This week the AI industry stopped asking "which model is smartest" and started asking "who runs it best." Anthropic shipped an agent that comes to you before you ask, Google showed that rewriting an agent's tooling can beat upgrading its model, Perplexity open-sourced the sandbox it uses to keep its own agents in check, Inception Labs launched a diffusion model that claims over 1,000 tokens per second, and Google and Amazon both moved to run PyTorch natively on their own chips. Here is what happened, with every claim verified against primary sources.
Key Takeaways
Anthropic opened a research preview of Claude Duet (Oct 21), a proactive agent app built on Claude Opus 5.5 that sends a morning brief and alerts when something needs attention, waitlist-only for US Pro and Max subscribers. (Reuters)
Google Cloud AI Research open-sourced Harness² (Apache 2.0), a framework that improves agents by rewriting their harness, prompts, memory, and tools while the model stays frozen, reporting up to an 11.5-point lift for GLM-5 on a workplace benchmark. (Ground Truth)
Perplexity open-sourced SPACE (Oct 15), its agent sandbox in Rust on Firecracker microVMs: ~60ms median spin-up and zero VM-to-host escapes in roughly 118 red-team attempts, now used by CoreWeave, Akamai, and Crusoe. (Forkast)
Inception Labs launched Mercury 3 (Oct 19), a diffusion-based model claiming 1,068 output tokens per second on Nvidia GPUs at a 50% launch discount: $0.10 per million input and $0.35 per million output tokens. (TechRepublic)
MiniMax shipped M3.1 Flash (Oct 19) at $0.15/$0.60 per million tokens with weights scheduled for Oct 23, while Mistral launched Timestral (Oct 15), its first time-series foundation models in four sizes from 5M to 209M parameters. (kie.ai)
Google open-sourced TorchTPU and Amazon shipped TorchNeuron, native PyTorch backends for their own chips, both aimed at loosening Nvidia's hold on PyTorch workflows. (MarkTechPost)
Sam Altman confirmed OpenAI "will definitely" build a humanoid robot, targeting a demo in 2027, with 27 robotics openings on OpenAI's careers page. (TBreak)
1. Anthropic Previews Claude Duet: The Agent That Comes to You
Anthropic opened a research preview of Claude Duet on October 21, an AI agent app that flips the usual script: instead of waiting for prompts, it watches your connected apps and comes to you with what needs your attention, reports Reuters. It is built around Claude Opus 5.5 and is available as a standalone iOS and Android app as well as on claude.ai.
The product shape is deliberate. Duet sends a morning brief, checks connected apps every few hours, and sends an alert when something needs action, then does the work in its own cloud environment, a virtual machine and browser, so it keeps running even when the user's device is off. Before taking consequential actions, sending messages, making payments, signing documents, or deleting files, it asks for approval.
Availability is narrow for now: waitlist-only, initially for US subscribers on Pro and Max plans, with Team and Enterprise plans not included. It works with more than 6,000 connectors such as Gmail and Slack, handles purchases through Stripe Link and Shop Pay, and runs on the subscriber's existing plan usage limits. The pitch is less "a smarter model" than "a coworker who notices things."
If you like quick video takes on how new open AI projects work, browse the repo signals for 30-second overviews.
2. Agents Get Open Tooling: Google's Harness² and Perplexity's SPACE
Two open-source releases this week attacked the same idea from opposite ends: the agent's software, not the model, is where the leverage is.
Google Cloud AI Research published Harness², an Apache 2.0 framework that improves an agent by rewriting everything around the model: the harness, system prompt, memory, tool guidance, subagents, skills, scripts, and plugins, while the model weights stay frozen, reports Ground Truth. The paper (arXiv 2610.14224) was submitted October 9, the initial release commit landed October 8, and the repository had collected roughly 5,900 GitHub stars by October 19. In one reported run, the approach lifted GLM-5 by up to 11.5 points on the WorkBuddy-Bench workplace benchmark, a striking claim for a technique that touches no weights.
Perplexity went the other direction and open-sourced the safety layer. On October 15 it published SPACE, its agent sandbox, under the Apache 2.0 license: Rust code running on Firecracker microVMs, reports Forkast. The numbers Perplexity puts forward: a median sandbox spin-up of about 60 milliseconds, pause and resume with rolling snapshots of up to a week, and zero VM-to-host escapes in roughly 118 adversarial red-team attempts, with credentials and network access controlled outside the VM. CoreWeave, Akamai, and Crusoe are listed as adopters, and the repository drew about 820 stars within a day of release. Together with Google's tooling release, this is the infrastructure week for agentic AI: one open tool to make agents better, one open tool to keep them safe.
3. Inception's Mercury 3: Diffusion at 1,068 Tokens per Second
Inception Labs launched Mercury 3 on October 19, a diffusion-based language model built for raw speed. The company claims 1,068 output tokens per second on Nvidia GPUs, roughly 4 to 8 times faster than OpenAI's GPT-6 Luna (about 130 tokens per second) and Anthropic's Claude Haiku 5.5 (about 240 tokens per second), reports TechRepublic.
The model carries a 1-million-token context window, up from 260,000 in Mercury 2.5, and Inception says it has 40 times the total parameters of Mercury 1, with quality competitive against GPT-6 Luna and Claude Haiku 5.5, though those quality claims are vendor-reported and mixed. The pricing is the other half of the story: $0.20 per million input tokens and $0.70 per million output, with a 50% launch discount ($0.10 and $0.35) carrying no stated end date. It is available through Inception's API as well as OpenRouter, Baseten, and Vercel. If the speed numbers hold up in independent testing, Mercury 3 makes the case that diffusion models are not a research curiosity but a pricing strategy: tokens so cheap and fast that agents stop being a luxury.
4. The Price War Continues: MiniMax M3.1 Flash and Mistral Timestral
MiniMax released M3.1 Flash on October 19 as an API, with model weights scheduled for October 23, reports kie.ai. The headline is the price: $0.15 per million input tokens and $0.60 per million output ($0.015 per million cached), which the company frames as 70% cheaper than the previous generation for agentic tasks, with about 3x faster inference and claimed speeds above 200 tokens per second (independent measurements land closer to 105-123). It carries a 1-million-token context window with up to 524,288 output tokens and text, image, and video input. It scored 43 on the Artificial Analysis Intelligence Index against 29 for M3, and the model had previously been running anonymously as "Space Bunny" with over 200 trillion tokens processed before its reveal.
Mistral took a different slice of the market the same week. On October 15 it launched Timestral, its first suite of time-series foundation models: four sizes from 5 million to 209 million parameters, doing zero-shot probabilistic forecasting with 21 quantiles per step and a 32,768-time-step context, notes the October 19 AI brief by Sean Weldon. The smallest model returns a one-day forecast in 10 milliseconds on an 8-thread CPU, the largest in 140 milliseconds, so these run on ordinary hardware, and the weights are open under a Modified MIT license with Apache 2.0 inference code. Mistral says the 209M model ranks first among pretrained zero-shot models on GIFT-Eval (MASE 0.658, CRPS 0.446), TIME, and fev-bench, its own benchmark claims. The message of the week for buyers: model prices keep falling, so cost estimates based on last year's prices are already stale.
5. Google and Amazon Both Bet Against Nvidia's PyTorch Monopoly
At the PyTorch Conference 2026 in San Jose on October 20, Google open-sourced TorchTPU, a native PyTorch backend for its Cloud TPUs, co-developed with Meta's PyTorch team, reports MarkTechPost. Existing code runs on TPUs through `torch.device("tpu")`, with roughly 450 ATen operations, eager mode, and torch.compile support; version 0.2.0 is on PyPI for Python 3.11 and up. The release ships as four repositories: TorchTPU itself, vLLM-TorchTPU, SGLang-TorchTPU, and TorchTitan. Google's own figures cite 52% model FLOPs utilization training Qwen 3 8B on 8,192 Ironwood (TPU7x) chips, and the company says TorchTPU replaces PyTorch/XLA for new workloads.
Amazon shipped its answer in the same window: TorchNeuron, a native PyTorch backend for its Trainium chips, reports Pondero, framing the pair as a direct move against Nvidia's grip on PyTorch workflows. The strategic point is not performance but portability: if training and serving code runs unchanged on TPUs and Trainium, the switching cost that protects Nvidia's CUDA moat gets smaller with every release. Expect this to matter most to cloud buyers renegotiating GPU contracts next year.
6. OpenAI Confirms the Humanoid: "We Will Definitely Do a Humanoid"
Sam Altman said it twice, once on the Sources podcast ("We will definitely do a humanoid", adding "We will do other form factors as well", "Both") and once to Fortune ("I think we'll have a cool demo in 2027", with robots on streets "a few years longer"), reports TBreak. The interviews themselves date to September, but the confirmation went wide in coverage on October 18-19. The hiring backs it up: OpenAI's careers page lists 27 robotics openings spanning motion control, state estimation, custom actuators, manufacturing, supply chain, and ML infrastructure for physical AI, reports TechTimes.
Altman's framing is that the "brain" is the hard part, not the body, which is a bet that OpenAI's models transfer directly into physical robots. What was not announced: the robot itself, any specifications, a manufacturing partner, or a price. Note the irony, noted by the same coverage: OpenAI once backed humanoid startups and now plans to compete with them. A 2027 demo date means two more years of promises, but the hiring roster says this is not a hobby.
How the Week's Stories Stack Up
Aspect | Story | Key Number | Source |
|---|---|---|---|
Proactive agent | Claude Duet (Anthropic) | 6,000+ connectors, waitlist | Reuters |
Agent tooling | Harness² (Google) | +11.5 pts on WorkBuddy-Bench (reported) | Ground Truth |
Agent safety | SPACE (Perplexity) | 0 escapes in ~118 red-team attempts (own figures) | Forkast |
Speed | Mercury 3 (Inception) | 1,068 tokens/sec claimed | TechRepublic |
Price | M3.1 Flash (MiniMax) | $0.15/$0.60 per 1M tokens | kie.ai |
Portability | TorchTPU (Google) + TorchNeuron (Amazon) | ~450 ATen ops, native backends | MarkTechPost |
Robotics | Humanoid (OpenAI) | Demo targeted 2027, 27 openings | TBreak |
The thread across all six: the model itself is no longer the story. This week was about everything around it, the agents that use it, the tooling that improves it, the sandboxes that restrain it, the price of serving it, the chips it runs on, and the bodies it might eventually inhabit.
Frequently Asked Questions
What is Claude Duet?
Claude Duet is Anthropic's research-preview agent app, launched October 21, 2026, built on Claude Opus 5.5. Instead of waiting for prompts, it sends a morning brief and alerts when something in your connected apps needs attention, working in its own cloud VM and browser. It is waitlist-only, initially for US Pro and Max subscribers.
What is Google's Harness²?
Harness² is Google Cloud AI Research's open-source (Apache 2.0) agent framework that improves an agent by rewriting its harness, system prompt, memory, tool guidance, subagents, skills, scripts, and plugins while the model weights stay frozen. One reported run lifted GLM-5 by up to 11.5 points on WorkBuddy-Bench.
How fast is Inception's Mercury 3 really?
Inception claims 1,068 output tokens per second on Nvidia GPUs, which it says is 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5. Those are vendor-reported figures; independent testing will show whether they hold.
Is MiniMax's M3.1 Flash really that cheap?
The published API pricing is $0.15 per million input tokens and $0.60 per million output ($0.015 cached), which MiniMax frames as 70% cheaper than the previous generation for agentic tasks. Model weights were scheduled for release on October 23.
What are TorchTPU and TorchNeuron?
TorchTPU (Google) and TorchNeuron (Amazon) are native PyTorch backends for Cloud TPUs and Trainium chips respectively, both released around October 20, 2026. They let existing PyTorch code run on non-Nvidia hardware with minimal changes, reducing dependence on Nvidia's CUDA ecosystem.
Is OpenAI actually building a robot?
Sam Altman confirmed OpenAI "will definitely" build a humanoid robot, targeting a demo in 2027, and the company's careers page lists 27 robotics openings. No robot, specifications, manufacturing partner, or price has been announced.
Conclusion
The center of gravity in AI moved this week, from the model weights to the machinery around them. Anthropic's Duet proposes a future where the agent initiates contact, Google's Harness² and Perplexity's SPACE suggest the real competition is in open tooling and safety infrastructure, Inception and MiniMax keep compressing the price of intelligence, and Google and Amazon are chipping away at Nvidia's hardware moat. OpenAI's robot, meanwhile, is a promise with a hiring page behind it. Next week will show which of these bets survives contact with users.
Read this article in Bahasa Indonesia: Berita AI: 23 Oktober 2026.