AI digest: cheaper models, sharper tools, and China's chip workarounds
Anthropic drops Claude Sonnet 5 at a lower price point, Google speeds up image generation, and Deepseek finds a way to squeeze more out of fewer chips.
A busy few days. The main thread running through everything right now is cost: cheaper inference, cheaper models, cheaper image generation. The race to the bottom is very much on.
Anthropic launches Claude Sonnet 5
This is the one to pay attention to. Claude Sonnet 5 sits below Opus on price but reportedly beats Opus 4.8 on at least one major knowledge work benchmark. That matters a lot for anyone building agentic workflows who doesn’t want to pay Opus prices for every call. This feels like a genuine shift, not just a minor update.
Deepseek’s DSpark boosts inference speed by up to 85 percent
DSpark uses a small model to propose token candidates that a larger model then checks in batches. The result is 60 to 85 percent faster per-user response speeds, without needing more chips. Given the tightening US export controls, this kind of efficiency work is exactly how China keeps moving forward. Worth watching closely.
Anthropic launches Claude Science for researchers
Claude Science is a dedicated workbench for computational research, with over 60 preconfigured skills covering areas like genomics and chemistry. There’s also a verification agent that checks citations and calculations automatically. Bundling workflow tools rather than just dropping a new model is a smart move, and it targets a user group that genuinely needs reliability over novelty.
Google launches Nano Banana 2 Lite
Four seconds per image at $0.034 is fast and cheap enough to matter for production use cases. Google is also pairing it with Gemini Omni Flash for video generation via API. Chaining the two to go from text prompt to finished video in one pipeline is an interesting pattern to experiment with.
Meta secretly tested rivals’ chatbots using thousands of minor-perspective prompts
This one is uncomfortable. Meta sent over 45,000 crisis prompts to OpenAI, Google, and Character.AI systems without telling them, using contractors posing as minors. Whatever the intent, running covert adversarial tests on competitors raises real questions about where the line sits between safety research and something else entirely.