Claude Opus 4.7 and Gemma 4
- Anthropic made Claude Opus 4.7 generally available on April 16, pitching it as a stronger coding model with better vision and the same API pricing. - Google DeepMind launched Gemma 4 on April 2 as an Apache 2.0 open-model family with multimodal input, 256,000-token context, and four sizes. - Together they show the split between premium hosted models and open local ones is narrowing fast. (anthropic.com)
Large language models are software that predict the next token, or text fragment, and newer versions are now splitting into two camps: hosted premium models and open local ones. Claude Opus 4.7 and Gemma 4 landed this month on opposite sides of that divide. (anthropic.com) (blog.google) Anthropic said on April 16 that Claude Opus 4.7 is now generally available, with gains over Opus 4.6 in advanced software engineering and better high-resolution vision. It is available in Claude products, the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. (anthropic.com) Anthropic kept pricing unchanged from Opus 4.6 at $5 per million input tokens and $25 per million output tokens. The company said developers can call it through the Claude API as `claude-opus-4-7`. (anthropic.com 1) (anthropic.com 2) Google DeepMind introduced Gemma 4 on April 2 as an Apache 2.0 open-model family meant to run on hardware from phones to workstations. Google said the line comes in four sizes: E2B, E4B, 26B Mixture of Experts, and 31B Dense. (blog.google) (ai.google.dev) Multimodal means a model can take more than plain text, the way a person can read a message and look at a photo at the same time. Google’s model card says Gemma 4 handles text and images across the family, while the smaller models also support audio and video features. (ai.google.dev) Context window is the model’s working memory, or how much text it can keep in view before it starts forgetting earlier details. Google said Gemma 4 supports up to 256,000 tokens and more than 140 languages, with native system prompts and function calling for agent-style workflows. (ai.google.dev) (developers.googleblog.com) That leaves developers choosing between a paid frontier model optimized for remote coding work and an open-weight family designed for local deployment and fine-tuning. Anthropic emphasized harder coding tasks and verification behavior; Google emphasized edge devices, licensing, and model efficiency. (anthropic.com) (blog.google) The open-source side is moving just as fast. The OpenClaw repository showed about 365,000 GitHub stars when checked on April 27, far above the roughly 302,000 figure in the prompt, suggesting the social-post snapshot was already stale. (github.com 1) (github.com 2) One detail in the prompt was also off on timing: Rust’s official release notes show version 1.95.0 on April 16, while the page is already tracking 1.97.0-nightly builds. I could not verify a Rust 1.96 stable release from the official sources I checked. (doc.rust-lang.org) The pattern across these releases is concrete: premium models are adding stronger coding and tool behavior, while open models are adding multimodal input, long context, and local deployment. For anyone building with artificial intelligence in April 2026, the shelf is changing every few weeks. (anthropic.com) (ai.google.dev)