OpenAI Launches GPT-5.6 Models and Aims to Conquer the Office Desktop with ChatGPT Work

GPT-5.6 Is Here and Sol Benchmarks Show Fable 5 Scores

On Thursday, OpenAI released GPT-5.6 Sol, its most powerful model to date, along with ChatGPT Work, an agent that independently operates spreadsheets, calendars, and email services. In standard benchmarks, Sol is roughly on par with Anthropic's top model Fable 5, according to Vals AI, and for real-world finance and legal tasks, Vals CEO Rayan Krishnan calls it state of the art. The release was delayed: The Trump administration had initially treated the new models from OpenAI and Anthropic as a national security risk, but allowed their public release after lengthy discussions. OpenAI is taking a more open approach than Anthropic, deliberately incorporating fewer guardrails so that companies can use the technology for their own defense. Anthropic, on the other hand, took its stronger model Mythos completely offline and blocked all foreign users by government order. Sol is expensive, in some cases even more so than Fable, which already costs about twice as much as Anthropic's next less powerful model. Meta released a weaker but significantly cheaper model on the same day. → www.nytimes.com

Synthszr Take: The real product decision here isn't in the benchmarks, but in the guardrails. OpenAI sells the thinner guardrails as a feature: Anyone who wants to find vulnerabilities in their own code needs a model that is allowed to name them. Anthropic has built the opposite with Fable, blocking sensitive answers and offloading them to the weaker Opus 4.8. This sounds cautious, but it blocks the very companies that want to secure themselves, not just the fraudsters. Anyone setting up SAST scans and security gates in their company this week should check which model is even allowed to respond when a gap in the authentication path appears. A day of setup versus months of damage—that calculation hasn't changed. Sol is expensive and not without risk, but a model that stays silent when it comes to defense is ultimately the worse choice.

ChatGPT Work Is a Frontal Assault on Claude Desktop

With the GPT-5.6 launch on Thursday, OpenAI not only showcased new models but also revamped its product setup. ChatGPT Work is the name of the new agentic tool for knowledge workers, built on Codex technology and in direct competition with Anthropic's Claude Cowork. Simultaneously, the previously separate ChatGPT and Codex desktop apps are merging into a single application, just as Anthropic did with its Claude app. ChatGPT Work starts tasks from documents and spreadsheets, pulls data from Slack, Teams, Google Drive, SharePoint, and CRM systems, and runs on web, mobile, and desktop. Scheduled tasks run in the cloud, so the laptop doesn't need to stay open. Notably, the classic chat window has moved to the sidebar; Work and Codex are now the two main modes. And the recently launched Atlas browser is being discontinued, replaced by a Chrome extension that brings ChatGPT directly into the Chrome sidebar. → thenewstack.io

Synthszr Take: The most interesting sentence in the entire announcement is almost an aside: Chat is no longer one of the two main modes. The conversational format that made ChatGPT visible to millions three years ago now sits in the sidebar next to scheduled tasks. The real engine is agentic work, and that came from Codex, not the chatbot. OpenAI is quite openly copying Anthropic's Cowork thesis. Both realized that their coding tools were long being used for non-coding work. The fact that Atlas is already dead after just a few months (we wrote about the Codex super-app transformation here in April, and now the standalone browser is history) shows the velocity of this market and the willingness to cannibalize one's own products. Anyone looking to test whether their knowledge work is ready for agents this week has two almost identical offerings in Work and Cowork for a direct comparison. The race is no longer decided by the model, but by who can embed themselves deepest into Slack, Drive, and calendars.

Meta Muse Spark 1.1: The Code Race Heats Up

Meta is opening up its AI to developers. After re-entering the race with its first in-house Muse Spark model in April 2026, the company is now releasing Muse Spark 1.1 along with a new Meta Model API. Meta calls it a “step-change”: better coding, detection and fixing of complex bugs, end-to-end agentic workflows across multiple apps, and native multimodality across images, videos, and documents. The model is now running in Thinking mode in the Meta AI app and on the website, with the API launching as a public preview for US developers. Every new account gets $20 in free credits. The launch follows directly on the heels of Muse Image, an image generation model that is causing trouble because it incorporates other users' Instagram content into its results. For Meta, it's about justifying the billions that have flowed into catch-up mode after prominent new hires and the restructuring of the last year, with the goal of achieving parity with OpenAI, Google, and Anthropic. → www.theverge.com

Synthszr Take: The $20 in free credits is an admission that no one is voluntarily asking for Meta's coding agent yet. Anthropic has occupied the command line with Claude Code, Cursor the IDE experience, OpenAI Codex the hard refactorings, and now Meta arrives with a second model and a fresh API in public preview for US developers. The real moat lies elsewhere: Whoever owns the developer workflow, owns adoption, and Meta is still on the outside looking in. The fact that Muse Image is simultaneously leaking by mangling other people's Instagram posts into its generations isn't just background noise; it raises the question of whether you want to trust this API with your codebase at all. Meta can afford this because its distribution network of WhatsApp, Instagram, and the smart glasses division gives its models a sales channel that no pure-play model provider has. Anyone setting up a 2-tool standard today should test Muse Spark 1.1 on the free credits against their existing stack and let the benchmarks decide, not the press release. A third serious provider pushes down prices, and that's a good day for any engineering team.

Perplexity Builds 'Teammate' for the Lucrative Coding Market

Perplexity, known as an AI search engine competitor to Google, has developed an internal coding tool codenamed “Teammate,” according to Business Insider, which has been used by its own engineers since May and could launch publicly later. The tool is designed to take over software projects end-to-end: owning projects, hunting bugs, monitoring services, according to the internal announcement. The key difference from Claude Code or OpenAI Codex: Teammate remains model-agnostic and isn't tied to any single chatbot. CTO Denis Yarats set the direction, urging his engineers to “stop looking at code” by the end of the year and just use the AI. The company reached a $20 billion valuation in its last funding round, giving it room to experiment. The report is currently based on a single source, and Perplexity has not commented or confirmed a launch. → The Next Web

Synthszr Take: The interesting move here is its model-agnostic approach. While Anthropic ties Claude Code to its own models and OpenAI pushes for workspace lock-in, Perplexity is building a service layer that can run on any model. This is the BYOK logic of Cline and Aider, scaled up to a $20 billion level and with enterprise ambition. Coding agents are one of the few areas where AI is actually making money today, which is why a search company is now pushing into this field. At Anthropic, 80 percent of the codebase now comes from Claude Code; the direction has been clear for a while. Anyone setting up their coding toolchain in 2026 should think in a model-agnostic way instead of chaining themselves to a single provider. A two-pronged standard of an IDE tool and a background agent, open to model changes, is the sensible bet, regardless of whether Teammate ultimately delivers.

Claude Launches Reflection Dashboard

Anthropic has launched a new beta feature for Claude that visualizes one's own usage. A reflection dashboard in the settings shows which topics you use Claude for and at what times of day, looking back over 1, 3, 6, or 12 months. It also includes quiet times, break reminders, and questions like, “What do you want to continue doing yourself, even if Claude could do it faster?” The whole thing is based on the 4D AI Fluency Framework: Delegation, Description, Discernment, and Diligence—which addresses what you delegate, how precisely you describe tasks, how critically you review the results, and who bears the responsibility. Anthropic draws a line when it comes to data privacy: Incognito chats, source files, and anything related to health integrations are excluded. The feature was developed with the MIT Media Lab (AHA program), the Digital Wellness Lab at Boston Children's Hospital, and the Family Online Safety Institute.

Synthszr Take: An AI provider suggesting you take a break is about as credible as a casino with a built-in exit counseling service. Nevertheless, there's a clever move here. Anthropic is no longer selling usage minutes, but usage competency, and the 4D framework is the real core of this. Anyone who learns to delegate properly and critically review results—Discernment and Diligence are the points where most teams are currently failing—remains the master of their own mind instead of a supplier for the autopilot. This week, you can test this without a major project: take an honest look at which tasks you've delegated that you should have thought through yourself. The suspicion that user retention is also a consideration remains, because a tool that knows your behavior also knows your weaknesses. But the ability to work with AI without outsourcing your own judgment is the skill that will be decisive over the next few years, and this feature at least makes it visible.

Seedream 5.0 Pro Can Do What Photoshop Can't

ByteDance Seed has introduced Seedream 5.0 Pro and is deliberately calling it Creation, not Generation: “Beyond Generation, It Understands Design.” The model independently builds logical layouts for infographics, UI mockups, and posters, packing a timeline, bar, pie, and line charts plus a real image into a single image hierarchy. The real leverage lies in Interactive Precision Editing: A finished poster can be broken down into over ten independent layers via text command, hidden backgrounds are automatically reconstructed via inpainting, and objects like a parrot can be swapped for a peacock. In practice, the AI delivers an editable PSD file instead of a flat image. It also features physically correct textures (reflection, refraction, skin structure) and native support for more than ten languages, including Arabic right-to-left script and culturally adapted architecture and clothing. With this, ByteDance shifts the creative's work from prompt writing to directing intent. → Trendium.ai

Synthszr Take: The breakthrough isn't a prettier image, but its decomposability. Until now, the biggest weakness of generative images was that the mock-up looked nice but was useless in actual production because you couldn't edit it. That wall is now coming down: When a result can be broken down into ten layers and the hidden areas are cleanly recalculated, it's no longer a shadow image for a demo, but a deployable asset that keeps the loop of design, client feedback, and partial correction going without the original file. This shifts the value of the creative from “how precisely do I formulate a prompt” to layout logic, editing instructions, and localization judgment. Anyone who dismisses the ten languages as a feature underestimates what this means for global campaigns: A multilingual key visual that considers script direction and cultural context saves production resources that are still spent per market today. Anyone running an agency pipeline this week should test the layer export against their own approval and correction loop before the next provider catches up. The craft isn't disappearing; it's becoming an engineering and directing job with better tools.

Google AI Overviews: When Hallucinations Become Industrial-Scale Disinformation

An analysis commissioned by The New York Times from the startup Oumi has examined Google's AI Overviews and found an accuracy rate of about 90 percent: 85 percent for Gemini 2 in October, 91 percent for Gemini 3 in February, measured against OpenAI's fact-checking benchmark SimpleQA across 4,326 search queries. Sounds solid, until you factor in the scale. Google processes over five trillion searches a year, and Sundar Pichai estimated the number of AI Overviews users at more than two billion per month in July. David Bader of the New Jersey Institute of Technology gets to the heart of the real concern: not the error rate, but the reach with which false answers are delivered with the visual authority of the world's most trusted search engine. Google is pushing back, calling the study methodologically flawed because one AI model (HallOumi) is grading another and the dataset was artificially generated. Researchers like Maarten Sap from Carnegie Mellon partially agree but emphasize that independent audits are essential as long as the providers do not provide controlled access for outsiders. The root cause remains technical: These models don't look up facts; they predict the next word. → The Deep View

Synthszr Take: The 90 percent is the wrong number everyone is staring at. When two billion people a month are served every answer with the authority of the Google logo, then ten percent errors become millions of false statements per hour, each one packaged in the appearance of authority. This fundamentally changes the role of search: The gatekeeper between question and answer now makes its own judgments, and it does so with a convenience that trains us not to double-check (I catch myself believing the box instead of clicking two links further). What's interesting is Bader's actual demand, and it can be implemented immediately: Anyone rolling out AI at this scale should fund independent audits instead of dismissing every external review as methodologically flawed. As long as Google doesn't disclose its own, independently verified accuracy data, rejecting the study is just a PowerPoint illusion of reliability. For brands, this means double the caution, because the machine not only decides who gets mentioned, but also what is claimed about them. Anyone who checks today how they are represented in the model gains a real advantage.

Claude Under Fire: Russian Propaganda and the Question of Sources

The Berlin-based think tank Agora has investigated where large language models get their news, and the result hits Anthropic where it hurts: In individual answers, Claude refers to content from the Russian Pravda network and treats it as a serious news source. Behind this is “LLM Grooming,” the mass flooding of the internet with identical content aimed not at human readers, but at AI systems. The study also shows how narrow the source base is: Of 544 identified domains, 248 are cited only once, while welt.de alone is cited 380 times. ChatGPT gravitates heavily towards “Welt” and “Bild,” which Agora links to the partnership between OpenAI and Axel Springer, while Claude and Perplexity also tap into more public and private sources. OpenAI disagrees: Which sources appear depends solely on usefulness and relevance, not on commercial agreements. The timing is sensitive, as a Munich court has just confirmed that Google is liable for its AI-generated answers, and a Yougov study confirms the rapidly growing influence of chatbots on public opinion. → MEEDIA Daily Update

Synthszr Take: The real problem isn't the Pravda network, but the number 248. When nearly half of 544 domains appear only once and one provider racks up 380 citations, what we have is a monopoly market for attention that AI doesn't break up, but rather cements. It's precisely into this narrow source base that the Russians are dumping their content, because a thin knowledge base is easy to manipulate. And the commercial deals exacerbate this: OpenAI can claim as credibly as it wants that the Springer partnership has nothing to do with citation density, but the correlation is still in the report. Anyone integrating an AI assistant into products today is also buying into the provider's source policy, and this responsibility cannot be delegated away. The pragmatic step is to test the source profiles of the models against each other now, and not blindly rely on a default engine. Vendor neutrality was long a technical argument; now it's becoming a question of trust and liability.

AI as a Geopolitical Chessboard: Washington and Beijing Tighten the Reins

Both of the world's largest economies now treat AI models as a matter of national security. Washington has restricted international access to Anthropic's Claude and OpenAI's GPT-4o after it became clear that Claude could help attackers find software vulnerabilities faster than ever before. The US imposed export controls and blocked the model for about three weeks so that authorities and corporations could harden their systems. Beijing is following suit: According to Reuters, China's Ministry of Commerce met with Zhipu AI, Alibaba, and ByteDance in June to restrict foreign access to the most advanced Chinese models and to slow down foreign investors. This could make waves in the US, as OpenRouter reported that over 30 percent of tokens used on Chinese models came from the US, and in some weeks of 2026, as much as 46 percent. Additionally, China's Ministry of Industry warned that Anthropic's Claude Code contains a backdoor vulnerability, and Alibaba banned its employees from using Anthropic tools. Earlier model generations and most open-weight models like DeepSeek, Qwen, and GLM are expected to remain unaffected. → The Deep View

Synthszr Take: The real news is in the 46 percent. American IT departments are flocking to Chinese open-weight models because the token bills from OpenAI and Anthropic are getting out of hand, and it is precisely this dependency that is now becoming a political lever. Anyone who has built their cost optimization on DeepSeek or Qwen has created a vendor dependency that can be cut off overnight by a ministerial decree. As was already seen with the great decoupling at the end of April, the blocs are hardening, while engineers underground continue to cheerfully copy from each other across the border. That's why a clean migration path belongs in every setup: a model adapter with an identical prompt layer so that a switch between Claude, GPT, Gemma, Nemotron, or GPT-OSS can happen without re-architecting. This can be decided this week and costs nothing but discipline. Anyone who ties their compute to a single flag is playing someone else's game.