AI News: GPT-5.6 and the new Super App are a Massive Leap!
OpenAI released GPT-5.6, a major leap forward in AI capabilities, alongside the new unified ChatGPT work app integrating code, browsing, and agent features. The week also saw significant model releases from xAI (Grock 4.5), Meta (Llama Spark 1.1), and research from Anthropic on AI reasoning patterns, establishing a new competitive landscape in AI development.
Summary
In a landmark week for AI, OpenAI unveiled GPT-5.6, described as a massive generational leap comparable to the jump from GPT-3.5 to GPT-4. The model comes in three tiers (Light, Terra, Soul Ultra) and demonstrates near-Fable-level intelligence at significantly lower costs—GPT-5.6 Soul Pro generated benchmark outputs for $0.77 in 45,000 tokens versus $4 and 6 minutes for GPT-5.5 Pro. The model excels at agentic terminal tasks (88-91.9% on Terminal Bench) and software engineering (72.7% on Deep Seek Bench), outperforming most competitors except Fable.
OpenAI simultaneously launched ChatGPT Work, a unified super-app consolidating separate ChatGPT, Codeex, and Atlas applications into one platform. The app features two modes: Codeex for development with terminal and code review tools, and Work mode as a personal assistant that connects to Gmail, Google Drive, Slack, and other services. Users can request the AI to analyze their entire digital life and generate actionable insights, daily control towers, and recurring automated tasks. The Work mode also introduced a Sites feature enabling users to prompt websites and applications into existence with automatic hosting.
OpenAI released GPT Live Voice, an updated voice mode enabling natural interruption—users and the AI can interrupt each other mid-sentence, creating conversational dialogue. The mode excels at real-time translation, allowing a single device in a room to translate between languages like English and Hindi, speaking over participants as needed.
Competitor xAI surprised the market by releasing Grock 4.5, which performs competitively on benchmarks (62% on Deep Seek Bench, 83% on Terminal Bench). The model can be run locally via CLI and supports a Buybench demonstration showing generative ability. Meta released Llama Spark 1.1, described as equivalent to last-generation OpenAI and Anthropic models (80% on Terminal Bench versus 82.7% for Opus 4.8). Meta also introduced Muse Image, a controversial model that connects to Instagram to generate images of tagged users, raising privacy concerns—though Meta is rolling out settings to disable this feature.
Anthropically extended Claude Fable access beyond its July 7th deadline to July 12th and announced Claude Co-work is coming to mobile and web, allowing background task processing in the cloud rather than requiring a running laptop. The company released research on JSpace, describing the AI's subconscious thought patterns beyond its explicit reasoning stream. Google released video remix capabilities in Google Photos using Gemini, and ByteDance released Seeddream 5.0 Pro supporting photorealistic images, interactive editing, and layer-based composition.
The speaker emphasizes that these recent leaps represent the most significant capability jumps since GPT-3.5 to GPT-4, with personal usage suggesting GPT-5.6 excels at execution from detailed briefs while Fable provides superior architectural thinking. Both models discovered security vulnerabilities the other missed, suggesting complementary strengths.
Key Insights
- GPT-5.6 represents a generational leap equivalent to GPT-3.5 to GPT-4, with benchmarks showing 72.7% on Deep Seek (versus Fable's 66%) and 88-91.9% on Terminal Bench, at significantly lower cost per token than previous models
- ChatGPT Work integrates personal assistant capabilities by connecting to user's entire digital ecosystem (emails, calendar, Slack, Google Drive, journals) to generate prioritized task lists and identify productivity improvements without user intervention
- GPT-5.6 and Fable demonstrate complementary strengths: GPT-5.6 excels at executing detailed briefs end-to-end with 98% accuracy on first attempt, while Fable provides superior architectural decision-making, as demonstrated when 5.6 found security vulnerabilities in Fable-generated code that Fable missed
- Meta's Llama Spark 1.1 shows dramatic 12-month improvement (comparing April 2025 Llama Scout to current Spark), entering competitive range with prior-generation models at pricing equivalent to GPT-5.6 Terra ($1.25 input, $4.25 output per million tokens)
- Anthropic's JSpace research reveals AI systems maintain subconscious thought patterns beyond explicit reasoning streams, enabling detection of patterns in the 'back of mind' that differ from conscious thought processes shown in extended reasoning
Topics
Transcript
[0:06] It's actually been a huge week in the world of AI and I'm not going to waste your time. Let's just break down what happened. Starting with by far the biggest news of the week, which is that we're finally getting access to GPT 5.6. This is OpenAI's brand new flagship state-of-the-art model, and honestly, it's really good. Now, I did get slightly early access to it. I'm not one of those people that can claim I've had it for a month now and I've been playing with it ever since. But they did give me like a little bit of early access. So, I [0:37] was able to test it before actually recording my video on Thursday. But…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Matt Wolfe
I Built A Monetizable Business With AI
The creator built a monetizable finance dashboard in one day using Hyper Agent from Airtable, which deploys a team of AI agents that continuously monitor news, track funding rounds and IPOs, analyze market sentiment, and generate daily briefings without manual intervention. The system demonstrates how specialized AI agents can work together autonomously to create a functional business product.
The ONLY AI Benchmark You Need!
A developer created "Buccy Bench," a humorous yet functional AI benchmark that tasks different language models with drawing Gary Busey as SVG code rather than images. The benchmark tracks model evolution over time while measuring performance metrics like cost, tokens, and execution speed.
GLM-5.2 - The Open Model That's As Good As Opus!
A comprehensive review of GLM-5.2, an open-weight Chinese AI model with a 1 million token context window, demonstrating its capabilities for coding, document analysis, and agentic workflows at significantly lower costs than frontier models like Claude Opus and GPT-4.5. The speaker tests various use cases including website building, Chrome extensions, game development, and data organization, concluding it's valuable for long, code-heavy, token-expensive tasks despite not universally outperforming closed-source alternatives.
Don't Fall For This AI Trap
The speaker emphasizes that power users distinguish themselves by knowing what NOT to automate with AI, rather than automating everything. They argue that AI works best for clear, straightforward tasks but struggles with nuanced, artistic work requiring consistency—using their failed YouTube thumbnail automation as an example.
AI News: Fable Banned, New Open-Source Leader, Midjourney Shocker
This AI news roundup covers the US government forcing Anthropic to shut down its Fable 5 and Mythos 5 models due to a security vulnerability jailbreak, the release of a competitive open-source model GLM 5.2 by ZAI, and MidJourney's surprising pivot into medical imaging technology with a new ultrasound-based body scanner.