If your AI/ML integration plan still names Google "AI Platform," it is now two rebrands behind. The bigger surprise for teams returning to this market in 2026 is that "Vertex AI" is stale too. The three hyperscaler platforms that most B2B software teams build on have all renamed, restructured, or consolidated their AI stacks within the past eighteen months, and in at least two cases the SDK you import changed along with the logo. That churn, not model quality, is the integration cost that actually shows up in sprint planning.
What these platforms are actually called now
Google has been consolidating Vertex AI into its wider Gemini Enterprise agent stack, and the naming in the documentation has been moving along with it. Rather than take any single summary on faith — including this one — check the official Vertex AI release notes and the console navigation directly. The product lineage runs AI Platform to Vertex AI to the current Gemini-branded agent tooling, and documentation paths, console labels, and support articles have been migrating as that consolidation proceeds. What matters for planning is less the exact brand string on any given week than the fact that it has already changed twice in this product's life and may change again before your renewal date.
Microsoft made a parallel move, and documented it more explicitly. The company's own documentation includes a mapping table: the brand formerly known as Azure AI Studio and then Azure AI Foundry is now Microsoft Foundry, and Azure AI Services is now "Foundry Tools." The rename came with real API surface changes rather than cosmetic ones. The Assistants API (Agents v0.5/v1) became the Responses API (Agents v2), monthly api-version parameters gave way to stable v1 routes, and a scatter of packages — azure-ai-inference, azure-ai-generative, azure-ai-ml — consolidated into a unified azure-ai-projects 2.x client hitting a single project endpoint. Azure Machine Learning still exists as a separate service for custom model training; Foundry is the agent and model application layer.
AWS is the stable one. Amazon SageMaker AI kept its name, and the tooling list on its pricing page reads much as it has for several years: Studio, JupyterLab, Feature Store, Ground Truth, JumpStart, HyperPod, MLflow, Canvas, and a set of inference modes. If naming stability across procurement cycles is worth something to your organization, that is a real differentiator and one of the few you can verify without a sales call.
How the pricing actually works
Published rates exist, so there is no reason to guess. Google's Vertex AI generative AI pricing page lists Gemini 2.5 Pro at $1.25 per million input tokens for prompts up to 200K tokens and $10 per million output tokens, with input rising to $2.50 per million above the 200K threshold. Gemini 2.5 Flash-Lite sits at $0.10 input and $0.40 output per million. Three structural levers matter more than the headline number: cached input tokens are discounted roughly 90 percent, batch and flex processing runs 50 percent off, and a Priority tier carries an 80 percent premium for prioritized handling. Tuned models bill at 1.5x the base model rate, and only requests returning HTTP 200 are charged.
Amazon's SageMaker AI pricing page is component-based rather than token-based for most workloads. Serverless inference is listed at $0.00004 per second of compute, data processing at $0.016 per GB in and out, and Feature Store at $1.25 per million write units and $0.25 per million read units. Model customization training is quoted at $0.5 per million tokens. Savings Plans offer up to 64 percent off in exchange for committed usage, applied across instance families and regions. The free tier is genuinely a trial rather than a standing allowance: 250 hours of ml.t3.medium Studio notebooks, 125 hours of m4.xlarge or m5.xlarge real-time inference, and 150,000 seconds of serverless inference duration, all within the first two months.
Databricks prices Foundation Model Serving in DBUs rather than dollars, which is the single most common source of budget surprises. Llama 4 Maverick is listed at 7.143 DBU per million input tokens and 21.429 DBU per million output tokens; provisioned throughput entry capacity ranges from 25 to roughly 85.7 DBU per hour. Because a DBU converts to dollars at a rate that depends on your tier, cloud, and region, and because underlying cloud compute is billed separately, a DBU figure is not a price until you multiply it by your own contract rate.
Why benchmark claims should not drive this decision
Vendor case studies and cross-platform performance comparisons in this category are frequently unsourced, and it is worth being blunt about that. MLPerf, run by MLCommons, is a real and rigorous benchmark suite, but its submissions measure hardware and system configurations on defined training and inference tasks. It does not score managed-platform developer experience, integration effort, governance tooling, or total cost for a B2B SaaS workload. A vendor citing MLPerf standing has told you something about silicon, not about whether their SDK will still exist next year. Any specific percentage improvement attributed to a named customer should be traceable to a case study you can open and read; if it is not, treat it as marketing copy.
A practical evaluation checklist
- Read the SDK deprecation notices before the feature list. Google's release notes flag the generative AI module in the Vertex AI SDK as deprecated, with a documented migration path to the Google Gen AI SDK. Microsoft's mapping table effectively says the same thing about its legacy packages. Those pages predict your migration work better than any feature matrix.
- Separate the naming change from the API change. Google's consolidation largely preserved existing endpoints. Microsoft's restructured the agent API and its terminology outright, moving from Threads, Messages, Runs, and Assistants to Conversations, Items, Responses, and Agent Versions. One is a documentation problem; the other is a refactor.
- Model your caching and batch ratio early. With a roughly 90 percent cached-input discount and a 50 percent batch discount on Google's platform, an architecture that reuses context and tolerates asynchronous processing can cost a fraction of the same workload run synchronously against cold prompts.
- Price idle capacity, not just calls. Provisioned throughput and always-on endpoints bill for availability. Serverless and batch modes bill for work performed. Match the billing mode to your actual traffic shape before committing.
- Weigh model breadth against platform lock-in. Microsoft Foundry advertises access to over 1,900 models from providers including OpenAI, Anthropic, Mistral, xAI, Meta, and Hugging Face, plus a catalog of over 1,400 tools. Breadth reduces model lock-in while deepening platform lock-in.
The durable takeaway
The lesson for B2B teams is procedural rather than a ranking. Pin your internal documentation links to official pricing and release-notes pages instead of to third-party summaries, re-check them each quarter, and assume that any comparison article, including the product names written into your own architecture docs, has a shelf life measured in months rather than years. The teams that got burned in this cycle were not the ones that picked the wrong platform; they were the ones whose runbooks still referenced a console screen that no longer exists.