Google’s flash ✨

Google Makes Gemini 3 Flash the Default

Speaking of Google…on Wednesday, it launched Gemini 3 Flash, a faster, cheaper model designed to be the new workhorse. It’s now the default inside the Gemini app and AI Mode in Search.

Here’s what’s new: 

• Stronger performance than the last Flash. Google says Gemini 3 Flash makes a big jump over Gemini 2.5 Flash and can even trade punches with frontier-tier models (including Gemini 3 Pro and GPT-5.2) on some evaluations.

• More multimodal-native behavior. The model handles mixed inputs (video, sketches, audio) better than earlier models and returns more “visual” answers (think tables and images).

• Pricing moved up slightly—but Google argues it nets out. It’s listed at $0.50 / 1M input tokens and $3.00 / 1M output tokens, up from 2.5 Flash. Still, Google claims it’s faster and uses fewer tokens than 2.5 Pro on “thinking” tasks, which could reduce total spend for certain workloads.

And that’s not all: Google is also threading “build with Gemini” deeper into the product. Its vibe-coding tool Opal is now integrated into Gemini on the web via the Gems manager, letting people create mini-apps from plain-English prompts. There’s also a visual step-based editor if you want to tweak the workflow without touching code.

OpenAI and Amazon Consider a $10B+ Investment

OpenAI is reportedly in discussion with Amazon about a potential investment that could exceed $10 billion, alongside an agreement to use Amazon’s in-house AI chips.

What we know so far: The talks are still fluid, but the deal would likely pair capital with infrastructure. Amazon would deepen its role as a compute provider by supplying its Trainium and Inferentia chips, while OpenAI gains another hyperscale partner beyond Microsoft. 

The timing matters: OpenAI finalized its for-profit restructuring in October, loosening Microsoft’s exclusivity and giving OpenAI more freedom to raise capital and partner across the ecosystem. OpenAI has already committed more than $1.4 trillion in long-term infrastructure deals, including a recent $38B capacity agreement with AWS.

GPT-5 Shows It Can Perform Novel Lab Work

GPT-5 has, for the first time, demonstrated the ability to carry out forms of novel laboratory work. The results suggest the model can go beyond analyzing papers or summarizing data to contributing directly to experimental workflows.

What it did: In controlled settings, GPT-5 was able to reason through experimental steps, generate hypotheses, and help design or interpret lab processes in ways that meaningfully supported human researchers. 

Nebius Token Factory—Post-training

Nebius Token Factory just launched Post-training—the missing layer for teams building production-grade AI on open-source models.

You can now fine-tune frontier models like DeepSeek V3, GPT-OSS 20B & 120B, and Qwen3 Coder across multi-node GPU clusters with stability up to 131k context. Models become deeply adapted to your domain, your tone, your structure, your workflows.

Deployment is one click: dedicated endpoints, SLAs, and zero-retention privacy. And for launch, fine-tuning GPT-OSS 20B & 120B (Full FT + LoRA FT) is free until January 9. This is the shift from generic base models to custom production engines.

Turn ideas into engineering specs

BrainGrid is an AI product management agent that converts plain-English product ideas into scoped tasks and production-ready prompts by analyzing your codebase.

How you can use it:

• Translate feature ideas into clear, testable engineering specs

• Generate acceptance tests and dependency mappings automatically

• Review and sanity-check code generated by tools like Cursor or Claude Code

• Keep AI outputs aligned with your existing repo and architecture

Pricing: Paid with a free trial

An AI workspace built for answers

Sup AI is a unified AI platform that automatically routes tasks across frontier models, combines their outputs, and surfaces answers with confidence scoring and citations. 

How you can use it:

• Ask complex questions and get cited, confidence-scored responses

• Run research workflows across multiple models without manual switching

• Store and query project-specific memory with secure team access

• Balance cost and performance with automatic model selection

Pricing: Free and paid plans available