Claude Mythos Benchmarks Shock the Competition, but Meta Stays in the Game

77.8 on SWE-Bench: Claude Mythos Pulverizes the Competition

With Claude Mythos Preview, Anthropic is showing what happens when an AI model becomes too good for the public market. The unreleased frontier model, the centerpiece of Anthropic's Project Glasswing cybersecurity initiative, achieves 77.8% on SWE-bench Pro—a 24-point jump from the publicly available Opus 4.6 (53.4%). On SWE-bench Verified, Mythos is approaching the 94% mark, and in multimodal coding tasks, it doubles the performance of its predecessor from 27.1% to 59.0%. The reasoning benchmarks tell the same story: 56.8% on Humanity's Last Exam without tools, where Opus 4.6 gets stuck at 40%. Mythos Preview remains restricted to a closed group of security partners and large enterprises—Anthropic cites the dual-use risks of its cybersecurity capabilities as the reason. → AI Secret

Synthszr Take: Anthropic is holding back Mythos Preview because the Glasswing consortium needs to close the security gaps that Mythos itself uncovered during its own development. The restriction to security partners and large enterprises is a grace period. What the benchmark numbers show is a different caliber. SWE-bench Pro: 77.8% for Mythos, 53.4% for Opus 4.6. Multimodal coding: from 27.1% to 59.0%. Humanity's Last Exam: 56.8% without tools, Opus 4.6 at 40%. These are leaps that can hardly be explained by incremental improvements. Anthropic seems to have crossed the tipping point where models actively accelerate their own further development. The model improves itself. If what the numbers suggest is true, the development pace is accelerating from within, driven by the models themselves; that is the truly astonishing part.

Claude Managed Agents: Anthropic Becomes an Infrastructure Company

Anthropic is launching the public beta of Claude Managed Agents, a service that allows companies to develop and operate AI agents directly on Anthropic's platform. Previously, the company primarily offered models that users could use to build their own agents. With Managed Agents, Anthropic abstracts away the entire infrastructure complexity: sandboxing, authentication, credential management, and multi-hour execution times are already integrated. The pricing is transparent: standard API token prices plus $0.08 per active session hour. Advanced features like multi-agent orchestration and self-evaluation will remain in a limited research preview for now. → thenewstack.io

Synthszr Take: Anthropic is making the same move Amazon did with AWS 20 years ago: instead of just selling technology, the operational complexity becomes the product. The months of infrastructure work for production-ready agents (sandboxing, checkpointing, tracing) has been the bottleneck for enterprise adoption until now. Anthropic is turning this hurdle into a subscription model at $0.08 per hour. This is reminiscent of the early cloud days when companies realized they no longer needed their own data centers. The strategic move is clever: while OpenAI and Google fight over model performance, Anthropic is building the toll road on which all agents will travel. The real power lies not in the best model, but in control over the execution environment.

Meta Stays in the Game: Muse Spark and the Battle for the Superintelligence Stack

Meta has unveiled a new AI model, Muse Spark, after months of rumors about performance issues and delays. On Threads, CEO Mark Zuckerberg emphasized its strengths in areas like health, social content, shopping, and gaming—a clear signal of its consumer focus. The stock market reacted with a 6.5 percent increase in Meta's stock. The team under ex-Scale AI CEO Alexandr Wang has delivered after billions in investment. However, the real challenge is aimed at OpenAI: while Anthropic dominates the enterprise market, advertising is the only remaining growth driver for OpenAI. Meta has a decisive advantage here: 3.5 billion users on its own platforms and decades of expertise in the digital advertising business. → Martin Peers

Synthszr Take: Meta is playing a different game than everyone thinks. While OpenAI and Anthropic are fighting over frontier models, Zuckerberg is building the actual superintelligence infrastructure: a stack of model, distribution, and monetization that finances itself. The 3.5 billion users are not just reach, but also a gigantic RLHF lab where every click trains the next model. OpenAI's dream of an advertising business seems like a Michelin-starred restaurant trying to compete with McDonald's in the burger joint business. The real innovation in Muse Spark isn't in the benchmarks, but in its integration: a model that is “particularly good” at shopping, gaming, and social content isn't a chatbot—it's a transaction machine. Meta doesn't need to have the best model, just the most profitable one.

Chinese Providers Continue to Push Down Token Prices

The Chinese company Z.ai has released GLM-5.1, a language model that achieves 94.6% of Claude Opus's coding performance—at only one-third of the cost. The 744-billion-parameter Mixture-of-Experts model was trained entirely on 100,000 Huawei Ascend chips, without a single Nvidia processor. The model works autonomously for up to eight hours on individual programming tasks, from planning to testing. With a score of 58.4 on SWE-Bench Pro, it surpasses GPT-4 and Gemini 1.5 Pro. The weights are freely available under the MIT license. → Unwind AI

Synthszr Take: Z.ai is demonstrating what is known in the semiconductor industry as a “second source strategy”: alternative suppliers drive down prices and break up monopolies. The 94.6% formula sounds like a classic B2B sales pitch (“almost as good, but significantly cheaper”), but this is about more. Huawei's Ascend chips may be technically inferior, but they are obviously sufficient for training and inference. This is reminiscent of the PC revolution of the 80s: IBM-compatible computers were never better than the original, just good enough and available. When AI performance becomes an interchangeable commodity, costs and supply chains, not benchmarks, will determine market share.

Token Prices are Falling, but Training Costs are Not

Last week, Anthropic cut off access for third-party tools to Claude Pro and Max subscriptions. The reaction was predictable: developers protested, commentators spoke of revenue protection. Luo Fuli, who leads the MiMo language model team at Xiaomi, saw something else: the inevitable collapse of a pricing structure designed for chat interactions with a few hundred tokens but subsidizing agent workloads with 10 to 100 times the token consumption. One Claude Max subscriber generated over $5,600 in API costs in a single billing cycle on a $100 monthly fee—a subsidy ratio of 25 to 1. The structural cause: third-party frameworks like OpenClaw fire off multiple tool calls with over 100,000 tokens for each user request, compressing all three steps of the context and thereby destroying Claude's caching system. The same pattern is emerging in China: Alibaba's Pro plan is sold out by 9:30 AM daily, Tencent's slots are permanently unavailable, while throttled speeds and quota walls point to a market under extreme cost pressure. → Hello China Tech

Synthszr Take: The AI industry is currently experiencing its own version of the tragedy of the commons, only here the pasture is made of GPU cycles and the sheep are code agents. The business model resembles an all-you-can-eat buffet where competitive eaters suddenly show up: the statistical distribution between casual and heavy users, on which every subscription model is based, collapses when every user becomes a power user through inefficient frameworks. China's overheated coding plan markets, where developers set alarms and write auto-purchase scripts, show the consequence: artificial scarcity as an emergency brake against economic suicide. Luo Fuli's analysis hits the core: pain breeds engineering—only when the true costs become visible does the pressure for efficient context management and cache optimization emerge. The industry must shift from a token price obsession to a token efficiency discipline, otherwise its own promises of efficiency will eat the business model alive.

Neuro-Symbolic AI: The Next Breakthrough in Token Costs?

Researchers at Tufts University have developed an AI system that reduces energy consumption by up to 100x while improving accuracy. The neuro-symbolic AI combines neural networks with human-like logical reasoning. In the Tower of Hanoi test, the system achieved a 95% success rate (standard AI: 34%) and required only 34 minutes of training instead of 36 hours. The key: instead of learning through endless trial and error, the system uses abstract concepts like shape and balance. Professor Matthias Scheutz explains that VLA (Visual-Language-Action) models no longer have to guess which shadow belongs to which block. AI systems already consume over 10% of U.S. electricity production, and this demand is expected to double by 2030. → TAAFT - There's An AI For That

Synthszr Take: The Tufts researchers are reviving a debate from the 1980s: symbolic AI versus connectionism, this time with a twist. While OpenAI and Anthropic are betting on ever-larger models (GPT-5 is rumored to be able to power a small city), the team around Scheutz shows that hybrid architectures are the real innovation. This is reminiscent of the development of airplanes: the Wright brothers relied on aerodynamic principles rather than increasingly powerful engines. Neuro-symbolic systems could become the “functions-on-demand” of AI, where reasoning modules are switched on as needed for specific tasks. The 100x energy saving turns a cost problem into a competitive advantage.

XPeng Ditches Nvidia: China's Automakers Write Their Own Chip History

XPeng has completed its transformation: the revised version of the Mona M03, its best-selling electric car, now runs on its in-house Turing processor instead of Nvidia chips. The switch affects the mid- and upper-tier model variants with intelligent driving functions and marks the end of Nvidia dependency across XPeng's entire core fleet. The Mona M03 launched in August 2024 at prices between 110,000 and 150,000 yuan and reached over 200,000 deliveries by October 2025 in just 14 months. The Turing chips are also used in a vehicle jointly developed with Volkswagen, which is produced at Volkswagen Anhui. In parallel, competitor Nio reduced costs per vehicle by about 10,000 yuan through its own chips, as the company announced in June 2025. → Caixin Global

Synthszr Take: XPeng is demonstrating what is becoming the new playbook in the automotive industry: chip sovereignty as a competitive advantage. The analogy to Apple's M-processors is obvious, but this is about more than just margins. Chinese automakers treat their ADAS chips like Coca-Cola treats its formula: too important to outsource. The Turing deal with Volkswagen shows that XPeng not only wants to get rid of Nvidia but also wants to become a supplier itself. What began as risk mitigation (less dependence on U.S. technology) is mutating into a business model. The irony: Nvidia has prepared the market so well that its customers are now building it themselves.

Website Size is Becoming Irrelevant to Google — The Timing is Suspicious

Google released a podcast in which Gary Illyes and Martin Splitt explain why the growing size of websites is not a problem. The central thesis: Page Weight is an unreliable metric because measurement methods vary and “Excess Weight” is often useful content. Splitt distinguishes between pure HTML (Googlebot crawls a maximum of 2 MB), compressed data on the network (5–6 MB), and decompressed data on the user's end (10 MB). A 15 MB HTML document is acceptable if “pretty much most of these 15 megabytes are actually useful content.” The crucial question is not the absolute size, but the ratio of markup to content—where metadata for third-party tools or regulatory requirements could also be considered “useful content.” → STACKED MARKETER

Synthszr Take: Google is downplaying Page Weight at the very moment its own products are becoming increasingly bloated. This is reminiscent of tobacco companies funding studies on the health benefits of nicotine. The argument follows a familiar pattern in the platform economy: first, set the standards (2 MB crawl limit), then claim authority over the measurement (“there is no uniform definition”), and finally, legitimize your own violations (“useful content”). What Google is selling here as technical clarification is classic gaslighting: the reality of a median page size of 2.3 MB is obscured by semantic confusion (“compressed” vs. “decompressed”). The real reason for heavy pages is not useful content, but tracking scripts, ad-tech, and the endless cascade of JavaScript frameworks. Google knows this—and profits from it.