Anthropic's Coding Position Faces Competition from China

ZCode Challenges Claude Code, Cursor & Co

Beijing-based Z.ai (formerly Zhipu AI) launched ZCode on Wednesday, a free desktop app for macOS, Windows, and Linux built as an “Agentic Development Environment” for its in-house GLM-5.2 model. The tool directly competes with Cursor, Claude Code, GitHub Copilot, and Google's Antigravity, supports BYOK configurations, and can be controlled from a mobile phone via WeChat, Feishu, or Telegram. The real kicker is the price: The GLM coding plans start at $16.20 per month and go up to $144 for “Max,” significantly undercutting Anthropic and Cursor. GLM-5.2 itself is a 744-billion-parameter MoE architecture with 40 billion active parameters, a true one-million-token context window, and was trained on 28.5 trillion tokens, released as open source under the MIT license on Hugging Face. On the Code Arena, the model ranked second worldwide in mid-June, just behind Anthropic's Claude 3.5 Sonnet. The crucial detail: It was trained entirely on Huawei silicon, without a single American chip, at an estimated total cost of $25 million. Gartner estimates the market for agentic coding tools will be around $10 billion by 2026. → Techpresso

Synthszr Take: $25 million in training costs on Huawei chips, and the result lands second on a coding leaderboard. That's the real story, not the app itself. In March, Cursor accused Anthropic of massively subsidizing its prices; now comes a provider that starts its coding plans at $16 and gives the model away as open source. Anyone establishing a 2-tool standard for their growth engineering in 2026 should seriously test GLM-5.2 as a BYOK option for regulated or cost-sensitive workloads, not dismiss it. The geopolitical question remains: a model on Chinese hardware, controlled via WeChat, is not a simple matter for European compliance. It requires clear guardrails and on-premise verification. But the price pressure still has an effect because it shifts expectations for all providers. Jevons paradox is in full effect here: the cheaper coding gets, the more of it will happen, and the advantage goes to those who build a model-agnostic architecture early instead of chaining themselves to a single vendor.

Sam Altman Offers Trump a 5% Stake

OpenAI has offered the U.S. government a 5% stake, valued at around $42.6 billion, according to a Financial Times report. The figure is based on the $852 billion valuation investors assigned to the company three months ago. Sam Altman reportedly wants Anthropic, Google, and Meta to cede a similar 5% share. Altman has discussed the plan with President Trump, Commerce Secretary Howard Lutnick, and Treasury Secretary Scott Bessent. Senator Bernie Sanders rejected the offer, instead calling for a one-time 50% tax on the holdings of OpenAI, Anthropic, and xAI. The move comes as political pressure on the ChatGPT maker grows in Washington. → Techpresso

Synthszr Take: Sam Altman is playing a familiar tune. By positioning oneself as quasi-public infrastructure, the goal isn't to be regulated but to be protected. A 5% government stake is the perfect bait: it turns Washington from a gatekeeper into a co-owner, with all the conflicts of interest that entails. Dean Ball is right with his rat analogy; once you make the regulator a shareholder, you can hardly ask them to judge independently. The contrast with Anthropic is interesting, as they demonstratively keep their distance and prefer to work on their own chips with Samsung, while OpenAI seeks political protection. Both are building a moat, but one is made of silicon and the other of lobbying relationships. For European companies using these models, this simply means the power struggle over the AI value chain is being decided in Washington, not Brussels.

Pay to Crawl: Cloudflare Enforces Bot Separation

Cloudflare is giving the AI industry a deadline: starting September 15, 2026, the provider will block so-called “mixed-use” crawlers by default on all ad-supported sites. This refers to bots that handle search, agent usage, and training all at once. Anyone who wants to train a large language model in the future will have to clearly separate their crawler from traditional search, or they won't get in. The rule applies to new customers, newly created sites, and all existing Free customers. CEO Matthew Prince cited a specific reason: bots have surpassed human traffic on the internet for the first time, a year earlier than expected. Cloudflare is visibly targeting Google, which it claims accesses “twice as much” data as others by intertwining search and AI training. “Pay Per Crawl” is becoming “Pay Per Use”: publishers should be paid when their content creates value, not just when it's collected. The first partners are Ceramic.ai and You.com. → AI Secret

Synthszr Take: Finally, someone is drawing a line in a field that everyone has been harvesting for free. The real masterstroke is the shift from “Pay Per Crawl” to “Pay Per Use”: payment is based on value creation, not retrieval. This is precisely the billing logic the open web has been missing for years, and it's coming late, but not too late. The figure of over 50 percent crawl traffic re-fetching unchanged pages is remarkable: a pure waste of bandwidth and compute that no one will miss. Cloudflare is positioning itself as a gatekeeper with a cash register at the door, hitting Google where it hurts most, as its Googlebot doesn't separate search and AI features. Any publisher still waiting for this problem to solve itself is flogging a dead horse. The clear separation of bot intentions is the lever that will decide in the coming months who still gets paid for their content in the AI era.

Palantir: US Government Clients Switch to Open Source

Palantir CEO Alex Karp is positioning himself as the gatekeeper between his clients' budgets and model providers OpenAI and Anthropic. After a heated CNBC appearance where he accused the two firms of scraping customer data and overcharging (there is no evidence for this; both prohibit training on customer data in their terms), he is positioning Palantir as a protective “Application Layer.” Specifically, some U.S. government clients have switched from proprietary models to Nvidia's open-source model Nemotron, with the Department of Defense being hinted at. Palantir has launched a product to help agencies securely operate Nemotron and continuously adapt it for their own environment. Karp expects every client to go open source “as soon as they see it as equivalent.” Nvidia's Jensen Huang clarifies that it's not an either-or but “proprietary and open”; according to Artificial Analysis, Nemotron lags behind closed frontier models and Chinese open-source models but is the leading U.S. open model. The stock rose 12% after the statements but is still down 22% for the year. → Applied AI

Synthszr Take: Karp is pulling the Red Hat maneuver of 2026. When Linux displaced the closed Unix derivatives from Sun, DEC, and HP, it wasn't the kernel that made the money, but the companies that made it usable. That's exactly where Palantir is positioning itself: Nemotron is freely available and cheap, but operating your own inference toolchain and tuning a model for your own environment remains hard work that few agencies can handle themselves. The moat shifts from the model weights to the service layer that orchestrates the whole thing securely and adaptably. Karp's salesman pitch about the model providers' evil intentions is marketing, but the underlying argument holds: as soon as open source reaches parity, no one has a reason to get into a proprietary lock-in. Anyone building an AI architecture today should build the model adapter to be replaceable and place the value creation in their own data layer. The 12% stock jump is a bet on this vector, not on Nemotron benchmarks.

TabFM: A Foundation Model for Tabular Data Without Fine-Tuning

Google Research has introduced TabFM, a foundation model for classification and regression on tabular data. The key feature: zero-shot predictions via in-context learning, completely without fine-tuning, hyperparameter optimization, or manual feature engineering. The model ingests the entire dataset (training examples plus target rows) as a single prompt and generates the prediction in a single forward pass. Technically, TabFM combines approaches from TabPFN and TabICL: alternating attention over rows and columns, compressing each row into a dense vector, and a transformer that operates on these compressed embeddings. With this, Google is challenging the classic tree-based methods like XGBoost, AdaBoost, and Random Forests that have dominated this field for years. TabFM is now available on Hugging Face and GitHub. → The Deep View

Synthszr Take: Tabular data is the backbone of almost every enterprise infrastructure, and that's exactly where the bottleneck has been. Getting an XGBoost model to work on a new dataset was never a single .fit() call, but weeks of manual labor: feature engineering, hyperparameter tuning, forcing domain knowledge into the model. TabFM commoditizes exactly this manual work, hitting an entire profession of data scientists right where their most expensive activity lies. The real value thus moves further up the stack: those who can clearly define which question the table should answer and which prediction actually drives a business process will win. Pulling the model takes minutes; defining the success criteria remains the real intellectual work. Anyone still approving budgets for months-long churn and fraud modeling should redo the math before a competitor does.

Cursor for iOS: Code from Anywhere

Anysphere is bringing its coding agent to the iPhone. The new Cursor app lets you start always-on agents directly from your phone or remotely control agents running locally on your computer. You get live updates on their progress and can merge pull requests from your phone without being at your desk. With up to eight parallel agents and its Composer mode, Cursor is already one of the strongest IDE-first tools in a field of 15+ serious players. The move to the smartphone decouples coding from the workstation and turns the IDE into a pure control room. → Cursor Team

Synthszr Take: The interesting movement isn't on the iPhone, but in the distribution of roles. When an agent independently writes database migrations, builds API endpoints, and adds tests while you just provide direction and approve PRs from your phone, the IDE has become a monitoring console. The human steers and corrects; the machine works in loops. This is exactly the transition that Cursor, Windsurf, and Devin have been making since late 2024: moving away from reacting and toward autonomous action. For teams, this means changing their mode of work instead of continuing to type lines of code. Anyone who still believes development is tied to a screen and keyboard risks flogging a dead horse. The real competence is shifting to the precise formulation of intent, and that can be done just as well from a train platform as from an office chair.

Higgsfield VFX: The Final Boss of Video Editing

Higgsfield has released a tutorial that takes VFX out of the toolchain of After Effects, 3D modeling, and tracking and moves it into a prompt pipeline. The core concept: Claude reads the uploaded footage frame by frame, the user describes the desired change in natural language (e.g., “when I snap my fingers, the background turns into a desert”), and a skill generates the complete prompt for Seedance 2.0 in 4K. Using a “lock header” concept, the face, clothing, lens, and camera movement remain fixed while the environment or individual elements are swapped out with precise timing. The examples range from simple background swaps to a six-step transformation of a hand into a robot arm, to handheld shots with a kraken attack, sauropods in the rainforest, and collapsing temples. 4K is not a luxury here but the threshold below which facial and lip-sync quality breaks down in 1080p. There are still limitations with complex movements and long clips, but the level is already practical for real-world use. → Trendium.ai

Synthszr Take: The interesting thing here isn't the kraken, but the division of labor. An LLM takes on the role that was previously the bottleneck: technically translating a creative intent into correct parameters. This is exactly the shift from a single model to an orchestrated pipeline that's happening across the entire stack (only now it's visible in consumer video). Higgsfield sits in the application layer, which is notoriously the one with the highest margins and the highest risk of substitution: Claude analyzes the frames, ByteDance's Seedance renders, and Higgsfield provides the skill logic in between. If you only serve one of these layers, you have to differentiate through workflow cleverness, and that's precisely what they're trying to do with the “lock header” approach. The real message is in the last paragraph of the tutorial: the emphasis shifts from filming to directing, from “what can I film” to “what do I stage.” For anyone producing video, this shifts the scarce skills. Craftsmanship in compositing becomes replaceable; the ability to precisely describe what should happen becomes the only real advantage.

Job Market and AI: The Human Returns

More and more companies that cut jobs in favor of AI are now reversing course and hiring people again, according to a CNBC report on Tuesday. Ford has rehired and promoted over 350 experienced engineers after automated quality control couldn't replicate the veterans' experience; AI is “a fantastic tool,” said hardware chief Charles Poon, but it's only as good as the training data. Australia's Commonwealth Bank replaced over 40 service staff with voice bots, only to reverse the decision due to rising call volumes. IBM plans to triple its entry-level hiring in the U.S. by 2026 because judgment, ethics, and complex decisions remain human tasks. An Orgvue study shows that 39 percent of executives have laid off people because of AI, and 55 percent now consider it a mistake. Robert Half reports that 32 percent of U.S. hiring managers who eliminated a position primarily due to AI later reinstated it. Meanwhile, Oracle, Meta, and Intuit continue to pump billions into AI while simultaneously announcing job cuts. → Techpresso

Synthszr Take: A 55 percent regret rate isn't a footnote; it's a mass phenomenon with a timeline. These companies booked AI as a replacement and then realized they bought an assistant. The Ford case nails it: the model responds in seconds, but the veterans' quality judgments are rooted in experience that no training dataset knows. This aligns perfectly with what IBM's own CEOs are saying: 83 percent see their AI success as more dependent on human adoption than on the technology. The bottleneck was never computing power, but proximity to context, the customer, the edge case. Those who lay off now only to rehire at a higher cost in six months are paying tuition twice and losing operational knowledge in the process. The sensible calculation is this: AI takes over the routine tasks, humans get the more demanding parts, and those who set this up early will save themselves the embarrassing U-turn.

Boris Cherny: Five Archetypes Instead of Job Titles

Boris Cherny, who leads the team behind Claude Code at Anthropic, describes in a widely shared post how classic roles are dissolving. Engineering, Product, Design, and Data Science are merging, to the point where old job titles barely explain anything. Instead, Cherny recognizes five archetypes in his own team: the Prototyper, who produces ideas on an assembly line (most die early); the Builder, who brings them to production; the Sweeper, who deletes and simplifies; the Grower, who hunts for product-market fit; and the Maintainer, who keeps a mature system alive at scale. These patterns describe what people actually do, not what's on their business card. For Cherny, this is the logical consequence of AI tools tearing down the boundaries between disciplines. → AI Secret

Synthszr Take: Cherny is describing the same shift that's becoming visible everywhere: code has become cheap, intent is expensive. When output costs nothing, input becomes the scarce commodity, and that input is judgment about the phase a product is in. The danger lies right here: a team with the wrong instincts will now build ten times faster, thus producing ten times the incoherence. Accelerating an unclear strategy entangles you in a thicket from which there is no escape. Organizing the org chart by craft is therefore ready to be retired, because the scarce resource is no longer distributed along job titles. It resides in the minds of those who sense whether it's time to perceive or to deliver. The team that assembles based on instinct rather than skill has already built in the decisive advantage.

Repo Radar: Five GitHub Projects Where Agents Do the Work

This week's GitHub trending radar shows five projects where agents are moving out of the chat window and into real production work. OpenMontage has cracked 31,003 stars and turns a coding assistant like Claude Code, Cursor, or Codex into a complete video studio: 12 pipelines, 52 tools, budget caps, and a decision log for every provider call (AGPL-3.0, difficulty 5/5). Google Labs' design.md introduces a format specification that links machine-readable design tokens in YAML with human-readable justifications in Markdown, so agents don't drift from the brand between sessions (Apache-2.0, difficulty 1/5, still marked as alpha). Strix by usestrix sends autonomous agent teams through the OWASP Top 10, intercepts HTTP, opens a shell, and writes working proof-of-concept exploits (29,913 stars). Alibaba's page-agent controls web interfaces using natural language, and MinerU turns chaotic PDFs into clean, LLM-ready Markdown. Each of these repos targets an agent at one specific task. → Marcus Schuler

Synthszr Take: Back in May, we noted here that agents are becoming more important than models. These five repos are proof that the debate has already moved a level deeper: it's no longer about whether an agent can write code, but what it delivers as a finished product that a human just needs to approve. OpenMontage is the most exciting case because a coding tool is leaving its home turf to produce video. This is the expansion of the coding agent into a general production layer, and that's where the value lies. Two things are worth considering before productive use: the AGPL-3.0 license for OpenMontage comes with real copyleft obligations, and the nice marketing results require a GPU plus cloud video models; the free path only covers local speech synthesis and free footage. If you invest two hours this week, clone design.md and Strix and test them on a real repo instead of waiting for the next model release. The tools are on the table; those who snap them into their workflow in 2026 will have the foundation for scaling in 2027.