Meta's Muse Wonder lands on the API at $4/$20, 1M-token context
Meta Superintelligence's new coder scores 52; the stronger Wonder Full is locked to $50-$200 plans, not the API. Plus DigitalOcean evals and QwenGyre.
Copy markdown
A frontier coder you can call today
Meta's Superintelligence Lab shipped Muse Wonder on an OpenAI-compatible Model API at $4/M input and $20/M output, with up to a 1,048,576-token context and cache reads at $0.20/M. It is built for coding, long-running agents that plan and delegate to sub-agents, and computer use, and it scores 52 on the independent intelligence index. Vision, tool calling, streaming, and strict JSON output are all supported.
The catch: the best tier never touches the API
The stronger Wonder Full (index 56) sells only through Muse Code subscriptions at $50 and $200 a month, so you cannot call it from the API at all. And by default the Muse agent trains on your conversations and tool calls, so check the data settings before wiring it into anything that touches private code.
DigitalOcean turns on agent observability
DigitalOcean opened Agent Data Services and Agent Learning Services - Insights, Signals, and Evaluations & Simulations - in preview: capture agent traces, flag failure modes, and simulate fixes without building your own eval stack. It pairs with Agent Droplets from $50/mo. Trace capture creates compliance work, so mind your retention settings.
Alibaba open-sources an agent-training framework
Alibaba reportedly released QwenGyre, an open-source framework for training tool-using agents - the RL and harness plumbing you would otherwise build yourself on top of Qwen. It is free to self-host; license terms were not stated at launch.