Synthszr — Synthszr is the daily AI news synthesis: business, design and technology — with a newsletter, AI product charts and company analyses.

Anthropic boosts quarterly revenue amid decelerating growth ahead of IPO

In the third quarter, Anthropic generated around $17 billion in revenue and slipped back into an operating loss. The figures are contained in an updated version of the S-1 filing that Anthropic submitted to the U.S. Securities and Exchange Commission (SEC) for its planned initial public offering. The company had published the first version of the prospectus a week earlier. It included financial data only through mid-year.

Quarterly revenue was up by approximately 1,127 percent year-over-year. Compared to the previous quarter, it grew by roughly 47 percent. In the second quarter, growth had stood at 144 percent, and at 132 percent in the first quarter. For the first nine months of the year, Anthropic reported revenue of $33.24 billion, roughly 1,200 percent more than in the prior-year period.

Anthropic recorded an operating loss of $536 million in the third quarter. Three months earlier, it had posted an operating profit of $974 million. Two special charges were decisive. Around $5 billion was spent on computing hardware that had to be purchased on short notice and at a premium due to unexpectedly high demand. Anthropic typically books capacity 12 to 18 months in advance. Another $2 billion stems from a non-cash program under which the company matches employee donations. Adjusted operating income, excluding stock-based compensation and donations, rose by $300 million to $2.5 billion.

The net loss for the third quarter stood at $23.29 billion. Over nine months, it totals $104 billion, compared to $25.1 billion in the prior-year period. The company attributes the majority of this to changes in the fair value of convertible notes and derivative liabilities.

As of September 30, infrastructure commitments amounted to $895.4 billion. Nearly one-third of this is tied to leasing agreements with Google and Broadcom. In April, both companies pledged to provide Anthropic with several gigawatts of computing power based on Google’s TPU chips, which Broadcom co-developed. Data center contracts worth more than $100 billion each are in place with Amazon, Oracle, and SpaceX. In addition to leased capacity, Anthropic plans to build its own data centers. Purchase obligations increased from $55 billion at the end of 2025 to around $314 billion in the third quarter.

Anthropic faces competition from open-source models, OpenAI, and major tech corporations. → The Information, SiliconANGLE, CNBC

Synthszr Take: Even the loss is turned into a growth narrative in the prospectus. Five billion dollars for compute hardware at a premium reads like proof that demand is outstripping all projections. On top of that comes an adjusted operating figure that filters out donations and stock-based compensation. There is a reason for this staging: investors pay for revenue growth, so the curve has to stay steep. A quarterly gain of 47 instead of 144 percent compresses the multiple far more than any operating loss.

Anthropic warns in IPO prospectus of AI that could soon improve itself

Anthropic added new risk disclosures to its revised S-1A prospectus on Monday. According to these, the company could soon reach a threshold for Recursive Self-Improvement. Public attention initially focused mainly on the updated financial figures showing surging revenues and rising costs. The sections on self-improvement appear in the risk factors section. It states that internal estimates and expectations for future models suggest Anthropic could “imminently” meet the definition of “dramatically accelerated, AI-attributable progress” under its own Responsible Scaling Policy. The company stated that no current model has crossed this threshold, and the underlying data is being further reviewed.

As recursive self-improvement, observers define the point at which AI becomes so proficient at training itself that its capabilities increase in dramatic leaps. Pessimists see this as a scenario in which AI slips beyond human control. Proponents expect it to deliver a massive boost to productivity and scientific discovery. When the term gained traction earlier this year, many viewed it as a buzzword or a milestone that could hardly be achieved, defined, or proven.

In the prospectus, Anthropic states that it already uses AI in its own research and development. This could lead to model capabilities advancing faster than the company can develop, validate, and implement safety, alignment, and security measures. Anthropic concedes that it only recognized a notable acceleration in its own development after the fact in past instances. This could happen again in the future. According to the company, using AI to build more capable AI carries a “substantial safety risk,” especially if the pace accelerates significantly.

At the same time, Anthropic announced plans to keep humans involved in the development process for the foreseeable future. The company does not expect to be able to safely rely on Claude to design future frontier models fully autonomously and without meaningful human oversight. This holds true even if it becomes technically possible. Consequently, the company is evaluating adjustments to its Responsible Scaling Policy to ensure control over model development remains in human hands. Critics counter that an IPO prospectus is always part disclosure, part marketing. Nonetheless, Claude’s capabilities in coding, mathematics, and cybersecurity have grown substantially this year. → Wall Street Journal, TechCrunch

Synthszr Take: A risk factor based on future models and internal estimates costs nothing in a prospectus. It shields against shareholder litigation while selling the steepest conceivable growth curve. The claim will only be verifiable with the next Claude release: if the threshold is “imminently” upon us, public benchmarks in coding and mathematics would need to show a leap within months that breaks the previous cadence. If it fails to materialize, the passage was simply a growth story disguised as a warning.

Anthropic curbs Claude Code on reverse engineering

Anthropic has confirmed that Claude Code, in combination with the Opus 5.5 model, is no longer supposed to create clones of existing video games. According to Tom’s Hardware, Anthropic employee Addy Osmani stated on X that two sentences were added to Claude Code’s prompt because intellectual property infringement violates the company’s terms of service. According to him, original games, genre titles, and games made from custom designs and assets should continue to work, with the rules being continuously refined. A screenshot circulating on X purportedly shows the wording, which prohibits replicating a specific game even if it is renamed or branded as original. In parallel, Tom’s Hardware reports growing accounts that Opus 5.5 in Claude Code is also refusing reverse-engineering tasks, such as analyzing product activations. → Tom’s Hardware

Synthszr Take: Two sentences in the system prompt won’t protect an indie game. The restriction only applies in Claude Code with Opus 5.5. Switching to Opus 4.8 or another harness like Ollama is already enough to bypass it. The clone factories that copy successful titles from small studios rely on open models and cheap teams anyway. The main entity being protected is Anthropic itself, against major rights holders and enterprise buyers. The bill is footed by legitimate developers whose reverse-engineering checks are suddenly being rejected.

Anthropic removes the five-hour limit for Ultra subscriptions

Anthropic has lifted the five-hour usage limit for verified subscribers of its most expensive Claude plan. At the same time, the plans were renamed, reports AlphaSignal. The previous Max 5x plan is now simply called Max and continues to cost $100 per month. Max 20x becomes Ultra at the unchanged price of $200 per month. According to the report, nothing changes for existing subscribers other than the name. Previously, the five-hour limit interrupted long sessions, such as during extensive coding tasks or ongoing background processes. → AlphaSignal

Synthszr Take: The weekly limit remains, so an Ultra customer doesn’t consume a single token more than before; they just consume it all at once. The five-hour limit was therefore a pacemaker for load peaks, with zero cost impact per token. Anthropic is now selling continuity as a product feature without giving up any additional compute time. The real cost protection lies in verification, which keeps account farms out. For Cursor and OpenAI, the benchmark thus becomes session duration rather than quota size.

Musk admits: Grok costs xAI more than the chatbot brings in

Elon Musk has publicly stated that running the AI assistant Grok is no longer financially viable for xAI. Techpresso reports this, citing NewsBytes. The trigger was complaints from users on X who hit their usage limits faster than expected this week. One user suggested resetting the limits due to high usage. Musk rejected this and did not hold out any prospect of additional access for paying subscribers either. → Techpresso

Synthszr Take: An entrepreneur looking for capital rarely complains publicly about costs unless the complaint is part of the negotiation. Musk’s reply to a user on X establishes two things: the limits stay, and scarce compute time is now treated as proof of demand. With an estimated 800,000 paying users, this reads like the justification for a more expensive tier. Investors hear the same story as a growth problem—and growth problems are solved with a new funding round.

UX author Jon Sabutis calls out the ubiquitous AI sparkle button

In the weekly newsletter of the UX Collective, author Jon Sabutis criticizes the sparkle icon software vendors use to label their AI features. His piece is titled “Nobody asked for the sparkle button” and leads the current issue. Sabutis describes how almost every paid product now features such an icon in the top right corner. Behind it, according to his description, a panel slides in from the right labeled “Ask anything.” This panel offers three suggestions the user never asked for. If closed, it returns after the next update, sometimes under a new name.

Synthszr Take: Every panel that returns after an update under a new name trains users in a single gesture: close. This conditioning is nowhere to be found in any AI budget, yet it costs the most because it tarnishes every future feature. If the product team delivers a genuinely useful feature the following quarter, it sits in the top right behind the same sparkle and is reflexively clicked away. Vendors that integrate AI into existing workflows without a sparkle icon retain this trust. The others squandered it on a roadmap slide.

Hackers heat up supermarket freezers across more than seven countries

According to Tech Brew, hackers switched supermarket refrigeration systems in seven European countries to defrost mode overnight. Temperatures rose to up to 122 degrees Fahrenheit—around 50 degrees Celsius—spoiling the food. It is the second attack of its kind in less than a month. In October, a nearly identical hack hit hundreds of French supermarkets, destroying inventory and reportedly causing fires in several locations. The CEO of one affected chain estimated losses on a single day at tens of millions of euros. The Pentagon is reportedly investigating whether six grocery stores on US military bases were also attacked this way in August. → Tech Brew

Synthszr Take: Manufacturers cut corners, retailers pay the price. A commercial freezer system operates far longer than its manufacturer provides security updates. Until now, this gap was an invisible cost advantage for the industry. A loss in the tens of millions in a single day makes it visible—on a supermarket chain’s balance sheet. In the future, insurers will tie their premiums to update commitments. When purchasing refrigeration technology next time, the duration of support will matter more than power consumption.

Seven out of ten early-career mathematicians doubt their academic future

Around 70 percent of young mathematicians now rate their chances of staying in academic mathematics worse than they did in the summer. That is according to an MIT survey published on October 29, reported on by Alberto Romero in his newsletter The Algorithmic Bridge. The survey questioned 268 early-career researchers from nearly 60 institutions, predominantly doctoral candidates and some postdocs. They were asked to indicate how the likelihood of an academic career following a doctorate or postdoc had changed since mid-July. Only 9 percent see their prospects improved. Romero links the shift in sentiment to recent advances by AI systems in mathematics and argues the thesis that mathematics will soon no longer be a human activity. → Alberto Romero from The Algorithmic Bridge

Synthszr Take: AI labs fish for their next generation of researchers in precisely this pond. Mathematicians with PhDs write the proofs, benchmarks, and verification procedures against which every model generation must be evaluated. If 70 percent of them write off their academic careers, this pool shrinks while OpenAI, Google, and others compete for the same talent. Prices for the remaining few rise. Romero’s thesis therefore has a built-in catch: A machine that replaces human mathematics requires humans who can verify its proofs.

MIT Uses GPT-6 Astra to Rewrite Qwen’s Weights, Boosting Math Score Sixfold

Researchers at MIT’s Computer Science & Artificial Intelligence Laboratory (CSAIL) have used OpenAI’s language model GPT-6 Astra to directly alter the weights of a small open-source model from Qwen. As a result, its mathematics performance reportedly increased sixfold. Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab present the work in a research blog post on November 2. They call the approach “Weight-Space Programming”. The test examined whether Astra can interpret and modify networks without calculating changes via traditional training procedures.

In the most striking example, Astra’s intervention raised the MATH-500 score of Qwen3.5-0.8B from 7.9 to 48.1 percent. Around 0.016 percent of the weights were modified. According to the researchers, the improvement stems primarily from correcting repetitive reasoning—meaning loops in which the model repeated itself. This interpretation shapes both the technical findings and the debate over their significance.

Methodologically, the approach differs from standard fine-tuning. There, weights are adjusted based on examples, feedback, or rewards via methods such as Gradient Descent that measure errors and compute corrections from them. In the MIT experiment, a separate model instead inspected the target network, analyzed its behavior, and proposed direct weight modifications. In some trials, Astra received the weights as text. For larger networks, it had access to a controlled Python environment.

The researchers themselves qualify the results. Their method also required compute, and in the Qwen experiment, practice problems as well as tests. Whether direct editing is permanently superior to conventional training in terms of speed, cost, or effectiveness has not been proven. Potential savings in time and computing resources remain unconfirmed. As an outlook, they outline systems that inspect, repair, and construct other neural networks—and may eventually do so to themselves.

Reactions have been mixed. Yuchen Jin, a technical staff member at Databricks not involved in the study, and Voyage AI researcher Chandan Singh expressed optimism on X. One skeptical commenter doubted that the results prove stronger model reasoning. Maxime Fournes, CEO of the organization Pause AI, warned that successful weight editing could accelerate dangerous advancements.

→ VentureBeat

Synthszr Take: The 0.016 percent of modified weights sounds like surgery, and surgery requires specialists. Astra suggests the cut. Someone still needs to verify why exactly these parameters end the repetitive loops and what they inadvertently shift along the way. Today, this judgment call is made by a handful of interpretability teams. Companies relying on open-weight models are therefore sticking with fine-tuning along with its tools, service providers, and evaluation routines. This market will only truly open up once a provider includes the verification process—and OpenAI already owns the editor model.