GPT-5.5 is here and the era of the great office wars begins

GPT-5.5 is here, the super app is not

OpenAI has introduced GPT-5.5, a new flagship model. The crucial point: The model was developed with the help of the in-house coding tool Codex and its predecessor model—a process the industry calls “recursive self-improvement” (RSI). OpenAI President Greg Brockman confirmed that GPT-5.5 will power the company's upcoming “super app,” with new features being rolled out gradually. The model shows significant improvements in long-term tasks, can better interpret vaguely formulated problems, and is said to perform “economically valuable work” in 44 professional fields. Despite its expanded capabilities, OpenAI promises no sacrifice in latency—a critical point given the increasingly sluggish performance of competing models like Gemini 3.1 Pro and Claude 4.7 Opus. → The Deep View

Synthszr Take: OpenAI is no longer building models, but model factories. Recursive self-improvement is reminiscent of the invention of the machine tool in the 19th century: machines that could build more precise machines triggered the Industrial Revolution. Here, Codex becomes a meta-tool that not only writes code but also develops the next generation of itself. The real innovation isn't in GPT-5.5's benchmark scores, but in the fact that OpenAI has established a self-reinforcing development cycle. While competitors are still manually tweaking models, OpenAI is already automating model development. The “super app” is almost an afterthought—it's just the user interface for a system that improves itself.

Office War (I): OpenAI offers new “Workspace Agents”

OpenAI is transforming ChatGPT from a dialogue system into an infrastructure for autonomous office workers. The new “Workspace Agents” can independently access Slack, Google Drive, Salesforce, and other enterprise apps, write code, analyze data, and work on tasks for weeks—without human supervision. For $20 per user on the Business plan, companies get agents that can remember processes, react to schedules, and continue their work over several days. The backbone is Codex, OpenAI's cloud-based coding environment with over 90 plug-ins that allow the agents to manipulate real files instead of just generating text. The service is free until May 2026. → Techpresso

Synthszr Take: OpenAI is turning the office into a franchise system: The headquarters (IT department) defines processes and permissions, while local teams run their own agent branches. This is reminiscent of the history of industrialization, except this time it's not craftsmen being replaced by factories, but knowledge workers by persistent code-execution environments. The crucial difference from previous automation tools: These agents have memory and initiative—they wake up on their own, resume interrupted work, and learn from corrections made by their human colleagues. What OpenAI is selling as a “Workspace” is actually a new organizational form: companies become orchestrators of agent swarms that write their own work history. The two free years until May 2026 is less about generosity and more about the time employees will need to realize they are currently defining their own replaceability.

Office War (II): Microsoft's Copilot can now do more than Clippy (1997)

Microsoft is turning Word, Excel, and PowerPoint into autonomous work environments: The new Copilot executes actions directly in documents instead of just making suggestions. After a year of development, the AI can now understand pivot tables, apply PowerPoint animations, and format Word citations precisely. Usage data from the first few weeks shows drastic increases: Excel is seeing 67% more engagement and 65% higher satisfaction, while Word users are 11% more likely to stick with it. Microsoft is thus solving a fundamental problem of the first Copilot generation: passive AI assistants that could only respond are being replaced by active agents that get hands-on. Control remains with the user; execution is handled by the machine. → techmeme

Synthszr Take: Microsoft is transforming Office into an agency where humans act as art directors and AI agents handle the execution. The principle is reminiscent of the division of labor in 1960s advertising agencies: the creative director sets the vision, and the teams implement it. Except now the teams consist of specialized models optimized for Excel formulas, PowerPoint animations, or Word formatting. The 67% increase in Excel engagement shows that people are willing to give up operational control if they retain strategic control. Microsoft is no longer creating tools, but digital employees with Office expertise. The line between software and service is blurring for good.

Office War (III): Google follows suit with Workspace feature

Google is integrating Workspace Intelligence into its office suite, automating routine tasks from email drafting to spreadsheet management. The system accesses user data from Gmail, Calendar, Chat, and Drive, with administrators retaining granular control over data access. Gemini reportedly fills Google Sheets 9 times faster than humans, converts unstructured data into tables, and creates documents in the user's personal writing style. The new features are clearly aimed at enterprise customers, where Microsoft, Apple, and startups are competing for the same high-paying target audience. Google is using its office infrastructure, already deeply rooted in companies, as a sales channel for AI upgrades. → TechCrunch

Synthszr Take: Google is selling the perfect office intern: always available, never forgets anything, never complains. Workspace Intelligence works like a digital ant colony, where each worker AI has access to the company's collective memory (Gmail, Drive, Calendar). The promise of being “9x faster than humans” in Sheets is reminiscent of the introduction of the calculator, except this time it's not individual calculations but entire work processes that are being automated. The real kicker is the mimicry feature: Gemini writes in your style, as if it had been looking over your shoulder for years. Google is betting that companies will replace their human interns with AI features that never spill coffee or ask for a permanent position

OpenAI is now better at redacting the Epstein Files

OpenAI has released Privacy Filter, an open-source model for detecting and redacting personally identifiable information (PII) in texts. The 1.5-billion-parameter model (50 million active) can be run locally and processes up to 128,000 tokens in a single pass. Unlike rule-based systems, Privacy Filter understands context: it distinguishes between public information that should be preserved and private data that must be redacted. The model recognizes eight categories of sensitive data, including private addresses, emails, phone numbers, account numbers, and API keys. OpenAI itself uses a fine-tuned version of the filter in its own data protection workflows and claims to achieve state-of-the-art performance on the PII-Masking-300k benchmark. → Casey Newton

Synthszr Take: OpenAI is turning data protection into a scalable business model. While regulatory requirements like GDPR were previously seen as a cost factor, compliance is now becoming a product category: a model that can be run locally, fine-tuned, and integrated into one's own pipelines. This is reminiscent of the emergence of the antivirus industry in the 90s, when a necessary evil became a multi-billion dollar market. The strategic move lies in the timing: companies wanting to deploy AI systems need tools right now to overcome regulatory hurdles. OpenAI is positioning itself as an infrastructure provider for the entire industry (hence open-weight rather than closed). Privacy Filter is becoming the transmission belt between AI innovation and legal reality.

Anthropic's market valuation surpasses OpenAI's on the secondary market

Anthropic has reached an implicit valuation of one trillion dollars on secondary markets, surpassing OpenAI as the most valuable private AI company. On Forge Global, one of the largest marketplaces for private shares, Anthropic is valued at about one trillion dollars, while OpenAI is at $880 billion. The main reason for this shift is Anthropic's explosive revenue growth: from $9 billion at the end of 2025 to $30 billion in March 2026, an increase of 233 percent in just one quarter. Demand is primarily driven by Anthropic's coding tools. The extreme valuation also stems from scarcity: employees and early investors have had few opportunities to sell, driving up prices on the secondary market. An IPO is planned for October 2026, with Goldman Sachs and JPMorgan anticipating a target valuation of $400-500 billion, about half of the current secondary market valuation. → Techpresso

Synthszr Take: The secondary market confirms what Synthszr has been writing in its corporate analysis of Anthropic for a long time: artificial scarcity drives prices. While OpenAI regularly pays out its employees (most recently for $1.5 billion), Anthropic is keeping the floodgates closed. This creates a bottleneck effect where every available share becomes a collector's item. The 233 percent revenue growth is impressive, but it's mainly based on one product: the coding assistants for enterprises. This is reminiscent of the early days of Salesforce, which also rose to a billion-dollar valuation with a single use case (CRM). The real test will come with the IPO: if the bankers are only targeting half of the secondary market valuation, the market is either betting on a spectacular surprise or a repeat of the 2021 hype. Anthropic is currently selling the most expensive ticket to a movie no one has seen yet.

Vercel Breach: It was Roblox cheating

Hosting giant Vercel has admitted that hackers had penetrated customer systems even before the April breach. An employee had downloaded an app from the startup Context AI that was infected with infostealer malware. The attackers systematically used the stolen credentials: they searched developer machines for API keys and tokens to gain access to Vercel accounts and other services. CEO Guillermo Rauch describes a recurring pattern: “As soon as the attackers have the keys, our logs show rapid and comprehensive API usage, with a focus on enumerating non-sensitive environment variables.” The hackers used the compromised employee accounts to access internal systems, including unencrypted customer data. Security researchers report that a Context AI employee allegedly caught the malware while searching for Roblox game cheats. → Techmeme

Synthszr Take: Vercel unwittingly demonstrates the fundamental problem of the AI agent era: agents need real access rights to be useful, but every token becomes a single point of failure. The infostealer campaign works like a parasitic organism eating its way through the ecosystem: an infected developer machine becomes a gateway into entire infrastructures. What used to require social engineering now runs on autopilot: malware scans for .env files, AWS-Credentials and API-Keys, while developers are googling for Roblox cheats. The irony is that the more we automate processes with AI agents, the more valuable the keys these agents need become. Vercel's breach is just the beginning of a new era in which every API key becomes a crown jewel.

Minimal Editing as a Metric: The community finally recognizes that less code is sometimes more

The coding community has developed a new benchmark for the “Minimal Editing” problem, which shows how AI models overdo it when fixing bugs. The study by @nrehiew_ specifically constructed minimally corrupted problems and measures excess edits using patch distance and additional cognitive complexity. GPT-5.4 has the strongest tendency for over-editing, while Opus 4.6 is the most restrained. Reinforcement Learning outperforms Supervised Fine-Tuning, DPO, and Rejection Sampling when it comes to learning a generalizable minimal editing style. The insight: models must not only fix bugs but also respect the existing code structure while doing so.→ AINews

Synthszr Take: Code editing is the new chess for AI models: a game where elegance counts more than raw power. The “cognitive complexity” metric is reminiscent of scoring in figure skating, where technical difficulty is weighed against artistic expression. It's no surprise that Reinforcement Learning shines here: the approach learns through trial and error what humans understand intuitively—that a surgical incision is better than an amputation. GPT-5.4's tendency for over-engineering reflects the Silicon Valley syndrome: if all you have is a hammer, everything looks like a nail that needs to be completely reforged. The real revolution is that we are starting to evaluate AI systems not just on correctness, but also on style.

GEO/AEO and the Fan-Out Effect

AirOps analyzed 17,000 queries and 354,000 pages to understand what content ChatGPT actually cites. The result is surprising: it's not the relevance of the content that determines the number of citations, but structural factors like heading hierarchy, recency, and domain authority. The “Fan-Out Effect” shows how the gap between a search query and the final citation is created. Companies like Webflow report 6% more AI-generated sign-ups within a few days after optimizing their content structure. Chime tripled its number of AI citations from 24 to 68 on priority questions. Carta and Ramp confirm: the quality of their content increased significantly through structural adjustments. → Techpresso

Synthszr Take: ChatGPT works like a librarian with OCD: it doesn't grab the best book, but the one that's sitting most neatly on the shelf. The Fan-Out Effect reveals the new currency of the AI age: structural excellence beats content brilliance. Companies are no longer optimizing their texts for humans or Google, but for the retrieval logic of language models. This is reminiscent of the early SEO days when keyword stuffing still worked, only this time heading hierarchies and recency signals are the new meta tags. The irony: while we expect AI to understand context, it rewards mechanical structure over semantic depth (more on this here).

After Chess and Go: Table Tennis Robot Beats Humans

Sony AI has developed a robot called “Ace” that can beat professional table tennis players. The autonomous robot uses event-based sensors that only analyze relevant image areas and was trained using model-free reinforcement learning over thousands of simulation hours. With eight joints and a reaction time of 20 milliseconds (humans need 230 ms), Ace won against five elite players, each with over ten years of competitive experience. Sony compares the breakthrough to Deep Blue's chess victory over Kasparov in 1997, only this time in the physical realm. The robot still looks like an industrial machine but is planned to be built in a humanoid form in the future. → MIT Technology Review

Synthszr Take: Sony is celebrating its table tennis robot as a historic moment, but the real story is that no one is applauding anymore. 30 years ago, Deep Blue's chess victory was a sensation; today, a robot beating humans in a sport is a side note between OpenAI's o3 model and Anthropic's Claude updates. The shift shows how much our expectations have normalized: machines surpassing us in isolated domains have become a commodity. What Ace really demonstrates is not the superiority of robots in table tennis, but the speed at which specialized AI loses its magic. The next milestone won't be when robots beat us in every sport, but when we stop reporting on it.