Meta Opens The Vault. Sort Of.
On Oct 15 Meta released Muse Spark Open, an 816B-parameter open-weight model with a 1M-token context window. Its best models stay closed.
Meta says Muse Spark Open is the most capable open-weight model in the US. That is Meta's own evaluation claim, and no scores were published in the coverage we read. It is built to run locally on a Mac Studio with an M5 Ultra, but Engadget reports it needs at least 1 TB of RAM, so this is not a laptop model.
The pattern matters more than the model: Muse Wonder Full, Meta's top model, is limited to Muse Code subscribers, and Muse Wonder costs $4/$20 per million tokens through Meta's API. Open weights below, closed frontier above. License terms were not stated in the reports.
Gemini 4 Argon: Announced Sep 30, Opening Up Now
Argon launched limited to vetted cyber defenders. One tracker says general availability began Oct 13, with intro pricing that doubles on Dec 1.
Trackers put Argon at 52.6 on the Artificial Analysis Intelligence Index and report an intro price of $2/$10 per million tokens, rising to $4/$20 on December 1, 2026. It can return up to 1M output tokens in a single response.
Caution: the same tracker contradicts itself on whether Argon has a general availability date. Check Google's own pricing page before building on those numbers.
GPT-6.1 Sol Ships. GPT-6.1 Astra Reportedly Doesn't.
At DevDay on Sep 29, OpenAI launched GPT-6.1 Sol at $2/$10 per million tokens. A report says the bigger Astra release was cancelled after safety tests.
Sol scores 51.8 on the Intelligence Index at about $0.72 per task. DevDay also brought an Ultrafast mode (up to 300 tokens/sec at 6x the price) and a $500/month Pro 500 tier.
The Astra cancellation is attributed to the Wall Street Journal, but we only found it secondhand through a tracker. Treat it as unconfirmed until OpenAI says something. GPT-6 Astra, released Sep 3, is still OpenAI's top shipped model.
Two Big Open Models Are Promised. Neither Has Weights Yet.
Mistral Large 4 is in preview with a trillion parameters. Alibaba announced Qwen 4 but is still training it.
Mistral Large 4 ("Le Chonk"), previewed Oct 6, has 1T total and 49B active parameters and is natively multimodal. It is only available through a guardrailed API preview, with weights promised by the end of October. No price or license has been published.
Qwen 4 was announced Sep 22 in four tiers: Max, Flash, Plus, and an open-weight 27B. As of early October there was no date, price, or weights.
P(doom): What AI Lab Leaders Say Out Loud
P(doom) is a person's guess at the chance AI causes a catastrophe. It is an opinion, not a benchmark, and it is not measured per model.
Figures publicly attributed to people (per Wikipedia's P(doom) page, which cites press interviews; dates vary): Dario Amodei, Anthropic, 10-25%. Elon Musk, xAI, roughly 10-30%. Geoffrey Hinton, 10-20%. Yoshua Bengio, 20%. Demis Hassabis, Google DeepMind, has said only that it is above zero.
These numbers are not comparable. People differ on the time frame, whether AGI is assumed, and what counts as doom. We show them because readers ask, not because they rank the models.