AITOP WeeklyIssue 72087 stories
Gemini 4 Argon debuts, Meta opens 816B model, Microsoft pivots chips
Singapore·12 Oct – 18 Oct·Ten minutes
Google launched Gemini 4 Argon on 13 October, scoring 97.5% on ARC-AGI-3, supporting 1M token outputs, and entering at US$2 per million input tokens and US$10 for output — with prices doubling on 1 December. Two days later, Meta open-sourced the 816-billion-parameter Muse Spark Open with 1M context, free for commercial use under the Muse Model License. Microsoft announced its in-house Maia chips will become its primary inference system, while Anthropic reported US$16.8 billion Q3 revenue with growth cooling to 45%. DeepSeek answered with V4.1-Flash-1017, cutting API prices and beating GPT-5.6 Sol on DeepSWE, 77.0 versus 73.0.
01
Five things this week
Enough to retell1
2
3
4
5
02
US vs China
The same week, both sides| Dimension | United States | China | What it means for ASEAN |
|---|---|---|---|
| New models | Google ships Gemini 4 Argon; Meta open-sources 816B-parameter Muse Spark Open | DeepSeek launches V4.1-Flash-1017; StepFun open-sources 600B Step 5 Preview | Open and closed models landed the same week — wider selection for ASEAN builders. |
| Prices | Argon API at US$2 input / US$10 output per million tokens, rising in December | DeepSeek V4.1-Flash-1017 cuts API prices in step with launch | Chinese price advantage persists — worth comparing before committing to US APIs. |
| Benchmarks | Argon scores 97.5% on ARC-AGI-3 and tops the Vals Index | DeepSeek hits 42.3 on Terminal-Bench 4.0 and 77.0 on DeepSWE | Chinese models now viable replacements for imported ones on coding tasks. |
| Policy | — | Epoch AI: open-weight models trail closed ones in cyberattack capability by 6–7 months | ASEAN open-model deployments need security and compliance risk review. |
2087Stories
345China
16ASEAN
251Sources
AITOP · generated by the editorial system