- NATURAL 20
- Posts
- ChatGPT Work Launches with GPT-5.6 for Agentic Workflows
ChatGPT Work Launches with GPT-5.6 for Agentic Workflows
PLUS: Reflect' Dashboard Debuts to Track Your Claude Usage, Seedream 5.0 Pro Upgrades Complex Image Generation and more.
Wes Roth
July 10, 2026
In partnership with
[Webinar] 8 levels of context maturity in AI-native engineering
AI is in your engineering workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem.
Join live July 23 (FREE) to learn:
The 8 levels of context maturity: where most teams are stuck and what the ceiling looks like at each stage
Why more MCPs, rules, and skills provide agents access but not understanding
How leading engineering teams are using a context engine to make the most of their agents
Today:
ChatGPT Work Launches with GPT-5.6 for Agentic Workflows
Grok 4.5 Drops as xAI's Smartest Model for Coding and Agents
Muse Spark 1.1 Advances Multimodal Computer Use
Reflect' Dashboard Debuts to Track Your Claude Usage
Seedream 5.0 Pro Upgrades Complex Image Generation
ChatGPT Work is a new agentic workspace powered by GPT-5.6. It brings together information from workplace tools, files, and desktop applications to turn scattered notes, research, and data into completed projects.
Instead of asking users to assemble everything manually, ChatGPT can gather context, plan the work, and produce polished spreadsheets, presentations, documents, dashboards, and reports. It can also follow existing templates and preferred formats.
Users can create interactive websites and web apps through Sites, including project trackers, launch calendars, prototypes, and live reports. More than 1,400 plugins allow ChatGPT to work with existing business tools and workflows.
The desktop app also includes a built-in browser with multiple tabs. This allows ChatGPT to work across websites, files, tools, and accounts during longer assignments.
Plan mode gives users more control by presenting a step-by-step approach for approval before work begins. One-time and recurring tasks can also be scheduled, monitored, and reviewed from a phone.
ChatGPT Work is available across all desktop plans. Web and mobile access is rolling out to Plus, Pro, Business, Enterprise, and Edu users.
Grok 4.5 has been released as SpaceXAI’s most advanced model for coding, autonomous tasks, and professional knowledge work. It was trained alongside Cursor and is designed to complete difficult engineering projects using fewer steps.
The model scored 62% on DeepSWE 1.0, 83.3% on Terminal Bench 2.1, and 64.7% on SWE Bench Pro. It led the listed models on SWE Marathon with a 29% resolution rate, although it did not finish first on every benchmark.
Grok 4.5 was trained using tens of thousands of NVIDIA GB300 GPUs. Its training covered hundreds of thousands of tasks focused on multi-step software engineering, science, mathematics, and other technical work.
Speed and efficiency are major selling points. The model runs at approximately 80 tokens per second. On SWE Bench Pro, it generated an average of 15,954 output tokens per task—about 4.2 times fewer than Claude Opus 4.8 in the published comparison.
The model can fix bugs, work with Rust and C++, and build complete applications from a single prompt. It is also the default model in Grok Build, where it can create multi-sheet Excel models, design PowerPoint presentations, and produce structured Word documents.
API access costs $2 per million input tokens and $6 per million output tokens. Grok 4.5 is available through Grok Build, Cursor, and the SpaceXAI API, with limited-time free access in Grok Build and Cursor. European Union availability is expected in mid-July.
Muse Spark 1.1 is a new multimodal reasoning model built for autonomous workflows, coding, computer control, and visual understanding. It is also the first Muse Spark model available to developers through the new Meta Model API.
The model can gather context, create a plan, and divide projects among several AI subagents working in parallel. Each subagent can follow a specific assignment and return control when it needs help or additional direction.
Muse Spark 1.1 supports a one-million-token context window. This allows it to remember earlier actions, retrieve details from long projects, and compress older information without losing the steps needed later.
Its computer-use abilities combine direct interface control with automation. The model can decide when to click through an application, write a script, or perform several actions together.
For software development, Muse Spark 1.1 can diagnose complex bugs, implement features in large codebases, perform software migrations, and use screenshots to find and validate visual problems.
The model can also understand images, video, audio, text, and PDFs. One demonstration showed it turning a smartphone product video into photos and information for a Facebook Marketplace listing, then operating the browser to create the listing.
Safety testing found stronger resistance to jailbreaks, prompt injection, and malicious instructions hidden in outside data. The unrestricted model showed higher chemical, biological, and cybersecurity risks, but the deployed version was rated “moderate or lower” after safeguards were applied.
Muse Spark 1.1 is available in Thinking mode through Meta AI and Meta.ai. The Meta Model API is currently available as a public preview.
🧠RESEARCH
Google introduced Gemma 4, a family of openly available AI models that understand text, images, and audio. Sizes range from 2.3 billion to 31 billion settings, with a thinking mode for harder problems. Tests show gains in mathematics, coding, vision, and long documents while using less memory and computing power.
DeepSeek created DSpark, a system that makes AI responses faster by letting a smaller model draft several words before the main model checks them. It adjusts how much text gets checked based on confidence and server demand. On DeepSeek-V4, it increased writing speed by 60% to 85% without lowering quality.
Researchers introduced SciReasoner, an AI model that studies three-dimensional structures of proteins, chemicals, and crystals, then explains how those shapes affect behavior. It achieved the best results on 67 of 86 tests, improved protein and chemical predictions, and explanations matched or surpassed top AI answers in 98% of reviewed cases.
📲SOCIAL MEDIA
🗞️MORE NEWS
Claude’s new Reflect dashboard shows which topics users discuss, what tasks they hand to AI, and when they use it most across one, three, six, or twelve months. The early test version also offers break reminders and quiet hours, and is available to Free, Pro, and Max users with Memory enabled.
Seedream 5.0 Pro can create and edit realistic photos, text-heavy infographics, storyboards, websites, and other professional visuals. It follows handwritten notes and sketches for precise edits, separates images into editable layers, and supports around 12 languages.
GPT-Live lets ChatGPT listen and speak simultaneously, allowing natural interruptions, pauses, short acknowledgments, and live translation. It can send harder questions to GPT-5.5 in the background while continuing the conversation, and is rolling out globally to Free and paid users, with developer access coming later.
AI 2040: Plan A presents a recommended path—not a prediction—for delaying AI that surpasses humans across most fields until 2040. It proposes an international agreement, public AI research, close tracking of powerful computer chips, and a verified pause at top-human-level AI until safety and human control improve.
Cursor is reportedly developing Sand, a personalized AI assistant that can respond to emails and texts, organize spreadsheets, and handle engineering work. Employees began testing it internally in late June, but no public release date has been confirmed.
NEO’s new robotic hands have 25 independently controlled movements, force-sensing joints, and touch-sensitive skin for handling delicate and heavy objects. They are waterproof, food-safe, and designed for tasks such as using tools, pouring drinks, picking up coins, and plugging in cables, with production capacity planned for 10,000 hands this year.
Meta is reportedly developing “super sensing” glasses that continuously capture audio and take photos every few seconds, allowing its AI to answer questions about what the wearer has seen and heard. Raw recordings may not be stored, but the recording light could remain off and collected information might train AI models, creating serious privacy concerns.
OpenAI audited SWE-Bench Pro, a widely used test for comparing coding models, and estimated that about 30% of its tasks are flawed because instructions and grading tests often conflict. It withdrew its earlier recommendation after AI investigators and five experienced software engineers reviewed the flagged tasks.
Robostral Navigate follows plain-language directions and moves wheeled, legged, or flying robots using one ordinary camera—without laser or depth sensors. It completed 76.6% of previously unseen route tests, beating systems with more sensors after training on 400,000 virtual trips across 6,000 environments.
What'd you think of today's edition? |
Reply