DeepSeek keeps pushing capability at the frontier
DeepSeek released v4-flash-vision-exp, adding vision capabilities to a model line that previously had people working around its inability to process images. The thread noted that DeepSeek's inability to handle Playwright screenshots was the main thing keeping some developers on Anthropic's Sonnet. That gap is now closing.
A separate thread surfaced 'Ox Alpha,' a mystery model appearing on OpenRouter with no clear provenance. The thread debated whether it was Anthropic or OpenAI based on guardrail behavior, with the working heuristic being: absurd guardrails point to US labs, reasonable or absent guardrails point to Chinese models. That framing says a lot about how the developer community has started to think about model identity.
The pattern across both threads: DeepSeek is moving fast enough that it is eliminating the specific capability gaps that were keeping developers on more expensive US models. Vision was one of the last meaningful holdouts.
So what?
If you are paying Anthropic or OpenAI rates partly because of specific capabilities like vision or multimodal reasoning, those gaps are closing fast. The calculus on cost versus capability is shifting every few weeks. Build your model abstraction layer so you can swap providers without rewriting your application.