GLM-5.3 Stuns with Performance Boost, and OpenAI Can't Catch a Break

Z.ai launches GLM-5.3 and stuns with performance boost through post-training

Z.ai introduced GLM-5.3 on Friday, a coding and agent model that, according to the company, uses the same base as its predecessor, GLM-5.2. The entire performance gain reportedly comes from an additional month of Post-Training: more training environments, greater task variety, and more compute time on the stack already built with GLM-5.2, consisting of IndexShare for long contexts, the RL method SAO, and the training framework slime. The weights are not yet public; they are expected to follow about two weeks after launch, once security auditing and hardening are complete. The Beijing-based company, founded by Zhang Peng, Tang Jie, and Liu Debing and formerly known outside China as Zhipu AI, is positioning the model directly against Anthropic and OpenAI in the coding segment.

The core of the described recipe is scaling the task environments rather than changing the architecture. According to Z.ai, these environments are built on real work units: in one example provided, the model is given the work environment of an ML infrastructure engineer with access to compute clusters, internal documentation, codebases, and experiment results, and is tasked with diagnosing bottlenecks, implementing optimizations, and delivering a measurable speedup. Individual tasks would correspond to several workdays for an experienced developer. To generate such environments en masse, Z.ai uses its own pipelines in which research agents translate task patterns from real work into executable environments, and a judge agent checks whether a task is solvable at all; verifiers are synthesized without access to the reference solution. Human post-processing is still said to be necessary.

The reported numbers are manufacturer’s claims. On Terminal-Bench 3.0, GLM-5.3 improves from 4.6 to 28.3 points compared to GLM-5.2, on DeepSWE v1.1 from 46.2 to 66.9, and on the CLI part of Agents' Last Exam from 23.8 to 28.5. On its in-house Z.ai Code Bench, the company claims a 50 percent improvement. The cyber scores are particularly strong: On CyberGym, Z.ai reports 84.5 points, on ExploitBench 54.4 compared to 24.4 for the predecessor, and on ExploitGym, 105 and 130 solved tasks respectively, instead of 29 and 39. The company itself writes that this capability has developed faster than expected. The comparison table includes Kimi K3, DeepSeek-V4 Pro, Qwen3.8-Max, Opus 4.8, and GPT-5.6 Sol, with the top Western models showing significantly higher scores on the exploit benchmarks.

The product’s status was unclear on the day of its release. Z.ai’s official release notes listed GLM-5.1 as the latest documented version, an API endpoint, a model identifier, or a download for GLM-5.3 were not listed, and the Hugging-Face access was marked as “Coming Soon.” This left developers without a confirmed access path and a reproducible production configuration; license and serving configuration are open. Regarding the company itself, the Hong Kong stock exchange prospectus lists 883 employees as of June 30, 2025, with 657 in research and development, a revenue of 312.4 million RMB in 2024 and 190.9 million RMB in the first half of 2025, as well as more than 12,000 institutional customers in the first nine months of 2025. → z, Bloomberg, The New Stack, South China Morning Post, Silicon Republic, Reuters, The Decoder, Caixin Global

Synthszr Take: Z.ai itself says that nothing about the base was touched, and in the same breath delivers a jump from 4.6 to 28.3 points on Terminal-Bench 3.0. If a single month of additional RL on synthesized work environments can achieve that, then the collection of executable and verifiable task environments determines the gap between a mediocre and a strong coding agent. These environments are the real asset, and Z.ai is now building them itself with agents that translate task patterns from real work into runnable environments, complete with a judge agent that first checks if the task is even solvable. Any organization that has its own workflows and testing criteria can replicate this: ticket histories and acceptance protocols are lying around unused in almost every large company. The price of this recipe is in the same post, because the same scaling that improves programming has more than doubled the exploit scores compared to GLM-5.2, and that’s precisely why the weights are coming two weeks later.

OpenAI (I): Employees question their own security culture after the agent hack

At OpenAI, an internal review is underway following the security incident involving an out-of-control AI agent, reports WIRED editor Maxwell Zeff in his newsletter Model Behavior. Zeff describes the event as a turning point for AI safety and cybersecurity and says he has never seen the company react so cohesively to a security incident. According to his account, many employees are internally asking what practices within the company led to the incident and how it can be prevented in the future. Zeff says he is, in his own words, “reasonably optimistic” that this will lead to tangible changes this time. Katie Drummond, Global Editorial Director at WIRED, places this in a series of recent news items where executives at AI firms are increasingly being held accountable for their agents' behavior. → WIRED

Synthszr Take: For security teams, the self-examination at OpenAI is a side note; their problem is the difference in tempo. An attacker with agents tries variants in series and overnight, while on the other side, someone is triaging a ticket and waiting for approval. Defense must therefore be automated where it still depends on individuals today: detection, blocking of credentials, revocation of rights.

OpenAI (II): Head of Sales Dresser leaves after less than a year

OpenAI’s Chief Revenue Officer, Denise Dresser, is leaving the company after less than a year at the helm of sales and Go-to-Market. Her successor as Chief Revenue Officer will be Dali Rajic, formerly President and Chief Operating Officer at the security company Wiz. In a blog post, OpenAI writes that Dresser wants to “pursue other opportunities”; during the transition, she will work closely with the business team to support customers. Dresser joined OpenAI in December 2025, after more than ten years as an executive at Salesforce, and was considered responsible for building the enterprise business. She had taken on additional responsibilities when COO Brad Lightcap moved into a new role focusing on “special projects.” → Business Today

Synthszr Take: Dresser arrived in December 2025 and is leaving before her first full sales year is through, and it is precisely within this window that the first major enterprise framework agreements with OpenAI were signed. Enterprise business depends on faces: the purchasing manager who negotiated pricing tiers, data protection addendums, and availability commitments with her is now sitting across from a man who addressed security budgets at Wiz and brings a different sales grammar. Everything that was verbally promised in those conversations and not written into the contract has effectively expired with her departure.

OpenAI (III): “Computer History” logs clicks and keystrokes on Mac

OpenAI has introduced a feature called Computer History for the ChatGPT desktop app on macOS, which makes past activity on the computer available for future conversations. According to Business Today, the feature processes interaction events from previously approved apps and websites: clicks, typing, keyboard shortcuts, and switching between applications. These events are periodically converted into text summaries and local memory files, from which ChatGPT generates a timeline of activity. Users can then ask about what they were working on before a break or where they saw a specific document. → Business Today

Synthszr Take: Clicks, keystrokes, keyboard shortcuts, and app switches: this is the data foundation that monitoring software typically works with, here with an added summarization step. The crucial half-sentence in the description is the periodic conversion into text summaries and local memory files, because locally stored does not mean locally processed as soon as these summaries appear as context in future conversations. The opt-in and pause button are only reassuring once the permissions are granular enough to keep password managers, email windows, and the browser tab with the performance review out of it.

DeepSeek completes V4 rollout: Pro version reaches agent level of Claude Fable 5

DeepSeek has completed the rollout of V4 and, according to MyClaw Newsletter, is making the model available in two variants: Pro and Flash. V4 Pro, according to the report, achieves the performance level of Claude Fable 5 on key agent benchmarks. Flash is positioned as a faster, more affordable option for everyday tasks. Both variants are now available on the MyClaw agent platform. → MyClaw Newsletter

Synthszr Take: The entire news story fits in three lines: Rollout complete, Pro on par with Fable 5 on agent benchmarks, Flash for everyday use, both now on MyClaw. This is what reporting looks like when the model cycle moves faster than the editorial deadline. Back in March, the talk was all about the OpenClaw hype around QClaw; five months later, the news value of a frontier model has shrunk to the question of how quickly the platform integrates it.

India’s central bank chief wants to use AI to grant loans to first-time borrowers and gig workers

The Governor of the Reserve Bank of India, Sanjay Malhotra, has urged banks to use artificial intelligence to approve borrowers who would be rejected by classic screening methods. Mentioned are first-time borrowers without a credit history, employees in the platform economy, and small businesses. As a data basis, Malhotra cites payment flows, tax returns, utility bill payments, and the applicants' digital footprints. This could expand the circle of creditworthy individuals at a lower cost than with traditional Underwriting. → MyClaw Newsletter

Synthszr Take: A credit model for gig workers without a credit history produces two bills: one for the defaults, which every bank sees immediately, and one for the rejections, which no one ever sees. Malhotra solves the clean part of the problem by leaving the liability with the bank and not with the provider of the scoring model. The hard part is explainability, because a system that combines electricity bills, tax returns, and digital footprints into a single number justifies its rejection as a probability, and a probability is not a justification before an ombudsman. Human oversight in practice means: log every rejection and have samples checked by humans, otherwise financial inclusion becomes automated sorting with a friendlier name.

X open-sources its “For You” algorithm and shows users their own ranking labels

X has published the source code of its “For You” timeline, including its core ranking system, on GitHub under the Apache-v2 license, as reported by TechCrunch. According to the company, the release now includes the model configuration, filters, and the parameters used to weight individual signals; this makes the disclosed codebase about ten to fifteen times larger than in the first disclosure step in January. In parallel, a page called “Under the Hood” is launching in the settings: users who have published at least ten posts in the last month can download their aggregated statistics as a JSON file and see if labels were assigned to their account or individual posts in the last calendar month, i.e., whether they were affected by a Shadowban. The feature is initially running as a pilot for a test group of accounts that are at least one year old. X’s head of product, Keith Coleman, told TechCrunch that external researchers were able to train and run the in-house scoring system Phoenix themselves with the open code beforehand. → Techpresso

Synthszr Take: Open code is symbolic politics as long as no one can check their own data against it; the JSON download with last month’s labels changes that. Ten to fifteen times more code than the first attempt, including the signal weightings: this is exactly the part that Meta, TikTok, and LinkedIn defend as a trade secret. The advantage lies in the reversed burden of proof, because with every reach complaint, X will now point to a repository and a file instead of issuing denials.

63% solo founders on Stripe Atlas: AI agents are replacing the co-founder

The proportion of single-founder new US startups has risen from 23.7 percent in 2019 to 36.3 percent in mid-2025, reports Linas’s Newsletter, citing Forbes. On Stripe Atlas, the incorporation service for US corporations, solo founders accounted for 63 percent of all newly formed C Corporations in the second quarter of 2026, an all-time high according to these figures. For 2025, Carta counts a 36 percent share of solo-founded startups among all those incorporated on its platform, about twice as much as a decade ago. The newsletter classifies this as a break with a convention that has been valid for over ten years: “Solo Founder” was considered a risk factor in venture capital, and accelerators actively pushed applicants to form teams. → Linas from Linas’s Newsletter

Synthszr Take: For a decade, the second person in the pitch deck was primarily a risk reduction that accelerators actively negotiated away through what amounted to matchmaking. At 63 percent solo founders in a single Stripe Atlas quarter, this heuristic is flipping, and selection processes and term sheets are not yet adjusted for it: what will be evaluated in the future is a person and their agent setup, not four résumés with a vesting plan. The effect on the job market is harsher than on the capital side, because for decades, the first support and junior development roles in a young company were the standard entry into the profession.

SpaceX acquires Cursor for $60 billion and trains Grok with its developer data

SpaceX completed the acquisition of AI coding startup Cursor for $60 billion on Friday. The two companies had entered into a partnership in April that gave Musk’s space exploration company the right to purchase; a binding commitment followed in June. Cursor announced on X that it would become part of the SpaceXAI team and continue to work on Grok Build, Grok Bot, Grok API, and Cursor itself. Musk replied to the post with one word: “Welcome.” On the August quarterly earnings call, he had withheld details of the partnership to avoid getting ahead of regulators.

Technically, both sides have been working together for some time. Cursor trained Grok 4.5 jointly with SpaceX, using what it described as trillions of tokens of Cursor data; it was the company’s first model not built exclusively for software development. Cursor was also involved in the Grok 4.6 released on Wednesday. In return, the startup gained access to Colossus, the SpaceX supercomputer with 200,000 Nvidia GPUs. Cursor writes that it now has the largest GPU fleet in the world, enabling it to offer more powerful models at lower operating costs; this is the company’s own statement.

Morgan Stanley estimates that Cursor could contribute about $2.5 billion to SpaceX’s revenue in 2026 and $13 billion in 2027. Analysts see this as the basis for a much more optimistic valuation if the coding platform maintains its growth pace. SpaceX went public in June, placing 555.56 million shares and raising about $75 billion. The IPO documents explicitly stated that part of the proceeds would be used to expand AI computing infrastructure.

In parallel, Musk asked the staff at an internal town hall to see themselves as “parents” of the company’s AI. The company wants to train Grok on the “sum” of its information, including the work and contributions of its employees, allowing the model to inherit their “thoughts, ideas, and beliefs.” Which data categories will be specifically included and how they will be handled has not yet been publicly disclosed by SpaceX; a request for comment went unanswered. In April, Meta’s Model Capability Initiative collected mouse movements, clicks, keystrokes, and screenshots from employees, with around 1,600 employees signing a petition against it.

In the same session, Musk stated that AI revenue could surpass that of all other SpaceX businesses as early as September. Ten gigawatts of computing capacity would, by his calculation, generate $300 to $500 billion in annual revenue. SpaceX currently has about 1.4 gigawatts and aims to reach ten gigawatts by the end of 2027. Starlink is now available in 167 countries and has about 22 million mobile users. → Bloomberg, Cursor, Fortune, Business Insider, The American Bazaar

Synthszr Take: Cursor delivered the purchase price in advance, in the form of the trillions of tokens from real developer work used to train Grok 4.5. The rest is a utilization issue: 1.4 gigawatts need to be paid for, and $13 billion in expected Cursor revenue for 2027 is the most plausible answer to that question that Musk could buy right now. For SpaceX, the coding assistant is primarily a machine that keeps Colossus busy every workday and, in return, provides a fresh stream of training data. Musk’s request to the staff to see themselves as parents of the AI belongs in the same calculation, except that at Meta, 1,600 people signed a petition against that very logic. If AI revenue, as announced, surpasses all other SpaceX businesses as early as September, this will have been the most expensive sales channel anyone has ever acquired for their own computing time.