Route coding agents cheaper, plan Gemini costs, gauge Nvidia risk
This week you get levers for budget, speed, and safety. Plus a sober take on Nvidia.
The most important AI news this week
- Google launches Gemini 4 Hydrogen as a cheaper model and makes it the base of the free tier, also for new search features. According to Simon Willison, Hydrogen costs 1.50 dollars per million input tokens and 7.50 dollars per million output tokens. Carbon and Argon double their prices on December 1. Today Hydrogen is only 25 percent cheaper. From December it would be 75 percent. Google says Argon uses fewer tokens per task than Carbon, despite the same price. For comparison, Willison cites Claude Haiku 5.5 and GPT-6 Luna at 0.10/0.50 dollars. What this means for you: You plan budgets and routing now with the price hike in view.1
- Anthropic explained its multi-chip strategy in the IPO roadshow video. Claude runs on AWS Trainium, Google TPUs, NVIDIA, and AMD. Since Wednesday Tom Brown has described how the internal system Mythos got Claude onto AMD in three days, including kernel porting. Anthropic stresses “compute multipliers”. These are methods that multiply-reduce effective compute per training target. That supports speed and resilience against supply shocks. For you this means: Vendors with real chip portability reduce dependencies.2
- NVIDIA NeMo Switchyard helps coding agents distribute model calls on Amazon Bedrock dynamically. The key metric is “cost per solved task”, not price per token. The bundled LiteLLM integration offers stage routing (chooses between two candidates based on recent tool history) and random routing for traffic splits. Note: The reference uses switchyard-server 0.3.0 as a demo and is not for production according to NVIDIA. You test this in a non-production environment and decide routing after that.3
- Multiple reports claim Jensen Huang’s resignation as NVIDIA CEO and tie it to export and cloud access debates. SemiAnalysis and ChinaTalk discuss RASA and the MATCH Act. They also discuss possible remote access limits, investigations, and smuggling allegations. Zvi Mowshowitz collects further allegations and reactions. An independent confirmation beyond these sources is not in the material. Outcome: You test scenarios for tighter China rules and their effects on supply chains, remote access, and rental models.456
- Dyania Health’s Synapsis AI runs inside the hospital firewall and compresses years of patient records into auditable clinical signals. The Cleveland Clinic used it for a study with over 7,800 aortic valve patients, including long-term echos, to model risks. Models are built with NVIDIA NeMo and optimized with TensorRT. Decisions remain with physicians. If you work in healthcare, on-prem with traceability is a verifiable path.7
- In the United States you can link Google Wallet with the Gemini app. Answers will then be informed by your spending patterns. This is opt-in. You get personal budget and analysis questions answered more directly.8
- Meta ports Cuttlefish, the high-fidelity Android Virtual Device framework, natively to Macs with Apple Silicon. Teams save on parallel-running emulators and get closer to real device behavior. That strengthens mobile testing without rare hardware.9
The week for your company
- Make coding agents cheaper. Set up a sandbox with LiteLLM plus NeMo Switchyard and compare two Bedrock models with stage routing against single models. Measure “cost per solved task” and abort rates. Then decide on standard routing for everyday tasks. Note: The reference is demo code, non-production only.3
- Actively manage model costs. Plan now for the price doubling for Gemini 4 Carbon and Argon from December 1. Check Hydrogen for execution steps and use stronger models only for hard sub-tasks. Compare with cheaper alternatives if quality is enough. This keeps costs controllable.1
- Automate legacy portals. Shield uses agents that operate old interfaces and take ticket work off your plate. Start with a narrow process with clear fields. You gain time for cases that need judgment. Mind UI brittleness and logging.10
- Pace product work with agents. OpenAI cites case studies. Gondola reports double revenue and 50 percent more output for the lead engineer. Amplitude ships improvements in 1 to 2 days. Holiday Extras lifts conversion. Numbers come from the vendor. Build your pilot from this: one target KPI, four weeks runtime, clear measurement before and after.111213
- Verify network segments mathematically. AWS Tiros validates WAN segmentation statically through modeling instead of sample traffic. That reduces blind spots and proves separation against the policy. For compliance teams this is a solid addition to packet tests.14
Try next week
- Set up the Switchyard sample in a test AWS environment and enable stage routing between two Bedrock models. For example, one capable and one efficient.3
- Run your typical coding agent workflow 20 times and record: solved vs. aborted, total cost per solved task, duration.3
- Compare against two single-model baselines. Choose the cheapest variant that passes your quality checks as the default route.3
A good find
- Microsoft Research shows this: Only 8 percent of long coding sessions need context compaction but cause 44 percent of costs. A background method (APEC) cut costs by up to 25.4 percent in tests and runtime by 9.6 percent. If you run coding agents, it pays to measure these 8 percent and compact asynchronously.15
Sources
- llm-gemini 0.37 (simonwillison.net)
- Anthropic's IPO roadshow video (simonwillison.net)
- Route coding agent requests to the right model on Amazon Bedrock with NVIDIA NeMo Switchyard (aws.amazon.com)
- Nvidia and the World Without Jensen (newsletter.semianalysis.com)
- Jensen Out, MATCH Act In? (chinatalk.media)
- Jensen Huang Resigns After Having Another Affair In Addition to the Chinese Spy (thezvi.substack.com)
- How AI Is Helping Care Teams Get Patients the Right Care, Sooner (blogs.nvidia.com)
- Connect Google Wallet to Gemini for personal finance insights (blog.google)
- Cuttlefish on Mac: Running Android’s Virtual Device Framework Natively on Apple Silicon (engineering.fb.com)
- Shield helps technicians hand off legacy portal work with OpenAI (openai.com)
- Gondola doubles revenue as its team builds more with Codex (openai.com)
- How Amplitude ships product improvements in 1–2 days with OpenAI (openai.com)
- Holiday Extras is rethinking what it can build with Codex (openai.com)
- Using mathematics to validate network configurations (amazon.science)
- Microsoft Research Focus – Edition #98 (microsoft.com)