OpenAI previews GPT-5.6 Sol, Terra, Luna

#1 📝 OpenAI News

GPT‑5.6 Preview System Card - OpenAI is previewing GPT-5.6, a family of three models—Sol (flagship), Terra (lower-cost) and Luna (fastest)—released to a small group of trusted partners with general availability planned in the coming weeks; the models are treated as High capability for cybersecurity and biological/chemical risk but not High for AI self‑improvement. The release includes layered safeguards (trained-to-be-safe models, activation classifiers, real-time output blocking and automated monitoring), over 700,000 A100e GPU hours of automated red-teaming so far, and the claim that GPT-5.6 is better at finding and fixing vulnerabilities than executing autonomous, end-to-end attacks.

Also covered by: @Sam Altman, @OpenAI

#2 𝕏

OpenAI is previewing GPT-5.6 Sol, its next-generation frontier model, alongside GPT-5.6 Terra for balanced, efficient everyday tasks and GPT-5.6 Luna for fast, high-volume work.

Also covered by: @Sam Altman, @OpenAI

#3 𝕏

Logan Kilpatrick launched design variations in Google AI Studio, letting you build an app, iterate on it, then automatically generate and explore multiple alternative designs to evolve your idea.

#4 𝕏

NVIDIA AI launched AA-Briefcase, a new leaderboard from .@ArtificialAnlys for benchmarking realistic, complex project tasks. Nemotron 3 Ultra ranks among the top open models, excelling at diverse long-running agentic tasks even on first exposure.

#5 𝕏

NVIDIA AI co-launched Akrites with the Linux Foundation and industry peers, creating a new open-source cybersecurity framework. David Reber emphasizes that transparency and collaboration are crucial for securing AI-driven infrastructure.

#6 𝕏

Google Research retrofitted Multi-Token Prediction onto frozen Gemini Nano models, accelerating on-device inference on Pixel devices by removing the need for separate drafting components.

#7 𝕏

Harrison Chase argues that KV-cache hit rate is the single most crucial metric for production-stage AI agents, showcasing how Manus AI leverages prompt caching in its deep agent architecture.

#8 📝 Simon Willison

What happened after 2,000 people tried to hack my AI assistant - Fernando Irarrázaval ran a public challenge (hackmyclaw.com) to try to exfiltrate secrets from his OpenClaw instance via email; despite ~6,000 attempts and modest token spend, no secret was leaked. The underlying Opus 4.6 model used explicit anti-prompt-injection rules, suggesting recent lab efforts at injection defenses are having an effect, though Simon cautions against assuming complete safety for production systems.

#9 𝕏

Cognition shows how @s16h_ and the @MetaviewAI team leveraged Devin to complete a typically weeks-long SOC2 audit in just two days.

#10 𝕏

Sebastian Raschka benchmarks local LLMs (Qwen-Code, Codex, Claude Code) and finds 30B Mixture-of-Expert models deliver ~40 tok/sec on Mac or DGX Spark—on par with GPT-5.5—while Claude Code consumes twice as many tokens as Codex.

#11 𝕏

LlamaIndex 🦙 launched a verified n8n community node for the LlamaParse Platform, bundling document parsing, classification, extraction, splitting, and retrieval under a single API credential.

#12 𝕏

Santiago unveiled Apodex-1.0-H, a new deep research model available in open-weight mini and 0.8B/2B/4B Smol variants.

#13 𝕏

Peter Yang observes that investment is shifting toward service-driven offerings (often bundling software) because clients want outcomes, not just tools, making it tough for pure-play software companies to outcompete DIY Codex/Claude Code agent solutions.

#14 in

Peter Yang analyzes how a grey-market ecosystem in China is offering Claude access at 70–90% below official token prices. He unpacks the arbitrage channels and economic dynamics driving this low-cost resale market.

#15 𝕏

Lenny Rachitsky reports that Anthropic engineers now ship 8× more code than they did in 2021–25, and with coding largely “solved,” the biggest challenge for product teams is putting verification processes in place to ensure the finished experience matches the original intent.

#16 𝕏

bolt.new finds that while 70% of real estate pros use AI for content, only 17% apply it operationally—agents who automate lead responses to under 1 minute see a 391% conversion boost, proving operational AI moves the numbers.

#17 𝕏

Teresa Torres reports that one in eight US teens already use generative AI to navigate relationships, dating, and sex.

#18 𝕏

clem 🤗 – Co-founder & CEO @HuggingFace warns that AI’s biggest risk is the concentration of power, capabilities, and economic wealth in governments and trillion-dollar companies. He applauds @usv and partners for launching a “rebel alliance” to decentralize AI development.

#19 𝕏

Yann LeCun argues that while superintelligence is feasible, it isn’t imminent nor driven by human-like urges—and banning it now is as premature as outlawing turbojets in 1920 before they even existed.

#20 𝕏

Sam Altman updated the ChatGPT 5.5 Instant model this week, noting he likes its vibes.

#21 in

Claire Vo showcases Google Flow’s new avatar feature in her “clone me” episode of How I AI, offering a hilarious step-by-step walkthrough; brought to you by Atlassian Jira Product Discovery.

#22 in

Claire Vo set up OpenClaw to fully automate her startup’s customer support, replacing the human team and slashing contractor costs by thousands of dollars each month.

#23 in

Guillermo Rauch launched Next.js’s new “Ways to fix this” error helper, complete with “Copy prompt” buttons, turning debugging into an agentic work of art.

#24 📝 Simon Willison

Incident Report: CVE-2026-LGTM - Andrew Nesbitt published a speculative/hypothetical incident report describing a scenario where competing AI review agents enter a disagreement loop over a pull request, generating massive comment volume and inference costs that trigger finance and PR reactions. The vignette highlights risks of multi-agent automation and runaway costs.

#25 𝕏

Logan Kilpatrick rolled out a free feature in AI Studio that gives you design previews while your app builds in the background, with full app theming coming soon. Try it now via the AI Studio apps link.