Claude API model guide gets a new top tier

Anthropic’s Claude docs now put Opus 4.7 at the top of the stack.

The latest Claude models overview is more than a product page refresh. It gives developers a clean decision tree for picking between Claude Opus 4.7, Claude Sonnet 4.6, and Claude Haiku 4.5, while also spelling out how IDs, pricing, and cloud endpoints work across Amazon Bedrock, Vertex AI, and Microsoft Foundry.

For teams shipping AI features, the useful part is the specificity. Anthropic now publishes model IDs, alias behavior, token limits, output caps, and knowledge cutoff dates in one place, which makes it easier to choose a model without guessing.

Opus 4.7 is the model Anthropic wants you to notice

Anthropic is pretty direct about where Claude Opus 4.7 fits: it is the company’s most capable generally available model for complex reasoning and agentic coding. The docs also call out a “step-change improvement” in agentic coding over Opus 4.6, which is strong language for a docs page and tells you where Anthropic thinks the biggest upgrade landed.

That matters because model choice is no longer just a price question. If your product needs long-context work, code synthesis, or multi-step tool use, Anthropic is steering you toward Opus first, then Sonnet if you want a better speed-to-intelligence mix, and Haiku if latency matters most.

• Opus 4.7: 1M-token context, 128k max output, $5 input and $25 output per million tokens
• Sonnet 4.6: 1M-token context, 64k max output, $3 input and $15 output per million tokens
• Haiku 4.5: 200k-token context, 64k max output, $1 input and $5 output per million tokens

There is also a practical signal buried in the comparison table: Opus and Sonnet now share the same 1M-token context window, so the old assumption that the premium model is the only one worth using for long documents is gone. The real split is now output depth, reasoning quality, and cost.

Versioning is getting stricter, and that helps developers

One of the most useful parts of the page is the explanation of model IDs. Anthropic says every Claude model ID is a pinned snapshot. If the ID contains a date, it is fixed to that release. Starting with the 4.6 generation, even dateless IDs are still pinned snapshots, not evergreen pointers.

That is a good thing for production systems. It means fewer surprises when a model updates under the hood, and it makes rollback planning much simpler. It also means developers need to pay attention to aliases, because older generations may expose alias names that resolve to dated IDs.

Anthropic also clarifies how the same model travels across platforms. On AWS Bedrock, the IDs use Bedrock-style names for some models, while Vertex AI uses its own endpoint conventions. The Claude Platform on AWS keeps the same model IDs as the first-party Claude API, which should reduce confusion for teams deploying in multiple clouds.

“The best way to use a foundation model is to pick the smallest one that can do the job.” — Dario Amodei, Anthropic CEO, in a 2024 interview on the Dwarkesh Podcast

That quote fits this page well. Anthropic’s docs do not push every developer toward the biggest model; they give enough detail to make a cost-performance tradeoff on purpose. In practice, that means a customer support bot, a code review assistant, and a research workflow can all land on different Claude models without needing separate mental models for each platform.

The cloud story is as important as the model story

Anthropic is also making it clear that model access is no longer limited to one API. The overview says current Claude models are available through the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. That distribution matters because procurement, region control, and cloud commitments often decide which AI model gets used before benchmark scores do.

Anthropic’s docs also highlight endpoint behavior. Starting with Sonnet 4.5 and newer, Bedrock offers global endpoints for dynamic routing and regional endpoints for geographic data routing. Vertex AI goes one step further with global, multi-region, and regional endpoints. For regulated teams, those distinctions are not trivia; they determine whether a deployment is even possible.

• Claude Platform on AWS uses first-party Claude API IDs, not Bedrock IDs
• Bedrock adds global and regional endpoint choices for newer Claude models
• Vertex AI adds global, multi-region, and regional endpoint choices
• The Models API exposes max_input_tokens, max_tokens, and a capabilities object programmatically

There is another subtle but important detail: Anthropic says the Max output values in the comparison table apply to synchronous Messages API calls, while the Message Batches API can go higher for some models with a beta header. That kind of footnote matters to anyone building batch workflows, especially if they are pushing long-form generation or offline processing.

What this means if you are choosing a Claude model today

If you are building against Claude now, the docs point to a simple decision path. Use Opus 4.7 when the task is hard and mistakes are expensive. Use Sonnet 4.6 when you want a better balance of speed and intelligence. Use Haiku 4.5 when cost and latency matter more than deep reasoning.

The bigger story is how much Anthropic has tightened the operational details around its models. You now get clearer pricing, clearer token caps, clearer versioning, and clearer cloud deployment rules in one place. That reduces the amount of guesswork in architecture reviews and makes it easier to compare Claude with models from OpenAI and Google on something more concrete than marketing copy.

My read: Anthropic is signaling that Claude is no longer a single flagship model story. It is a menu with sharp boundaries, and the next wave of adoption will come from teams that match the model to the job instead of defaulting to the biggest one. If you are already on Opus 4.6 or older, the practical question is simple: does your workload justify moving to Opus 4.7 now, or can Sonnet 4.6 cover it for less?

For a deeper look at prompt behavior and output quality, see OraCore’s Claude 4 prompting guide. If Anthropic keeps this level of documentation up to date, model selection gets easier, and production mistakes get cheaper.