Space Bunny Alpha: a free 1M-token stealth model lands on OpenRouter

Multimodal, always-reasoning, 524K output tokens, $0 in preview. Plus Aion 3.5 doubles context at flat pricing and Vercel sandboxes reach private networks.

Nowline SEP 25 2:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A free 1M-context model you can hammer today

    OpenRouter is routing Space Bunny Alpha, an anonymous stealth model: a 1M-token context window, up to 524K output tokens, native image input, and reasoning on every call — all at $0 in and out. The catch: the unnamed provider may retain your prompts and completions, so keep proprietary data out and a tested fallback wired in. Past "Alpha" stealth listings have consistently turned out to be real labs, which makes this a free frontier-scale testbed while the preview lasts.

  • Aion 3.5 doubles its context and keeps the old price

    Aion Labs shipped Aion 3.5 and 3.5 Mini with a 262K-token context window (up from 131K) at unchanged pricing — $3/$6 per million for 3.5 and $0.70/$1.40 for the Mini, both with cached-input discounts. Reasoning runs on every call at low, high, or max effort, and the Mini swapped its DeepSeek base for a GLM one. No public SWE-bench or Terminal-Bench figures exist yet, so benchmark it on your own workload before switching.

  • Vercel sandboxes can now reach your private network

    @vercel/sandbox 3.5.0 adds --network-id on the CLI and networkId in the SDK to attach an agent sandbox to a Secure Compute network — so an ephemeral agent can hit private databases and internal APIs, not just the public internet. The 4.4.x line also tags AI-agent traffic in the user-agent header and switches size readouts to binary units (KiB/MiB).

  • An open MIT agent harness that undercuts Codex

    Unreal Labs open-sourced Unreal Agent, a model-agnostic Go harness that fires tool calls asynchronously so the model keeps working — and the prompt cache stays warm — while tools finish in the background. It reportedly scored 57.9% on Terminal-Bench 4.0 for $1,428, roughly 39% cheaper than a comparable Codex run, with OpenAI, OpenRouter, Fireworks, and Ollama backends, and drew around 900 GitHub stars in two days.