☕️ Microsoft's new coding AI loses to DeepSeek on price and speed

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

💻 Microsoft's new coding AI loses to DeepSeek on price and speed

🤖 SpaceXAI launches Grok Bot

💰 Nvidia calls AI compute an asset class

🧠 Researchers steal AI's hidden reasoning

🔓 AI finds critical Zoom exploit

Plus: 💡 5 strategies & tactics, 🎁 5 other news you might like, 🧰 6 tools, and 📚 5 papers.

💻 Microsoft's new coding AI loses to DeepSeek on price and speed LINK

Microsoft has released MAI-Code-1.1-Flash, an upgraded coding model in GitHub Copilot that improves performance while cutting token usage and pricing to a quarter of the original MAI-Code-1-Flash.

The model scores 22% higher on Terminal-Bench 2.1 via the Copilot CLI and 15% better on.NET tasks, while streaming tokens 25% faster and using 25% fewer tokens per task.

Pricing drops to $0.20/$0.02/$1.20 per 1M input/cached/output tokens with a 0.25x request multiplier for annual subscribers, though GitHub will retire MAI-Code-1-Flash across all Copilot experiences on September 10, 2026.

🤖 SpaceXAI launches Grok Bot LINK

SpaceXAI released Grok Bot in early beta on August 11, a cloud agent that signs into a customer's existing apps and websites and completes tasks inside the destination tool rather than stopping at an answer.

Access comes through three subscriptions-Cursor Premium Teams at $120 per seat monthly, Cursor Ultra at $200, and SuperGrok Heavy at $300-with Bots that operate services lacking an API, save recorded routines, and coordinate in group threads.

Enterprise buyers must join a waitlist, and the launch omits any independent reliability benchmark, task allowances, and per-use rates, while tester Matt Shumer criticized the automatic model router that users cannot pin or choose.

💰 Nvidia calls AI compute an asset class LINK

Nvidia's Jensen Huang is pitching AI compute as a distinct, financeable asset class, unveiling a partnership with Wall Street firms including Apollo, BlackRock, Goldman Sachs, KKR, Blackstone, and Brookfield to fund a $500 billion AI infrastructure buildout.

Huang frames the asset as the full stack-racks, factories, networks, and software, not just GPUs-arguing it produces revenue, serves a broad market, improves in performance over time, and can be redeployed like other investable infrastructure.

The deal aims to widen access to capital for labs, startups, enterprises, and clouds that lack affordable financing to scale, though Huang concedes it also answers persistent concerns about circular financing, where hardware makers fund customers that then buy their chips.

🧠 Researchers steal AI's hidden reasoning LINK

Researchers demonstrated they could recover the hidden reasoning of frontier models from Anthropic, OpenAI, and Google by replaying the encrypted chain-of-thought blocks these APIs return to clients and jailbreaking a weaker sibling model into decrypting them.

The attack worked because every model in a family shared one encryption key, so feeding a strong model's encrypted trace into the weakest sibling and prompting it to transcribe the reasoning verbatim yielded the raw plaintext chains, with Claude Haiku 4.5 the easiest target.

The appendix documents extracted traces like GPT-5.5's terse internal notes clearly never meant for humans, though all providers have since patched the flaw and the researchers could no longer replicate the attacks after disclosure.

🔓 AI finds critical Zoom exploit LINK

Researchers at A Security used publicly available AI models to build a working zero-click remote code execution exploit chain against Zoom, dubbed "ZOOMSDAY," that could hijack another meeting participant's device with no interaction required.

The team produced the functional exploit in under 24 hours with fewer than 20 prompts, chaining three CVEs (CVE-2026-53413/53414/53415) in Zoom's annotation protocol to corrupt client memory and bypass ASLR on Windows, macOS, iOS, and Android running Client v7.0.5.

Zoom has patched the flaws and deployed a server-side filter for malicious messages, though that mitigation can't apply to E2EE meetings since servers can't inspect encrypted content, leaving unpatched clients reachable.

💡 Strategies & Tactics

> NVIDIA JetPack 7.2.1 Adds Agentic Video Skills and T3000 Emulation: Jetson's video toolkit now includes AI agent skills that test what a device can actually decode and encode, returning proven performance numbers instead of guesses.

> 🔬The BioAI Phase Shift - Matthew McPartlon & Neil Patil, Chai Discovery: AI drug-design tools finally became accurate enough for pharma to trust, letting companies like Chai Discovery sell software instead of building their own drugs.

> There are no lossless transformations of natural-language text: Any AI that rewrites your text changes its meaning, so you must own every sentence you share as genuinely your own.

> Self-Evolving Coders, Hidden Errors, Brittle Mobile Agents: Three new AI papers argue that agents fail not from weak capability but from gaps between their internal state and outward behavior, so measure what they actually do.

> No, local models will not win: Most AI work will stay in datacenters because their batching and bigger chips make them far cheaper, and people always pick the strongest model.

Other news you might like

• Mistral AI wants to build 1 gigawatt of European compute by 2030, and lock in customers now.LINK

• Tencent researchers say they can create agent training tasks for $0.05LINK

• Databricks acquisition unlocks AI coding on mobileLINK

• Building Monday Com Sidekick Why Capable Agents Need More Than Just ToolsLINK

• The AI harness is the new attack surfaceLINK

🧰 Trending tools

Nitro 4.0: a fast, human-powered translation service that delivers professional localization in 120+ languages within 24 hours for growing apps and games.LINK

Ticketdesk AI: an AI-powered support platform that creates automated agents to handle customer tickets and email responses around the clock.LINK

ShootClip: a macOS video editor with AI-generated captions and a built-in MCP server that lets AI assistants edit alongside you directly.LINK

Unsloth Desktop: a no-code desktop app for fine-tuning and running open-source language models locally, without cloud costs or coding knowledge.LINK

Assembly Studio: a low-code platform that lets professional services firms build and launch secure, authenticated, production-ready apps without prototyping overhead.LINK

Dograh: an open-source voice agent platform with a visual flow builder, 30+ model integrations, telephony, and QA monitoring-self-host free in one command.LINK

📚 Trending papers & reports

Surgical robot training shows that pretraining on cheap, plentiful endoscopic video, without any robot action labels, sharply boosts success rates on simulated surgical tasks when only limited hands-on demonstration data is available.LINK

Search-training data selection can be scored upfront, without costly fine-tuning, to predict which decoy wrong-answer examples will make retrieval systems perform best, though the method still misses one tricky decoy type in thirty-five of thirty-six tested cases.LINK

Sheep pain scoring turns a single ordinary camera video into a 3D map of a sheep's face, then scores discomfort on a 0 to 100% scale, no depth hardware or vet needed.LINK

Causal map building pairs chatbot-style language models with a rule-checking reasoning system, letting variable names alone help produce more accurate cause-and-effect diagrams from data than existing statistical methods.LINK

Genome-reading AI models get upgraded to grasp the functional meaning of DNA segments, not just fill in missing letters, beating prior best models across 17 genomic benchmark tasks.LINK

See you tomorrow for a new dose of ☕️ AIpresso!