Google's Gemini 4 Carbon takes first place on the Artificial Analysis Intelligence Index

Google's Gemini 4 Carbon now leads an independent intelligence index, and Apple is opening Siri AI to third-party apps. OpenAI, Anthropic and Deep Principle shared new model and tooling results, while new research tested routing and robot policies.

1. Gemini 4 Carbon takes first on the Artificial Analysis Intelligence Index at 60

   Gemini 4 Carbon moves up from 59 to 60 to lead an index that Artificial Analysis scores with its own evaluations rather than lab-reported numbers.

   Models: Gemini 4 Carbon

2. Apple opens Siri AI to third-party apps with iOS 27.3 Developer Preview

   Any app can now plug into Siri AI, and app skills chain several tasks into one action without users naming the app, per Apple's developer documentation.

3. Deep Principle releases open-weight Hydrogen-1.0 science model built on GLM-5.3

   An open-weight science model on GLM-5.3 claims a FrontierScience Research lead over GPT-6 Astra, though the scores and cost figures come from its own evaluation setup.

   Models: GPT-6 Astra, GLM-5.3

4. OpenAI says GPT-6 Astra designed a hardware device for Codex in under a week

   GPT-6 Astra produced an enclosure, circuit board and firmware in under a week, but manufacturing tolerance misses meant several revision rounds before production.

   Models: GPT-6 Astra

5. Google will let schools disable its AI tools but lose access to Search

   Because AI is now built into Search, schools that switch off Google's AI tools also give up Search, which leaves them a hard choice.

6. Anthropic releases Message Threads to cut time to first token in long conversations

   Sending only changed content each turn cuts time to first token by about 35 percent for Messages API requests of 500,000 to one million tokens.

7. Route Anything paper studies model and policy routing across omni-modal AI systems

   Across 374,401 scored runs, learned routers beat the best fixed candidate by 0.3 points on 16 benchmarks while using 24% less policy time.

8. RoboColosseum distributed benchmark tests VLA robot policies and frontier agents across four embodiments

   In 5,699 trials at 124 sites, frontier agents rival specialized VLA robot policies, but every system tested still struggles with open-world tasks.

Written by Claude (claude-opus-5-5) from the stories below.