Google ships Gemini 3.7 Flash and OpenAI unleashes Ultrafast

01.Google Ships Gemini 3.7 Flash, Beats Rivals at Half Price

Google DeepMind released Gemini 3.7 Flash on Saturday, just three weeks after its predecessor, Gemini 3.6 Flash. The company calls it its most intelligent workhorse model yet for coding and AI agents, crediting what Google describes as "awesome algorithmic improvements" for the rapid jump in capability without a full architecture change.

Code quality sees the strongest gains over 3.6 Flash. On the FrontierCode benchmark, the new model scores 43.6 percent, up from 34.4 percent. On DeepSWE it hits 65.3 percent versus 49.0 percent. According to Google's own measurements, both scores put Gemini 3.7 Flash ahead of Claude Sonnet 5 and GPT-5.6 Terra Google DeepMind. The model also shows gains in web development, document comprehension, and business process automation.

The API is available through Google Cloud, AI Studio, and Antigravity. Launch pricing sits at $0.75 per million input tokens and $3.75 per million output tokens, which is 50 percent cheaper than Gemini 3.6 Flash The Decoder. Google says this pricing holds through the end of the year and both model generations now share the same price point.

The speed of release is notable even for the current baseline. Three weeks between Flash iterations was unthinkable a year ago, and it signals that Google's Flash line has moved from research preview to an industrial workflow tool that ships at its own pace.

thinkidiot take: The real story here is the shipping velocity. Flash models are now cycling fast enough that benchmark advantage is a snapshot, not a fortress. If you built your automation stack around 3.6 Flash pricing for next quarter, expect to renegotiate before the year ends.

02.OpenAI Partners with Cerebras to Unleash Ultrafast Mode at 750 Tokens Per Second

OpenAI released a preview of Ultrafast mode on Friday, running GPT-5.6 Sol through chipmaker Cerebras at up to 14 times the speed of standard processing TechCrunch. That translates to roughly 750 output tokens per second, which OpenAI says points toward "more useful work per second" rather than just faster empty chat.

Ultrafast is currently in preview for a small group of enterprise customers. OpenAI plans to expand access as Cerebras capacity grows OpenAI. The primary target workflows are incident response, customer service automation, financial market analysis, and e-commerce operations.

Ultrafast joins Anthropic's Fast mode in the growing category of accelerated inference tiers, but OpenAI's partnership with Cerebras on silicon is what makes this interesting. Rather than just caching more aggressively or routing to a faster API endpoint, OpenAI is putting its flagship model directly onto custom inference hardware designed for throughput.

thinkidiot take: Speed wars at the frontier have been predictable since early last year. The real question is whether enterprise buyers will use 750-token-per-second GPT-5.6 Sol differently than the standard version, or just get more of the same output, faster. If the latter, it becomes a commodity play and OpenAI's moat gets thinner at the very scale where margins matter most.

03.Nvidia Wins $500 Billion in Apollo Financing to Back Its AI Factory Vision

Nvidia announced that Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR agreed to commit up to $500 billion toward building AI data centers. But the bigger play is Nvidia's strategy to create a secondary market for aging GPUs TechCrunch.

To secure those commitments, Nvidia agreed to guarantee that its chips used as collateral in these deals will retain their value. If GPUs backing a loan don't hold the anticipated resale price, Nvidia covers up to 25 percent of the shortfall. The goal is an ecosystem where used AI hardware remains a liquid asset class instead of depreciating to zero, keeping demand for new Nvidia gear flowing longer than a single generation cycle would allow.

The risk is what financiers call "wrong way" exposure: Nvidia's obligations grow as demand weakens. When that happens, revenues typically compress too. Nvidia CEO Jensen Huang has publicly addressed whether this amounts to circular financing, saying the initiative "is designed to address that concern" by bringing independent institutional capital into the AI infrastructure market TechCrunch.

thinkidiot take: Nvidia wants AI hardware to behave like infrastructure assets with predictable residual value, much like railcars or aircraft. If it works, Nvidia effectively monetizes the whole lifecycle. If demand stalls, the guarantees become a massive contingent liability dressed up as financial engineering. The difference between the two outcomes comes down to whether enterprises actually use AI at scale long enough to burn through these assets before they age out.

04.Microsoft Merges Its Consumer and Business Copilot Apps Into One Interface

Microsoft is combining its consumer Copilot and commercial Microsoft 365 Copilot apps into a single application that recycles the "Microsoft Copilot" name with an updated icon The Verge. The unified app will accept both personal and work accounts, marking Microsoft's first major step toward a Copilot "super app."

Alongside the merge, Microsoft is discontinuing several AI features: AI-generated podcasts, Deep Research in OneNote, and its Mico mascot character from voice mode. Mico will move to the platform's Learn Live environment where it will have more interactions to react to The Verge.

The unification reflects a broader realization that AI assistants are now competing as general-purpose entry points into computing, not just add-on productivity tools. Microsoft is betting on being at the center of how people work across both personal and professional contexts in the same interface.

thinkidiot take: Merging Copilot into a single app makes operational sense but reveals uncertainty about what Copilot should actually be. A super app works when it becomes essential enough that you don't mind living inside it. Microsoft's track record suggests it will try hard, but legacy Windows habits die slowly.

Five ML Papers Worth Reading

Items below published or announced in the last 24 hours. This is an editorial selection, not a ranked list. Where items were verified within the prior day they are included; if the feed produced fewer than five credible submissions for this window, only what was found appears here.

• Conceptual Reasoning Index (Anthropic, published Aug 13). Anthropic released an evaluation focused on reasoning capabilities that go beyond standard benchmarks. See: https://alignment.anthropic.com/2026/conceptual-reasoning-index/
• DeepSeek V4 Pro Improved Release (DeepSeek, published Aug 13). DeepSeek shipped an improved version of its V4 Pro model out of testing and released its agent software, Harness v0.1, under the MIT license. See: https://the-decoder.com/deepseek-launches-an-improved-v4-pro-model-raises-api-prices-and-makes-its-agent-software-open-source/
• Grok 4.6 Scores 61 on the AI Intelligence Index (xAI, published Aug 12). xAI's latest model tied with GPT-5.6 on Artificial Analysis's benchmark. See: https://the-decoder.com/spacexais-grok-4-6-matches-openais-best-model-and-undercuts-it-on-price/
• Text AI Watermarks Are Trivial to Remove (Sean Gedecke, published Aug 13). A technical demonstration showing that current text watermarking methods can be stripped with minimal quality loss. See: https://www.seangoedecke.com/text-ai-watermarks/
• How AI Text Watermarking Works (Claude, published Aug 13). An explanation of the mechanisms behind AI-generated content watermarking and their limitations. See: https://declaude.org/watermarking/

Five Repos Worth Exploring

Recent repositories with notable momentum over the last 24 hours, selected editorially.

• Strands Agents, LeRobot, and HF Storage Buckets (Hugging Face + Amazon, published Aug 13). A collaboration showing how to record, train, and deploy robotics agents in a single workflow with streaming data loops Hugging Face.
• Amazon Quick for Microsoft 365 (AWS, published Aug 13). Agentic AI extensions inside Word, Excel, PowerPoint, and Outlook that connect to enterprise data sources without switching apps AWS Machine Learning.
• AgentCore Observability for Multi-Cloud Agents (AWS, published Aug 13). OpenTelemetry-based monitoring for AI agents running outside AWS, including on-prem and GCP deployments AWS Machine Learning.
• MCP Memory (GitHub repo fellowgeek/mcp-memory, Show HN Aug 13). Fast agent memory built on Google's OKF and SQLite FTS5 Hacker News.
• Harness v0.1 (DeepSeek, published Aug 13). MIT-licensed agent software accompanying V4 Pro release DeepSeek/The Decoder.