them through Google Cloud Marketplace, and deploy them inside your own Virtual Private Cloud, the private network segment you control on Google Cloud. Part 2 covered self-deploying open-weight models like Llama, where you already had the weights. This is different: these are proprietary models you could not host yourself before, now running on your network under your VPC Service Controls policy.
The launch shipped with eight partners: AI21 Labs with Jamba Large 1.6, CAMB.AI with the MARS7 speech model, Contextual AI with a reranker for retrieval, CSM with a two-dimensional to three-dimensional image model called Cube, Mistral AI with a self-deploy build of Codestral, Qodo with the Embed-1 code embedding model, Virtue AI with the VirtueGuard safety model, and WRITER with Palmyra X4. You find them in Model Garden under Model Collections, then Self-deploy partner models. You buy the license, click deploy, and a managed inference service stands up the endpoint on machine types you choose.
Billing is pay-as-you-go on the compute, and your committed-use discounts and reservations apply, so this is not a separate pricing island. What you are buying is not lower cost. It is a promise that the prompt never leaves your perimeter, and a license you can actually sign. Here is how the two paths compare on the things that decide the choice.
| Question | Managed API (MaaS) | Licensed, self-deploy in VPC |
|---|---|---|
| Who runs the servers | Google, serverless | You, on machine types you pick |
| How you pay | per token | per compute hour, plus license |
| Where the prompt is processed | Google-managed endpoint | inside your VPC, under VPC-SC |
| Scaling to zero | automatic, no idle cost | manual or autoscaled, idle costs money |
| When to pick it | almost always, to start | a residency or contract rule forces it |
Cost and latency, the trade you are making
Per-token billing feels cheap until volume climbs, because you pay for every token whether the model was busy or idle a second earlier. A self-deploy replica is the opposite. You pay a fixed cost to keep the machine warm whether it serves one request an hour or a thousand. Somewhere between those two shapes is a break-even, and knowing roughly where it sits keeps you from over-engineering a low-traffic feature or overpaying on a high-traffic one.
Worked example
Take an illustrative managed rate of 6 dollars per million output tokens, and an illustrative self-deploy replica that costs about 2,600 dollars a month to keep warm. Managed cost grows with volume: at 100 million output tokens a month you pay 600 dollars, well under the replica. At 433 million tokens the two lines cross. Past roughly 430 million output tokens a month, the always-on replica is cheaper, and it gets cheaper the more you push through it. These figures are illustrative, not quoted Google prices. Plug your real rate and machine cost into the same shape and the crossover moves, but the shape does not.
Latency runs the other way from what people assume. A warm self-deploy replica in your own region can beat a managed call, but only once it is warm and only while it is not saturated. A managed endpoint has no cold start you own and absorbs bursts for you. For spiky traffic, managed usually wins on the tail latency that users actually feel. For steady high volume in one region, a self-deploy replica can win on both cost and median latency. The break-even chart and the latency answer point the same way most of the time: start managed.
Gotchas that will catch you
Two more that cost people time. First, provisioned throughput, the reserved-capacity option that guarantees you served requests under load, is not available on the global endpoint, so a team that bought capacity and then coded against global quietly gets none of it. Second, the reseller rule on Claude is a hard block, not a warning. If your Google Cloud billing runs through a prohibited reseller, no amount of IAM will let you enable Claude, and you find out at the worst time, mid-integration. Check billing ownership before you promise Claude to a stakeholder.
When a partner model beats Gemini, and how to run it
Here is my rule after building on this catalog. Reach for a third-party model when it clears a bar Gemini does not for your specific job: a Claude version your agent code already targets, a Codestral your team benchmarked on your repo, a reranker that lifted your retrieval scores in a real test. Do not switch on reputation. Run the managed API first, always, because it costs you nothing to stand up and nothing when idle. Move to a licensed VPC deployment only when a named requirement, a data-residency clause or a contract that forbids a shared endpoint, forces the traffic inside your perimeter. Preference is not a requirement.
Next in this series I cover Vertex AI against the Gemini API in Google AI Studio, and why the same model can feel like two products depending on which door you walk through. Before you read it, do one thing: open Model Garden, filter to partner models, and check which of your target models even serve from the region your workload must run in. That single check has saved more launches than any benchmark.
References
- Vertex AI partner models for MaaS, Google Cloud documentation
- Self-deploy proprietary models in your VPC with Vertex AI, Google Cloud blog, October 2025
- Vertex AI Model Garden overview
- Cross-series: Azure Foundry Model Catalog and Models-as-a-Service
About The Author
DrJha