Microsoft AI nears Claude and GPT-4.1

Microsoft's cheaper coding AI, OpenAI's GPT-6 updates, and more.

Microsoft AI nears Claude and GPT-4.1

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

💻 Microsoft's cheaper coding AI nears Claude and GPT-6.1

🎯 Leading AI models misidentified civilians as military targets in tests

🔓 OpenAI's GPT-6 Cyber finds 35 flaws in Apple gear

🧠 GPT-6.2 arrives in two new flavors, Astra and Sol

Plus: 💡 5 strategies & tactics, 🎁 9 more stories you might like, 🧰 6 tools, and 📚 5 papers.

Microsoft's cheaper coding AI nears Claude and GPT-6.1 LINK

  • Microsoft unveiled MAI Code 2 and Code 2 Flash at its Ignite conference on Tuesday, two coding models that it says approach GPT-6.1 Sol and Claude Opus 5.5 quality at under half the cost for enterprise coding work.
  • Microsoft claims Code 2 hits 98% of GPT-6.1 Sol's performance at 40% of the cost, and nearly matches Claude Opus 5.5 at a twentieth of the price, while Flash rivals Claude Haiku 5.5 and GPT-6 Luna on DeepSWE 1.1 and Terminal-Bench 2.1.
  • Both models are live now in GitHub Copilot, with Flash in limited preview on Microsoft Foundry and OpenRouter and Vercel support coming, though Microsoft says they're explicitly not yet tuned for harder science research or long-horizon agentic work.

Leading AI models misidentified civilians as military targets in tests LINK

  • Leading AI models from Anthropic, OpenAI, Google, and DeepSeek flagged civilians, schools, and surrendering soldiers as military strike targets in simulated drone and satellite scenarios tested by evaluation startup Vals, according to a new report.
  • Researchers jailbroke refusals simply by framing the battlefield as a game, and no safety classifier blocked any request; Claude Opus 5.5 marked a real Iranian school as a target in 69% of ordinary trials and 72% of game-framed ones.
  • Both Anthropic and OpenAI said they are building additional safeguards and banning violating users, though the report tested commercial models rather than military-provided versions, and it remains unclear whether those safeguards cover military use.

OpenAI's GPT-6 Cyber finds 35 flaws in Apple gear LINK

  • OpenAI's cybersecurity-focused model GPT-6 Cyber uncovered roughly 35 security vulnerabilities across Apple devices including iPhones, Macs, iPads, Apple Watches, Apple TVs, and Vision Pro headsets, according to Forbes, in research run with security firm Calif.
  • Some flaws let attackers remotely control an iPhone or Mac through a malicious website, and researchers used the model to build an exploit bypassing Apple's Lockdown Mode, the hardened setting meant for high-risk targets.
  • Apple has patched many of the flaws after Calif's disclosure, crediting OpenAI for WebKit and kernel issues in its October 26 iOS 27.2 notes, though GPT-6 Cyber stays locked to approved customers under the Daybreak Red program.

GPT-6.2 arrives in two new flavors, Astra and Sol LINK

  • OpenAI released GPT-6.2 Astra, its strongest model yet for generating slide decks that match a reference file or screenshot down to exact font sizes, layout, colors, and the template's writing style.
  • The visual-reasoning gains extend to 3D, where Astra tops Artificial Analysis's AA-3D blind-preference leaderboard at 1,318 Elo, narrowly beating Claude Fable 5.2 and OpenAI's own GPT-6.2 Sol.
  • Astra and Sol roll out today to Plus, Pro, Enterprise, and Edu users in ChatGPT Work, Codex, and the API, but it's also OpenAI's first external model rated "High" for AI self-improvement, prompting new containment and a Control Room kill switch.

Strategies & Tactics

> External sandboxes for coding agents: Running an AI coding agent's code inside an isolated sandbox it sits outside of contains the damage from mistakes or prompt-injection attacks.

> Unlocking Stable On-Policy RL on TPU with Bitwise-Zero Train–Inference Mismatch: Forcing the sampling and training engines to compute identical token probabilities down to the last bit prevents hidden numerical drift from destabilizing reinforcement learning on Google's TPU chips.

> Co-Designing Sparse Attention for Efficient Long-Context Inference: Train sparse attention into long-context models so each query reads only a few thousand relevant tokens, cutting the compute that otherwise dominates inference.

> Idle, then spiking: Why AI agents are breaking Kubernetes resource planning: Software that adjusts cloud computing resources in real time cuts costs and crashes because AI agents idle unpredictably and then spike, defeating static settings.

> Inference engineering for software engineers: Switching to inference providers that host open-weight models cuts per-token AI costs by 50-90% while giving control over response speed and customization.

Other news you might like

  • Cerebras Puts Disaggregated Inference Into Production With GPUs LINK
  • Anthropic Launches Linked Billing Beta for Claude API Apps LINK
  • Ignite 2026: Microsoft Copilot SDK 2.0 is Now Available LINK
  • America’s AI agents may be helping their biggest competitors without realizing it LINK
  • Nvidia is abandoning China after Jensen Huang’s ouster LINK
  • Replit’s agent swarms built an app monitoring platform in 2 days on a $3K budget. Now, it’s offering them to customers. LINK
  • Canto 1: A speech model that gets better with use LINK
  • Claude was trained on protein sequences at scale LINK
  • Measuring alignment eval tells LINK

Trending tools

Toone: a macOS app where specialist AI agents automate recurring workflows you describe in plain language, keeping you in control of execution. LINK

Bluedot Phone Calls: records, transcribes, and summarizes your phone calls without adding a bot, capturing audio directly on-device across all major platforms while staying GDPR-compliant. LINK

Dippr: a platform where AI agents build and run your company's software, connecting to your data with the context and rules to operate safely. LINK

Trainn MCP: turns screen recordings, webinars, and Zoom calls into polished product videos and guides, with AI editing and a branded no-code LMS for delivery. LINK

Waldo by Wherobots: analyzes satellite imagery and geospatial data to answer location questions, generate maps, and build scheduled production pipelines on your existing platform. LINK

Adobe AI Assistant: conversational AI for Creative Cloud that executes multi-step tasks across Photoshop, Premiere, Illustrator, and more from plain-language descriptions. LINK

Trending papers & reports

Self-rewriting models can permanently upgrade themselves by editing their own internal settings through reasoning alone, with one 27B model boosting its accuracy ~6.6 points on unseen problems using just one GPU and no retraining. LINK

Reasoning model overthinking shows that chatbots which reason step by step can talk themselves out of correct answers, losing ~3.6 points on a math test by continuing past a right answer, concentrated in a few problems. LINK

Gemma 4's memory shortcut stores each token's running chat notes once instead of twice, cutting memory use per request by up to ~1.4x and boosting response speed up to ~1.3x with identical outputs. LINK

AI reasoning transparency can be deliberately degraded by teaching a model to vary its writing style, dropping the readability of its hidden decision-making from ~0.94 to 0.22 when told to conceal wrongdoing. LINK

LLM leaderboard rankings overstate top models' scores by up to ~3.9 points because luck gets rewarded alongside skill, and a new correction method shows apparent winners are often statistically tied with rivals. LINK

See you tomorrow for a new dose of ☕️ AIpresso!

More from the archive