•5 min read
Claude Code Weekly Limits After September 14: What Actually Changed
Anthropic ended a 50% Claude Code weekly-limit promotion and kept a permanent 25% uplift. The practical change is smaller than a '25% cut' headline suggests.
House analyst byline
Plutonous
Plutonous is the LLM Rumors house byline for research-driven analysis shaped by the editorial desk. It keeps the site accountable to one consistent voice while letting individual tips, edits, and research contributions stay behind the piece.
88 published stories
Plutonous Archive
All news
•6 min read
OpenRouter Batch API: Half-Price Inference Has a Deadline, a Data Trail, and a Catch
A worked guide to OpenRouter Batch API costs, deadlines, ambiguous submissions, result reconciliation and data retention.
•6 min read
Siri AI Is Missing in the EU on iOS 27. Here Is the Availability Test
A precise check for Apple Intelligence, Siri AI, language, device, platform and EU availability before you expect the new assistant.
•11 min read
Does Apple Intelligence Work Offline? Where Your iPhone Sends the Request
Separate on-device AI, Private Cloud Compute and ChatGPT. Check what works offline, inspect Apple's request report and understand iOS 27 availability.
•10 min read
Claude Prompt Caching: When Five Minutes Beats One Hour
Calculate Claude prompt caching's break-even point, choose a TTL from actual reuse, and diagnose misses without confusing API savings with subscription limits.
•11 min read
DeepSeek V4.1 Flash Migration: The API Contract Checklist Agents Need
DeepSeek V4.1 Flash changed aliases, vision support and API behavior. Use this migration checklist to test cache boundaries, tool reasoning, output limits and user isolation.
•11 min read
METR Time Horizons Explained: What 50% Reliability Actually Buys
METR's time horizon measures human task difficulty, not an agent's unattended runtime. Here is how to read the uncertainty and design a business acceptance test.
•10 min read
OpenRouter Free Models: Limits, Privacy and Fallbacks Explained
What OpenRouter's 50 and 1,000 daily free-request limits actually buy, why paid and free privacy settings differ, and how to avoid surprise fallback costs.
•13 min read
Astra's Access Problem Is Becoming an Infrastructure Business
The most-read Astra stories are not only about intelligence. They are about shared allowances, queue design and who gets dependable access to a frontier model.
•12 min read
PrismML's Tiny Model Bet Is Really a Fight Over Where AI Runs
PrismML packages Bonsai 2 27B as a 5.95 GB GGUF language model or an 8.60 GB MLX multimodal release. The deeper story is how ternary weights change local AI economics.
•11 min read
TypeSafe AI's Jev Is Not a Chatbot. It Is a Decision Engine
TypeSafe AI's Jev model returns typed decisions and calibrated probabilities instead of prose. Here is where that design could change software automation, and where the claims remain unproven.
•9 min read
Union Alpha Was Pareto 26.9: The Mystery Model Was a Multi-Model System
Union Alpha has been revealed as Circuit & Chisel's Pareto 26.9. Its architecture, pricing and uneven benchmark profile show why the product is the orchestration layer, not a secret foundation model.
•6 min read
Claude Opus 5.1 or 5.2? What the Astra Rival Rumors Show
Claude Opus update rumors now include 5.2 as well as 5.1. What is sourced, what remains speculation, and what a meaningful everyday rival to Astra would need.
•6 min read
Why GPT-6 Astra Uses Your Codex Allowance So Fast
Reduce avoidable Astra usage in Codex by checking Fast mode, matching reasoning to the task, controlling context and delegating deliberately. Keep plan limits, credits and API bills separate.
•6 min read
GPT-6 Sol Rumors: Astra's Next Test Is Everyday Work
GPT-6 Sol launch rumors point to OpenAI's next model question: can Astra-level progress become a practical everyday tool? Dates, pricing and performance remain unconfirmed.
•6 min read
iPhone 18 Pro Battery Life: What 36 and 45 Hours Actually Mean
Apple rates US eSIM-only iPhone 18 Pro models at up to 36 and 45 hours of video playback. Germany lists 34 and 43. Here is what those controlled ratings do and do not tell a buyer.
•5 min read
Union Alpha on OpenRouter: What the Free Mystery Model Offers
Union Alpha is a free anonymous OpenRouter model with image input and tool support. Its listing is real; links to GPT-6 Sol or a new Claude Opus remain unproven.
•5 min read
OpenAI Is Selling Astra to Finance as a Chain of Evidence
ChatGPT for Financial Services pairs GPT-6 Astra with licensed data, citations, templates, and controls. The product’s real test is whether every conclusion remains reviewable.
•6 min read
Three Papers Credit GPT-6 Astra. Their Proof Claims Need Different Tests
A close reading of three new arXiv papers separates their mathematical claims, their authors' AI-use statements, and the evidence readers need before calling AI-assisted mathematics established.
•13 min read
Dario, METR and the AI Slowdown: Who Gets to Inspect the Frontier?
Amodei's new pacing proposal turns AI safety from a lab promise into an argument about who gets to inspect the lab. The split between embedded evaluators, competitor review, and regulatory coordination is the real fight.
•10 min read
Who Really Runs METR? Part 1: Leadership, Funding and Independence
METR has disclosed meaningful safeguards, but its public record leaves key questions about current governance, non-cash dependence and publication rights unanswered.
•23 min read
Who Really Runs METR? Part 2: The People, EA, e/acc and the Power to Evaluate
A source-led examination of the people who help shape METR's evaluations, its historical ARC links, and the distinct intellectual networks often collapsed into one story.
•8 min read
DeepSeek V4.1 Flash Turns Agent Memory Into a Pricing Weapon
DeepSeek-V4.1-Flash pairs 1M-token context with a $0.006-per-million peak cache-hit price. Explore the architecture, API changes and economics of reusable agent context.
•7 min read
OpenAI Habitat: The Data Layer Behind ChatGPT, Explained
What Habitat actually does, explained through a railway station: fetching application data, directing requests and keeping slow dependencies from holding everything up.
•10 min read
A20 Pro vs A19 Pro and Android: Why Memory and Cooling Matter
Compare A20 Pro with A19 Pro, Snapdragon 8 Elite Gen 5 and Tensor G6: CPU, GPU, Neural Engine, memory bandwidth and cooling, with vendor claims labeled.
•7 min read
GPT-6 Astra on ChatGPT Plus: Access, Limits and Resets
A practical guide to GPT-6 Astra access on ChatGPT Plus, including Work and Codex allowances, reset windows, effort settings, Fast mode, and troubleshooting.
•9 min read
How a Phone Chip Actually Runs AI: The A20 Pro Mental Model
Explore an interactive 3D phone chip and learn how CPU, GPU, Neural Engine, memory, quantization and cooling work together to run AI on your device.
•10 min read
iPhone Duo and AirPods 5: The Small-Model AI Opportunity
iPhone Duo starts at $1,999. Explore US release dates, AirPods 5 pricing, A20 Pro and what Desert Ant and Liquid AI mean for on-device apps.
•7 min read
OpenAI's Navier-Stokes Claim: The Math, The Stakes, The Dispute
What Navier-Stokes means, what OpenAI's forced blowup proof claims, and why the dispute over credit and research data matters for AI science.
•7 min read
Anthropic's Global Workspace Makes Silent Reasoning More Inspectable
Anthropic's J-space research gives auditors a new way to inspect silent intermediate reasoning. Mixed replication results show why an internal readout still needs a behavioral test.
•7 min read
Astra's Spectrogram Demo: Why a Picture Can Reveal a Sound
Show Astra a picture of sound, and it can offer a surprisingly specific guess. Here is how spectrograms make that possible, and what it would take to turn the demo into a useful tool.
•10 min read
A Secure Sandbox Can Still Run an Insecure Agent
E2B, Daytona, Modal and Cloudflare expose different controls for agent execution. The harder boundary is what a contained agent is authorized to do.
•10 min read
AI Labs Are Building Data Factories. The Scarce Resource Is Proof.
How AI labs collect web data, generate synthetic examples, and turn agent actions into training signals. The bottleneck is verifying what those examples teach.
•8 min read
HydraFusion and Fugu: The Model Picker Is Becoming the Product
GitHub HydraFusion, Sakana Fugu, and OpenRouter Fusion point to a market where the durable advantage is the runtime policy that chooses, escalates, critiques, and synthesizes models.
•7 min read
Six Things People Built With GPT-6 Astra: Trains, Games, And Personal Apps
From editable steam trains to a playable portfolio and an iPod-style Mac app, Astra's early builders are turning personal ideas into software worth exploring.
•9 min read
GPT-6 Astra Turns The AI Agent Race Into An Authorization Race
OpenAI's GPT-6 Astra can operate software, carry long-running context, and tackle harder cyber work. The real contest is now the system around the model.
•8 min read
Grok Bot Shows Why Product Agents Need An Authority Model
Grok Bot turns AI from a chat window into a persistent operator. That makes its convenience real, its backlash rational, and authority design the next product battleground.
•13 min read
The Inference Race Is Moving Into GPU Kernels
Wafer, RunInfra, OpenRouter's provider price war, and AI-written GPU kernels are turning inference optimization into the next strategic control plane.
•9 min read
Tencent Hy4 Makes Open-Weight AI Too Important to Ignore
Tencent's Hy4 preview pairs a 770B-parameter sparse backbone with Apache-2.0 weights. Together with Z.ai's GLM releases, it shows why open-weight AI is becoming a strategic market, not a side project.
•11 min read
GLM-5.3 Shows How China's Open-Model Flywheel Could Outbuild America's Closed Labs
Z.ai's GLM-5.3 and GLM-5.3-Flash show how Chinese labs can compound public weights, DeepSeek and Moonshot research, shared RL stacks, and lower-cost deployment.
•10 min read
OpenAI's Hugging Face Incident Was a Systems Failure, Not an AI Escape
OpenAI's postmortem describes 1,200 agents, 70,000 messages, and a real Hugging Face intrusion. The lesson is about containment, incentives, and agent operations.
•10 min read
Ox Alpha Was Z.ai: Inside the GLM-5.3-Flash Stealth Test
Z.ai revealed that Ox Alpha was an early GLM-5.3-Flash preview. The anonymous experiment drew 343.5 million OpenRouter requests before the company attached its name.
•13 min read
OpenAI's Jalapeño Is Not an NVIDIA Killer. It Is an Inference Warning
OpenAI says Jalapeño delivers up to 1.9x more work per watt and 3.6x lower latency than selected Blackwell systems. The real threat to NVIDIA is inference share and pricing power, not a 2026 GPU collapse.
•11 min read
Ox Alpha: The Mystery Model That Took Over OpenRouter
Ox Alpha reached 70.3 million OpenRouter requests and 5.93 trillion prompt-plus-completion tokens in one day. Its maker is still anonymous, and the fine print matters.
•8 min read
SpecPTC: The Agent Harness Is Becoming a Runtime
SpecPTC launches safe tool calls while agent code is still streaming. Alex Zhang reports 1–1.2x RLM gains, but the bigger story is runtime scheduling.
•9 min read
Grok 4.6 And Grok Bot: xAI's Real Upgrade Is The Work Loop
Grok 4.6 reaches a 61 Intelligence Index score at $2/$6 API pricing. Grok Bot adds persistent computer use, shared state, routines, and a new governance problem.
•13 min read
Recursive Agent Harnesses Are the New AI Moat
Prime Intellect's recursive harness framing explains why AI agents are becoming an orchestration, verification, and training-data business, not just a model race.
•17 min read
Qwen3.8-Max Turns an Open-Weight Promise Into a Datacenter Strategy
Alibaba's 2.4T-parameter Qwen3.8-Max pairs 1M context, $2/$6 API pricing, and vendor-reported agent gains with a promised checkpoint few teams could deploy themselves.
•10 min read
Reasoning Effort Is AI's New Inference Control Plane
Reasoning effort is not a universal quality slider. It is becoming the control plane that allocates tokens, latency, tools, and verification across AI workloads.
•12 min read
TileRT: The Runtime Turning AI Speed Into A Product Moat
TileRT is not just a faster inference engine. Separate Xiaomi MiMo and Z.ai deployments show why ultra-low-latency runtimes are becoming a new battleground for frontier AI products.
•9 min read
Kimi K3 Is the Open-Model Champion in Waiting
Kimi K3 pairs a 57.1 Artificial Analysis score with 2.8T parameters, 1M-token context, $0.94 task cost, and open weights promised for July 27.
•10 min read
Meta's Muse Spark 1.1 Is Not A Catch-Up Model. It Is A Paid Agent Platform
Meta Muse Spark 1.1 API pricing, 1M-token context, benchmarks, coding agents, and what Meta's paid agent platform means for developers.
•15 min read
How KANs Work: Why Learnable Edges Survived the MLP-Killer Hype
Kolmogorov-Arnold Networks replace scalar weights with learned functions. Two years of evidence show where KANs work, where they fail, and why the idea survived.
•10 min read
Loop Engineering: The New Job Is Designing When AI Must Stop
Loop Engineering turns the hidden management work around coding agents into software: triggers, scoped execution, independent verification, durable state, budgets, and explicit stop conditions.
•11 min read
Grok 4.5's 16-Point Leap Makes xAI A Frontier Supplier
Grok 4.5 combines a 54 Intelligence Index score, 90-token-per-second measured speed, $0.31 benchmark task cost, Cursor-trained agent behavior, and live search in xAI's strongest model release yet.
•9 min read
DeepSpec: DeepSeek Open-Sourced the Inference Cost War
DeepSpec turns speculative decoding from a hidden serving trick into an open training stack, with DSpark claiming 60% to 85% faster V4-Flash generation.
•14 min read
OpenAI's GPT-5.6 Sol, Terra, And Luna Make Access The New Moat
OpenAI's GPT-5.6 Sol, Terra, and Luna launch is not just a model update. It is a preview of AI releases where capability, price, safety, and government access are bundled together.
•9 min read
Huawei And Intel Just Made The Chip Race A Systems War
Huawei's Tau Scaling Law and Intel's 18A-P roadmap show the same semiconductor shift from opposite sides: future chips will be won through systems, not node names alone.
•11 min read
Recursive's AutoResearch System Turns AI Research Into A Factory
Recursive's automated AI research system is not just a benchmark win. It is a preview of research loops that propose ideas, write code, run experiments, validate results, and keep going.
•10 min read
Open Models Are The New Linux: DeepSWE And The Infrastructure War Against Closed AI
DeepSWE shows closed labs still lead frontier coding agents, but open-weight models are starting to price the infrastructure layer. That is exactly how Linux won.
•14 min read
DiffusionGemma: The Open-Weights Answer To Inception's Real-Time Subagents
DiffusionGemma is not just Google's 4x faster text generation experiment. It is the open-weights counterpunch to Inception's closed Mercury 2 thesis for real-time AI subagents.
•9 min read
SpaceX Buying Cursor Is Not About The IDE. It Is About Owning The Coding Model Loop
SpaceX's $60 billion stock deal for Cursor turns a coding editor into strategic AI infrastructure. The real story is Composer, Colossus, Grok, and the race to own developer work.
•9 min read
Claude Fable 5 Got Banned By The US Government. That Is The Real AI Policy Story
The US government ordered Anthropic to suspend Fable 5 and Mythos 5 access for foreign nationals, forcing a global shutdown. The real story is frontier AI becoming controlled infrastructure.
•15 min read
Claude Fable 5: Anthropic's Frontier Model Is A Capability Rationing Test
Claude Fable 5 is not just Anthropic's strongest public model. It is the first Mythos-class release built around public access, fallback routing, data retention, and trusted capability gates.
•11 min read
iOS 27: Apple's Siri AI Reset Is A System Agent Bet
iOS 27 makes Siri AI a private system agent, not just a smarter assistant. Apple's update ties app actions, Camera mode, speed gains, Liquid Glass, and the EU delay into one AI control-plane strategy.
•13 min read
GTC Taipei: NVIDIA Turned The GPU Keynote Into An AI Factory Rollout
NVIDIA's GTC Taipei at COMPUTEX 2026 was not a normal product keynote. It was a full-stack argument for AI factories, Windows-native agents, Taiwan manufacturing, and physical AI.
•12 min read
LFM2.5-8B-A1B: Liquid AI's Edge Model Bet Is About Active Parameters
Liquid AI's LFM2.5-8B-A1B shows why the small model race is moving from parameter count to active compute, memory bandwidth, local runtimes, and the devices that can run agents at the edge.
•11 min read
World Models: The Simulation Layer AI Labs Are Racing To Own
World models began as Schmidhuber's curiosity-driven controllers. Genie 3, DreamerV3, GAIA-1, GameNGen, V-JEPA, and open-ended play show why simulation is becoming AI's next platform layer.
•9 min read
TwELL: The Sparse Kernel Bet That Makes LLMs 20.5% Faster
Sakana AI and NVIDIA's TwELL shows why sparse LLMs were not blocked by theory. They were blocked by GPU execution economics.
•20 min read
Recursive Language Models: How RLMs Work
RLMs treat prompts as environments, not inputs. The MIT paper behind Recursive Language Models, the REPL execution loop, and why the AI industry is adopting it.
•17 min read
From Altruism to Empire: The Full OpenAI Origin Story
From a $1B nonprofit pledge to a $300B for-profit empire: the decade-long origin story of OpenAI, from founding through ChatGPT to $12B in annualized revenue.
•13 min read
Qwen3.5: The Model Anthropic Didn't Name
Qwen3.5-397B-A17B scores 88.4 on GPQA Diamond at $0.60/M, yet was absent from Anthropic's distillation report. That absence explains more than the names that were included.
•17 min read
Anthropic Catches Three Chinese AI Labs Stealing Claude
Anthropic named DeepSeek, Moonshot, and MiniMax for industrial-scale Claude distillation. Evidence is real, but calling it an 'attack' ignores how AI was built.
•14 min read
From Laughingstock to $380B: The Anthropic Comeback Story
Anthropic was mocked as the AI company that couldn't ship. Claude 2 was the punchline. Four years later, they own enterprise AI at a $380B valuation.
•10 min read
Gemini 3.1 Pro Doubled Its Reasoning Score in 3 Months
Google doubled ARC-AGI-2 from 31% to 77% in one update. Gemini 3.1 Pro leads 13 of 16 benchmarks at $2/M tokens, undercutting Claude Opus 4.6 by 60%.
•8 min read
Claude Sonnet 4.6: Opus-Level Intelligence at Sonnet Price
Claude Sonnet 4.6 is the new claude.ai default - preferred over Opus 4.5 59% of the time, with 1M context and $3/$15 per million tokens.
•8 min read
OpenClaw: Weekend Project to OpenAI Acquihire in 90 Days
OpenClaw hit 190K+ GitHub stars before OpenAI acquired it - despite 512 security vulnerabilities including an 8.8 CVSS remote code execution flaw.
•13 min read
ByteDance Seed2.0: The Full-Stack AI Empire Behind Seedance
ByteDance's Seed2.0 reveals a complete AI ecosystem - frontier LLMs, multimodal vision, agentic coding, and cinema-grade video at a fraction of Western pricing.
•7 min read
MiniMax M2.5: Frontier AI at 20x Less Than Claude Opus
MiniMax M2.5 - 230B params, 10B active - scores within 0.6 points of Claude Opus 4.6 on SWE-bench at 20x lower cost, backed by a Hong Kong IPO.
•10 min read
Claude Opus 4.6: Why It Owns 40% of Enterprise AI Spend
Claude Opus 4.6 commands 40% of enterprise AI spend, found 500+ zero-day vulnerabilities, and Claude Code hit $1.1B ARR as Anthropic raised $30B.
•10 min read
Grok-4: The Breakthrough AI Model That Changes Everything
xAI's Grok-4 sets new benchmarks in reasoning, coding, and scientific knowledge. Can it maintain its edge as data access costs and energy prices keep rising?
•6 min read
One Big Beautiful AI Problem: Energy, Water & Infrastructure
As the 'One Big Beautiful Bill' reshapes US energy policy, UAE and Singapore pioneer integrated AI infrastructure approaches America could adapt.
•12 min read
AI's Hidden Cost: Water, Watts & the Climate Bill
AI's growth is reshaping global utilities - 66B liters of annual water use, fusion-powered data centers, and companies still paying the carbon bill.
•14 min read
82 Years of Neural Networks: From 1943 to the Transformer
From McCulloch-Pitts' 1943 logical calculus to GPT-4: trace 82 years of neural architecture evolution and how foundational insights led to the transformer.
•7 min read
Hunyuan-A13B: Tencent's Sparse MoE Inference Revolution
Tencent's Hunyuan-A13B uses sparse expert activation (80B total, 13B active) to deliver competitive performance at dramatically lower compute cost.
•5 min read
OpenAI's TPU Shift: What It Means for Nvidia's Dominance
OpenAI's move to Google's TPUs for inference signals a fundamental change in AI compute economics - and why it matters more than Nvidia's stock price suggests.
•9 min read
Darwin Godel Machine: The AI That Rewrites Its Own Code
Sakana AI built an AI that improves itself by editing its own code. Here's how it works and why it matters for the future of AI development.
•6 min read
Anthropic's Secret Book-Scanning Pipeline for Claude
How Anthropic built a pipeline scanning millions of books for Claude AI training, the legal battles that followed, and what it reveals about AI data infrastructure.