AI Frontier Daily
Anthropic Unveils Claude Fable 5.2 and Mythos 5.2 with Major Coding and Reasoning Gains
Today's AI Intelligence Briefing: 12 distinct events rigorously curated from 13 cross-verified sources. Frontier models continue to advance hybrid reasoning and multimodal synthesis, while open-source tooling and autonomous agents accelerate real-world execution. Reports on the same event are merged into a single entry, and covered events do not reappear in later issues.
Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, two models built on the same underlying architecture with significant gains in coding, reasoning, and vision. Fable 5.2 is available to all users with built-in cybersecurity and biosecurity safeguards, while Mythos 5.2 opens to safety and life-science organizations through a verification program, cutting sandbox escape rate from 39% to 3.1%.
In pre-release tests, Fable 5.2 showed strong technical capabilities: 1,000 agents formalized Fermat's Last Theorem in 30 hours, versus 11 days for the first human formalization, with proof length reduced to one-seventh. The model also completed previously infeasible software migrations, designed three working electronic devices from scratch, and solved two open problems in black hole physics.
Commercially, Fable 5.2 input pricing drops to $8 per million tokens, response speed improves 30%, and it supports zero data retention (ZDR) for enterprise compliance for the first time.
Frontier Models & Architecture Breakthroughs
3 picksTracking flagship foundation models, novel architectures, and reasoning leaps.
Meta Launches Muse Ember Lightweight Model with Input Price as Low as $0.10 per Million Tokens
Meta Superintelligence Labs has released Muse Ember, its lightest and most cost-efficient model yet, designed for high-frequency agent tasks including tool calls, coding subtasks, routing, information extraction, and long-document parsing. Input pricing is as low as $0.10 per million tokens and output $0.50 per million tokens, with 1M token context and native multimodal understanding.
Meta benchmarks show it outperforming Anthropic Haiku and OpenAI Luna on cost and intelligence on Artificial Analysis, competing on price with open-source rivals like DeepSeek V4.1 Flash and GLM 5.3 Flash. Ember is positioned as the agent execution layer paired with the larger Muse Wonder model, which handles planning while the lighter model handles high-frequency execution.
Cross-verification shows Muse Ember scores 36 on the Artificial Analysis Intelligence Index, close to Claude Haiku, and tops Code Arena WebDev with 1608 points, surpassing Claude Haiku 5.5 and GPT-6 Luna as the strongest small model. API pricing is identical across the three, but Ember's median cost per task is $0.03, about half of Luna and one-third-point-seven of Haiku.
Meituan Open Sources 1.6T-Parameter LongCat-2.5 Model
Meituan's LongCat team has open-sourced LongCat-2.5, a native multimodal model with 1.6 trillion total parameters and about 48 billion activated per token, supporting million-level context under the MIT license for commercial inference and further training.
The model quantizes visual signals into multi-level discrete tokens that share backbone parameters with text, rather than using an external vision module. It also uses W4A4 4-bit quantization to adapt to domestic AI chips with limited memory and bandwidth.
ByteDance Releases Seedream 5.5 with Native Transparent Background Output
ByteDance's Seed team has released Seedream 5.5, an image editing model whose core breakthrough is native support for transparent backgrounds with an alpha channel and pixel-level invariance in non-edited regions. Official demos show details remain precisely aligned after 100 consecutive edits of the same person, significantly reducing image drift in multi-turn editing.
Technically, the model finely processes contour edges and transparent areas, preserving difficult edges like feathers, so generated elements can enter design workflows as standalone assets without post-processing cutouts. Serious text error rate drops from 2.8% to 0.4%, and editing a 2K square image costs about 0.28 yuan, less than half of Seedream 5.0 Pro.
The company says it outperforms GPT Image 2.5 Sunburst in dense small-text rendering and realistic lighting and materials. The model is rolling out on Jimeng, Doubao, and Doubao Work, and developers can call it via Volcano Ark as Doubao-Seedream-5.5, with a Pro version planned.
Industry Ecosystem & Compute Economics
4 picksCapital deployment, compute supply chains, and industrial transformation.
AI Startup Proves Quantum Computing Conjecture Using Private Claude Model
AI for Science startup Physical Superintelligence used an internal research system running an undisclosed Anthropic model to find a complete proof within three days for a quantum computing speedup conjecture open since 2008, posed by Scott Aaronson and Andris Ambainis, verified by the Lean proof assistant.
This is a landmark case of an AI system autonomously expanding the boundaries of human knowledge, and it signals that synthetic researchers may produce results faster than humans can understand them. The conjecture concerns when quantum computers can achieve exponential speedup.
DeepX Tapes Out Second-Gen Edge AI Chip Running 100B Models Under 5 W
South Korean edge AI chip startup DeepX has taped out its second-generation AI accelerator, DX-M2, which uses Samsung's custom processing-in-memory (PIM) technology to run models with up to 100 billion parameters on-device within a power budget under 5 W, at 10-30 tokens/s.
Compared with the already mass-produced vision-focused first-generation DX-M1, the key leap in DX-M2 is triple the memory bandwidth, targeting generative AI edge devices such as robots and drones. DX-M2 supports models with 20 billion to 100 billion parameters (about 3 billion to 20 billion activated).
OpenAI Builds Control Room to Halt Model Workloads in 35 Seconds
OpenAI has disclosed Control Room, a central control system in its research infrastructure that uniformly tracks all inference workloads including training, evaluation, and data processing, and can stop model output within 35 seconds of a safety decision.
The system integrates with misalignment monitoring and has fault isolation, triggering a stop even if it goes down itself. It is now fully deployed on the research side.
IN BRIEF
Agentic Systems & Autonomous Workflows
4 picksMulti-agent orchestration, tool usage, and complex task execution.
Berkeley-Led Study Finds Self-Evolving Agent Harnesses Lose Solved Tasks
A new paper from UC Berkeley and other institutions finds that letting agents rewrite their own prompts, skills, tools, or code to improve benchmark scores can mask the loss of tasks they previously solved.
Testing 4 evolution methods with GPT-5.4 on 2 scaffolds and 2 benchmarks, all 16 evolved agents showed trade-offs: one improved reliability by 5.2 points but lost 10 original tasks, and 8 of 11 more reliable agents saw higher per-run costs.
The authors recommend tracking lost tasks, running costs, and performance on unseen tasks rather than just average scores.
Baidu Unveils ERNIE Evolve to Auto-Evolve Training and Inference Code
Baidu has released ERNIE Evolve, an automated research system that uses a director plus parallel evolution trees with independent evaluator and reviewer loops to automatically edit training and inference code, keeping only verified gains.
The system achieved 0.8763 BPB on NanoChat AutoResearch and 38.9 seconds on NanoGPT Speedrun Track 1, and boosted inference speed on Kunlun P800 chips to 623.6% of baseline. The system itself is not open-sourced; only final benchmark code and run logs are public.
UC Berkeley Supervisor Rebuilds Coding Agent Teams to Beat Expert GPU Kernels
A new UC Berkeley paper proposes Evolving Agent Teams, where a supervisor diagnoses stagnation bottlenecks in coding agent teams and dynamically reorganizes roles such as profiler, debugger, and knowledge integrator, with a shared file system to accumulate cross-team experience.
On an MLA decoding kernel optimization task, the method reached 0.74x FlashInfer speed under the same token budget, versus 0.31x for a fixed 4-role team; with more budget it reached 1.21x and won 36 of 47 production workloads.