Reading

When Anthropic released Claude Fable 5.2 on November 9, it set its standard API price at $8 per million uncached input tokens and $40 per million output tokens, which is lower than OpenAI's GPT-6 Astra ($10/$50). This price cut was widely promoted as a cost advantage for Fable 5.2. However, The New Stack technical author Jessica Wachtel conducted independent tests using three coding tasks and found that a lower list price does not necessarily mean lower actual spending.

Wachtel ran each model once via the respective vendor APIs with default settings to test three tasks simulating real development work: fixing 10 planted bugs in a 2,600-line billing codebase, reviewing an EU invoicing package for 45 manually planted defects, and writing validation code from scratch for VAT IDs in 27 countries and IBANs in 34 countries. Each task was scored out of 50 by an automated script.

In the bug-fixing task, both models fixed all 10 bugs and passed all 50 hidden checks, each requiring 12 conversation turns. However, the processes differed significantly: Fable 5.2 consumed over 17,000 reasoning tokens, while Astra used only 819. Fable 5.2 cost approximately $2.51 and took 326 seconds, whereas Astra cost about $0.98 and took 122 seconds. For the same score, Fable 5.2 cost about 2.6 times more than Astra.

Code review was the only task where scores differed. Fable 5.2 scored 49, while Astra scored 45. Fable 5.2 identified 44 of the 45 defects with no false positives, while Astra identified 41 and incorrectly classified one Spanish ID as valid. However, Fable 5.2 spent $6.05 and 993 seconds to achieve this, compared to Astra's $0.95 and 227 seconds. Fable 5.2's extra 4 points came at a cost exceeding six times that of Astra and four times the time.

In the code-writing task, both models achieved a perfect score of 50, yet Fable 5.2 still cost 2.3 times as much as Astra. Across all three tasks, Fable 5.2's total cost was approximately $13.39, while Astra's was about $3.99—despite Fable 5.2's lower unit price, the total cost was roughly 3.3 times higher. Wachtel noted that Fable 5.2 consumed about five times as many tokens as Astra across these three runs.

These results corroborate OpenAI President Greg Brockman's statement at the Astra launch that price per task is key. Wachtel concluded that for such tasks, she would start with Astra, which achieved the same scores on bug fixing and code writing at about 41% of the cost; she would only consider Fable 5.2 when careful code review was needed, as it found three additional defects in that specific instance. She also acknowledged that both models missed some defects and that each model was run only once, so the four-point gap could shift in another run.

NVIDIA has introduced two new algorithms, Ozaki-II and 2M, in the latest release of its CUDA Toolkit 13.4.2 for the cuBLAS math library. These technologies aim to resolve a long-standing contradiction in high-performance computing: how to leverage the faster integer Tensor Cores—designed primarily for low-precision tasks—while maintaining the full accuracy of double-precision floating-point (FP64) calculations. This update allows scientific simulations dependent on high-precision dense general matrix multiplications (GEMM), such as those in molecular chemistry and materials science, to automatically gain speed without requiring any changes to application code.

Traditionally, GPU FP64 units offer high precision but limited throughput, while Tensor Cores provide extreme speed for lower-precision data. The new cuBLAS methods employ "floating-point emulation" techniques, decomposing each double-precision number into multiple integer segments, performing exact integer multiplications on the Tensor Cores, and reassembling the result to meet strict FP64 standards. Specifically, the Ozaki-II algorithm utilizes the Chinese Remainder Theorem (CRT) to significantly reduce the number of required integer multiplications compared to its predecessor, Ozaki-I. Meanwhile, the 2M algorithm optimizes complex matrix multiplication by reducing the operation from four real multiplications to two, thereby lowering computational overhead.

According to NVIDIA's official technical blog, these algorithms demonstrate substantial performance improvements in real-world applications. In quantum chemistry, DGEMM (real double-precision matrix multiplication) using Ozaki-II runs 5.5x faster than native FP64 on NVIDIA VR NVL72 systems and 3.9x faster on GB200. For Coupled Cluster Singles and Doubles (CCSD) calculations—considered the "gold standard" in quantum chemistry—the end-to-end speedup reaches 2x on VR NVL72 and 1.4x on GB200. When employing mixed-precision modes, these gains extend to 5x and 3.5x respectively.

In materials science, ZGEMM (complex double-precision matrix multiplication) within the BerkeleyGW software package achieves a 5x speedup on GB200 when using the 2M algorithm. Tests on the BerkeleyGW Epsilon module show an end-to-end speedup of 6.6x on VR NVL72 and 3.9x on GB200. Notably, all tests verified that results remain within FP64 accuracy bounds, and users do not need to manually adjust APIs or rewrite kernels, as cuBLAS automatically selects the optimal path.

Furthermore, the upcoming NVIDIA Rubin GPU architecture introduces the new TI16 numeric format and the tcgen05.mma instruction, which will further accelerate Ozaki-series computations. Existing users can enable these features simply by downloading and installing the latest CUDA Toolkit. Developers can refer to the floating-point emulation section of the cuBLAS documentation for detailed API control options, such as cublasSetEmulationStrategy(), to fine-tune the balance between precision and performance for specific workloads.

According to TechRadar, Russia's communications watchdog, Roskomnadzor, has reportedly officially deployed machine learning filters to choke virtual private networks (VPNs) across the country. This move marks a sharp escalation in the Kremlin's digital barrier. The state censor poured 2.27 billion rubles into the project, according to government digital transformation documents reviewed by Forbes Russia and reported by The Moscow Times.

Unlike previous attempts to decrypt user content, the newly deployed AI tool is a predictive model that searches for behavioral fingerprints in connection metadata: packet sizes, transmission intervals, session lengths, and hosting server patterns. Dmitry Kuteynikov, an expert at digital rights organization RKS Global, explained that the tool helps staff decide what to block and warned that as the model comes into use, blocks are likely to come faster, making VPNs less reliable.

A developer working for Russian provider Amnezia VPN confirmed that this kind of model could affect a wide range of protocols because it targets the general characteristics of VPN traffic rather than any single one. While obfuscation hides a protocol’s signature, the traffic pattern of a tunnel stays visible, which is exactly what a behavioral model looks for.

The automated filtering is already causing disruption across Russian business networks. Although Russian law technically exempts internal corporate VPNs, enterprise connections look nearly identical to circumvention tunnels under automated inspection. As reported by Kommersant, telecom sources noted that traffic analysis algorithms misclassify traffic in 5% to 6% of cases, causing unrelated websites and corporate cloud applications to collapse when entire hosting subnets are blacklisted.

Earlier roadmaps showed that Russia drafted plans for AI that could block 95 percent of VPNs. While official targets were quietly revised, funding continues to surge. The Kremlin is preparing to double its internet censorship budget in coming years, with another 10.2 billion rubles allocated to upgrade deep packet inspection boxes in 2026 alone.

The departure of Nvidia CEO Jensen Huang has triggered widespread reactions across Asia. Some governments have moved quickly to confirm chip supply commitments, while others are monitoring the impact on local projects or addressing ongoing investigations.

In South Korea, the Ministry of Science stated that Nvidia had reaffirmed its plan to supply 260,000 advanced GPUs to the country and to continue joint AI projects. Samsung Electronics and SK Hynix, major suppliers of memory chips for Nvidia's AI processors, maintain close ties with the company. Additionally, South Korea's government AI strategy council held an emergency meeting on Friday to assess the potential impact of Huang's departure on the country's AI industry.

In Singapore, the Ministry of Home Affairs confirmed that the investigation into Alice Huang, former CEO of Megaspeed, remains ongoing. Authorities seized her passport in June 2025 to investigate possible false declarations regarding server purchases. In February 2026, police found insufficient evidence to establish offenses related to these declarations but noted she knowingly provided inaccurate information about her role at Megaspeed. She was issued a warning and barred from re-entering Singapore. Singapore also mentioned receiving limited assistance from U.S. authorities despite repeated requests for information about server maker Supermicro.

In China, the Foreign Ministry rejected allegations linking a Chinese businesswoman to espionage as "groundless." A Reuters review indicated that Huang's exit has raised uncertainty over more than $3 billion in Chinese AI computing projects reliant on Nvidia's restricted processors. Many of these chips are subject to U.S. export controls, making them difficult for Chinese companies to obtain.

In Taiwan, Economy Minister Kung Ming-shin said Huang's departure was unlikely to significantly change the growth of the global AI industry. The Taipei city government stated that Nvidia's planned overseas headquarters in the city remained on track despite the leadership change, noting the company has met its contractual obligations so far.

In Japan, the Industry Ministry said the government would closely monitor Nvidia, describing the chipmaker as an important player in Japan's AI semiconductor industry. Deputy Chief Cabinet Secretary Masanao Ozaki said it was too early to assess the impact on the global AI industry, while stressing Japan's commitment to ensuring a stable supply of semiconductors.

Traditional lakehouse formats split metadata into numerous manifest files scattered across object storage. For every query, the engine must traverse these files to determine which Parquet data files to read. Pedro Holanda, lead developer at DuckLabs, argues that this is essentially a relational database problem that does not need to be solved using file system methods.

His explanation is straightforward: you issue a query, and the server answers with which Parquet files to read. The Parquet files themselves do not change; only the catalog changes. This is why DuckLake stores table metadata in a standard SQL database rather than scattering manifest files. Users are already familiar with SQL and schemas, so they do not need to learn an additional logic for traversing metadata.

DuckLake 1.1 was released in October alongside DuckDB 2.0.0 LTS, bringing file compaction and inline/flush mechanisms tailored for streaming data lakes. Regarding file compaction, when the number of files becomes excessive, DuckLake can merge multiple files into one to reduce request counts; this operation can span multiple snapshots, further compressing the total file volume. Regarding inlining, actual data representations are stored directly as tables within the catalog. Holanda stated that while building DuckLake, they realized that since they already had a database system and a catalog, why not store the data itself in tables?

This resolves the small-file problem in streaming scenarios. When multiple data sources write simultaneously, traditional lakehouses generate a large number of fragmented files. Queries then require traversing metadata and reading many small Parquet files, slowing down both processes. DuckLake absorbs these fragments into the catalog through inlining and flushing, deferring compaction operations until they are needed.

Weston Pace, an engineer at LanceDB, supplemented DuckLake's design trade-offs from the perspective of concurrent writes in another technical blog post. He categorized table-format coordination mechanisms into two types: zero-infrastructure atomic operations (relying on cloud storage's 'put-if-not-exists' semantics, requiring no extra services but limited to approximately ten commits per second) and centralized coordination (where DuckLake uses a Postgres instance as a commit coordinator, synchronizing writers via mutexes or CAS primitives). Centralized schemes may achieve higher transaction rates and apply to filesystems that do not support atomic creation, at the cost of maintaining a long-running service.

Both DuckDB and DuckLake are open-source under the MIT license, overseen by the independent DuckDB Foundation. DuckLake can run without massive server clusters, partly because the Parquet and DuckDB storage formats are naturally similar—both are columnar and organized by row groups, minimizing data movement steps.