U.S. spot Bitcoin ETFs saw a net inflow of $174.6 million yesterday.

Beating AI News Brief: Microsoft AI has launched MAI-Image-2.6-Flash, now available in public preview on Microsoft Foundry. The model is an accelerated variant of its flagship MAI-Image-2.6, supporting text-to-image, image editing, multi-image reference, real-time web information access, and automatic aspect ratio selection, with a core focus on cutting generation latency and costs. Speed is its key selling point: MAI-Image-2.6-Flash generates images 2.8 times faster than GPT-Image-2 Medium, with 72% higher GPU utilization efficiency. That said, it does not trade off image quality solely for speed. In Artificial Analysis’s current image editing leaderboard, MAI-Image-2.6-Flash ranks third with an Elo score of 1311, slightly higher than GPT Image 2 high’s 1309; it places eighth in text-to-image with an Elo of 1297. The flagship MAI-Image-2.6 performs better overall, currently ranking first in image editing and second in text-to-image respectively. On Microsoft Foundry, MAI-Image-2.6-Flash is priced at $19 per million tokens for image output, while the flagship MAI-Image-2.6 costs $38 per million tokens. After adjusting for representative images per Artificial Analysis, Flash’s cost is roughly $19.5 per thousand images, compared to GPT Image 2 high’s approximately $211 per thousand images.

Beating AI Insight Flash News: GitHub has added a multi-model orchestration system called HydraFusion to Copilot. The system first determines how to handle a task before calling different models, operating in three modes: simple tasks are assigned to a single model for direct completion; complex tasks start with a lower-cost model, which is escalated to a more powerful model if the result is unsatisfactory; for tasks requiring review, one model completes the work first, while another model from a different family is tasked with identifying errors, after which the first model revises its output. GitHub compared HydraFusion with Claude Opus 5 across three programming agent benchmarks: it scored 4.9 percentage points higher than Claude Opus 5 on TerminalBench 2.1, 1.5 points lower on DeepSWE, and only 0.1 points lower on CheckpointBench (nearly on par); meanwhile, cost reductions reached 67%, 36%, and 65% respectively across the three benchmarks. This approach is very similar to Sakana AI’s Fugu. Both shift the focus from "which model to choose" to "how to organize multiple models". The key difference is that Fugu is a trained orchestrator model that learns to call different agents and can even recursively invoke itself; HydraFusion, by contrast, currently only selects among three fixed execution modes: Single, Cascade, and Critique, and is essentially a multi-model scheduler integrated directly into Copilot. HydraFusion is now available as a research preview for all GitHub Copilot plans, and can be enabled via the experimental features in Copilot CLI. GitHub also noted that the system is currently best suited for single-turn programming tasks that can be clearly defined at once, with multi-turn long tasks still under optimization.

Beating AI Express: Since OpenAI launched GPT-6 Astra on September 3, multiple model benchmark evaluation metrics have been continuously adjusted. Some changes have boosted Astra’s performance, while scores of some competing models have dropped, sparking external skepticism about AI "benchmark manipulation" and evaluation transparency. Specifically, Astra’s hallucination rate was once lowered from 4.2% to 2%, while GPT-5.6 Sol’s rate fell from 12.2% to 9.4%, before both rates returned to 4.2% and 12.2% respectively. In math evaluations, Anthropic’s Fable 5.1 score also dropped from 87.8% to 78%, and has since rebounded to 83%; GPT-5.6 Sol’s score fell from 83% to 80.5%, then returned to 83%. Additionally, Astra’s score in the ARC-AGI-3 benchmark rose from 98.6% in its pre-release draft to 99.99% on the final page, while its programming evaluation score was also slightly adjusted upward from 57.7% to 57.9%. OpenAI stated that evaluation results are influenced by factors including model version, tool configuration, inference level, and test runs, adding that the adjustments were made to ensure the data more accurately reflects the model’s optimal performance. However, Stanford University researchers argue that frequent re-running of evaluations may involve so-called "benchmaxxing" – the practice of adjusting test conditions to maximize benchmark scores. Industry insiders note that as competition among AI models intensifies, evaluation data has become a key tool for measuring model capabilities and competing for market share, with improving the transparency and reproducibility of benchmark tests drawing growing attention.

Beating AI News Flash: According to Reuters, citing people familiar with the matter, AI firm Anthropic’s IPO plans are being adjusted to mid-October. Roadshows could kick off as early as mid-October, with the listing targeted to complete before the U.S. midterm elections in November. The company had originally planned to publicly release its IPO prospectus as early as next week, but that timeline is now pushed back to the end of September. This timing shift means Anthropic could aim for a listed valuation of up to roughly $2 trillion. If achieved, the IPO would rank among the largest in history and serve as a key case for capital markets to test the AI industry’s growth potential. Sources also revealed that Anthropic is finalizing a $15 billion revolving credit facility, with participating bank analysts expected to meet with the company ahead of the prospectus filing. Morgan Stanley, Goldman Sachs, JPMorgan Chase, and Citigroup are among the institutions involved in preparing for the IPO. The offering is expected to draw significant market attention and could proceed around the same time as potential listing plans of other AI firms like OpenAI.

Tesla’s stock price plunged 5.92% in Friday’s close, wiping roughly $88 billion (about 591 billion yuan) off its market capitalization in a single day. While Tesla has positioned its Robotaxi as the next growth engine, the commercialization of its Cybercab is facing a “double whammy” of regulatory scrutiny and market expectations. The U.S. National Highway Traffic Safety Administration (NHTSA) has launched an investigation into around 1,000 Cybercab units, focusing on the compliance procedures and technical data Tesla used when it determined certain federal motor vehicle safety standards did not apply. Since the Cybercab lacks a steering wheel, accelerator, brake pedal and traditional side mirrors, the probe could impact its subsequent deployment timeline and expansion of operating regions. Meanwhile, Tesla’s Cybercab commercialization launch event held in Austin was deemed “below expectations” by Wall Street: it lacked specific deployment scale, production targets and timelines, and Elon Musk himself did not attend. Analysts from Evercore ISI, Wells Fargo, Barclays and JPMorgan all noted the event lacked sufficient new details, which could put pressure on Tesla’s stock in the short term. Still, Wall Street has not completely shifted its stance on Cybercab’s long-term potential. JPMorgan forecasts Tesla’s Robotaxi fleet could expand to around 9,000 units by the end of 2027, while the Royal Bank of Canada (RBC) projects the number of Cybercabs in the U.S. market will reach roughly 40,000 by 2030.

According to Lookonchain monitoring, after Shivon Zilis reposted a post containing the SLINK contract address and Elon Musk subsequently replied, SLINK's market cap surged to $80 million at one point. After Shivon deleted the relevant post, SLINK's price plummeted more than 80%. Analysts noted that in this market movement, over 30 suspected insider wallets bought SLINK in advance and sold quickly, reaping a total profit of over $4.7 million, with the highest return exceeding 1,200 times. Meanwhile, numerous traders who followed the trend to buy suffered losses.