GPT-6.1 Terra continues to advance the cost-intelligence frontier across our model families. It’s a significant step up over GPT‑5.6 Terra, with 50% lower standard API prices. In ChatGPT Work and Codex, your usage allowance goes further with GPT‑6.1 Terra than with GPT‑5.6 Terra—and far more than with GPT‑6.1 Sol or GPT‑6 Astra. For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra.

New from our STEM Fellows Program: Designing more efficient quantum circuits with Claude. Given plain-language instructions, Claude Mythos 5 designed quantum circuits that improve on the best published ones, sometimes by an order of magnitude or more. More efficient circuits reduce the resources needed by quantum algorithms, which could bring quantum computing applications in chemistry and materials science within reach sooner. Read more: https://t.co/QCvoJLDmBo

L6N5401-0.75 seems to be the 'DJ neuron' - it puts everything into the context of a DJ set https://t.co/hTZjOnBhBe

Announcing Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Carbon represents a step up across most benchmarks, delivering strong frontier performance in autonomous coding and cyber defense. We’re first making Carbon available to a subset of trusted cyber defenders in our Fairwind program. We’ll expand access as soon as possible.

No pull requests, no code review. We push to main. Our agents read production logs, watch deploys, grant feature flags, and check their own work end to end. Here's how we do that: https://t.co/G4r37Lgoxc

Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems. This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution. The emerging rules of thumb: → Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions. → Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data. → Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with. The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed. Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up. As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete. https://t.co/DNq60Skkmc

I am transitioning away from using CLI agents for development. On my last project I transitioned to using grokbots.

A Turkish physicist's confession: he spent a year on ~500 pages of hand calculations for quantum gravity in his PhD thesis, but Opus 5.5 did it in 10 hours, including the article and GitHub repo. 😱 It even found the errors. Luckily, the result didn't change!

Energy-Based Models can now scale better than Diffusion models. Introducing TEM: Transformer Energy-based Models TLDR: • TEM outscales Diffusion without guidance • First scalable contrastive Energy-Based Models • Better generalization, test-time scaling, and OOD detection than Diffusion 🧵

BREAKING: We broke the wall for Integer Multiplication! Conditional κ = 0.00302 (+80%) For days we were stuck in incremental improvements, but today zxv changed the game. zxv’s insight: stop throwing away work. Helper registers that finish holding a completed transform of their own input now count as output, instead of being undone. He brought us to 0.0023(+37.5%), and immediately after 1337isnot1337 pushed that to 0.00302. The game is afoot :)

after much experimentation, we landed on a single set of rails that powers our agentic systems across the company - background agents, async agents, live chat: 1. stack - accounting harness (https://t.co/FP0HUI8zNl) 2. glass - general purpose coworker (https://t.co/Q7X7SGdBE0) 3. inspect - coding agent 4. our support chat bot, Ask Ramp 5. headless agents that do document OCR 6. ramp text (HIRAMP/447-267) agent works with all models, supports all modern harness features - background tasks, sandboxes, tool hooks, tool search, durability, forking, compaction, queueing, streaming, and observability laminar, our new harness, sits on top of ramp router, and brings enormous leverage to our agent developers across the company more info on our blog ->

How do you _guarantee_ your agent's code is correct? You prove it. Introducing Proof Driven Development Kit (PDD Kit) from @MSFTResearch. PDD Kit helps turn your plan into a formal spec, then proves the agent's code meets that spec for every input. No vibe verification. PDD Kit is integrated with @GitHubNext's Chopin, so you can plan as a team. PDD Kit takes an app's login flow from intent to a verified pull request: It discusses your code, infers your intent as it goes, and proposes a formal spec in Verus, a proof language for Rust. Once you approve the spec, the Verus verifier automatically proves the code meets it. Where proof isn't possible, it falls back to empirical verification with tests and computer-use agents (CUAs). It works on Rust today. No new language, no change in how you produce the underlying code, making formal methods more accessible. The goal is a verified software factory, where engineers review specifications instead of code. Come see it in action at GitHub Universe, Oct 28–29. Swing by the Speakeasy and ask for GitHub Next!

Introducing Simlock: a free, open-source tool that turns your machine into a simulator cloud that manages itself, so every coding agent gets its own iOS or Android device, simulator or real phone. https://t.co/iymaqe5TVp https://t.co/fxJtaXUk68

Introducing Kino: make videos end-to-end with Claude Code and Codex. Drop your footage, get highlights, edit together, send links for feedback, and let your agents cook. AI is becoming a great creative partner. Kino is a new home for your team and agents to collaborate. https://t.co/AJStZkrAS1

Introducing Qwen Image 2.1 Realtime ⚡️ I'm releasing V1 of my *2-step* distilled LoRA, and it runs in *realtime* on ZeroGPU ⚡️⚡️ (live editing part of the video was not sped up!) > DMD2, rank 32 > 2 steps, with quality competitive with Turbo (8-step) and Base (40-step) > realtime on ZeroGPU (half of an RTX Pro 6000) LoRA: https://t.co/dZQC6MOxbx Space: https://t.co/6fHURqMoZ3

+1 more

Pocket TTS has been getting updates those last few days. While it was ~100M parameters, we release smaller versions, down to 1.9M: an 18.2M model with the quality of the original, a 6.9M model, and a 1.9M model that runs 103x faster than real time on one CPU core. All with voice cloning.

I agree with one thing. Making linear attention work at scale takes talent, and not every lab has it. The rest doesn’t hold up. In order: (1) “Why are the Western labs choosing to drop context instead of compressing it?” The first rigorous, controlled, large-budget test of compression was what we did at NVIDIA’s in June 2024(https://t.co/OQ3pkwmXXV). We trained 8B Mamba, Mamba-2 and Transformer models on the same data for up to 3.5T tokens. Additionally we trained an 8B hybrid of 43% Mamba-2, 7% attention and 50% MLP and we showed it beats the 8B Transformer on all 12 standard tasks (+2.65 points on average). Interestingly this hybrid variant was up to 8× faster at generation, and matched or beat the Transformer across 23 additional long-context tasks. This is seven months before MiniMax-01, fifteen before Qwen3-Next, sixteen before Kimi Linear. And since then NVIDIA has been shipping a hybrid architecture, with modifications, repeatedly. Good examples are Nemotron-H, Nemotron Nano 2, and Nemotron 3 (Nano, Super, Ultra) which is a Mixture-of-Experts hybrid Mamba-Transformer family with context up to 1M tokens, post-trained with multi-environment RL. Nano keeps exactly 6 attention layers out of 52 and the other 46 are Mamba-2 and MoE blocks. Super is 120B total / 12B active, pre-trained on 25 trillion tokens. IBM’s Granite 4.0 (October 2025) interleaves Mamba-2 and attention layers at 9:1. To this list, you can also add Jamba (AI21, March 2024, a 52B production hybrid), RecurrentGemma (Google, April 2024), Codestral Mamba (Mistral, July 2024) and Phi-4-mini-flash (Microsoft, July 2025). (2) “SWA is a lossless local transformation, therefore superior.?” SWA is lossless inside the window and zero outside of it and it is exactly why it fails so badly. (https://t.co/sY3p9p9e6u) for example shows that larger sliding windows hurts the long-context performance. This is because a precise local softmax lets the model avoid training the long-term memory it would otherwise rely on. They show that a hybrid with a 128-token window beat the same hybrid with a 2048 window by 16 NIAH points (88.9% higher recall), and at 131k the short-window models still retrieved about 30% of needles while the 2048-window model was near 0%. And the same degradation from larger training windows shows up in local-global stacks that alternate SWA with full attention ( e.g. the GPT-OSS/Gemma recipe). (3) “Compression is strictly harder to serve.” This was true in 2024. But now vLLM, FlashInfer, SGLang and TensorRT-LLM serve hybrid linear models extensively and without any issues. The most exciting part is the recent FlashInfer additions that let you use chunked gated-delta-rule prefill and gated-delta-rule decode kernels, with an SM100 (Blackwell) path. It even supports much newer models like Gated DeltaNet-2 and also lets you use efficient kernels for training. Also, let's not forget ReplaySSM (https://t.co/zSL7l5o8kO) which enables a very optimized speculative decoding path for linear attention. (4) “If you write all the kernels anyway, why not roll your own attention in the meantime?” On NVIDIA hardware, linear attention is just as free as attention. The Mamba CUDA kernels exist. flash-linear-attention, the Triton library that the GDN and KDA layers in Qwen3-Next, Qwen3.5 and Kimi Linear use, was written for NVIDIA GPUs, and vLLM and SGLang use it directly instead of reimplementing it. Megatron has supported Mamba-2 training since 2024. A Western lab on H100s that wants a hybrid doesn’t need to write a kernel. It can just import it. So kernel cost can’t be what keeps them on using the SWA. Also consider that decode is memory-bandwidth-bound, so the cost of every generated token is dominated by reading the KV cache, which grows with context length. A constant-size state removes that growth. That’s why Nemotron is mostly Mamba-2, and why Nemotron 3 Super gets 2.2× the throughput of GPT-OSS-120B at 8K in / 16K out. (5) “For your first release, it’s a much safer option to ship SWA. Everyone knows it’s working just fine.”? Safe compared to what? “Safe” should mean proven in production at scale. By that standard the hybrids are the safe choice and SWA is the bet. Look at what has shipped on Gated DeltaNet alone. Qwen3-Next, then Qwen3.5 from 27B up to 397B-A17B, then Qwen3.6. Kimi Linear, then Kimi K3 which is a 2.8 trillion parameters, open weights, 1M-token context, a 3:1 hybrid of Kimi Delta Attention and gated MLA. GLM-5.3-Flash on the same KDA layout. KDA is Gated DeltaNet with channel-wise instead of scalar gating. Right now, three frontier labs run their model in production with 1M context, RL post-training, etc. all on the same layer. Meanwhile SWA is the architecture with the documented issue. It fails past the window and drags any hybrid it sits in down to pure-SWA recall. And when MiniMax reached for an SWA hybrid as the fallback in M2, it failed and they went to full attention instead. In my view, shipping SWA for a first release in late 2026 isn’t conservative. It’s picking the layer with a known failure mode because it needs no new code.

good time to announce that i’ve joined @GoogleDeepMind as a forward deployed engineer i’ll be partnering with startups and researchers across the ecosystem to co-design benchmarks. if you’re building interesting evals, i want to hear about them!

Today we're releasing Blackthorn in Private Alpha: The first Private AI inferencing stack that removes the CPU TEE from the trust chain. Blackthorn can load and run any AI model privately with little performance overhead compared to plaintext inference. https://t.co/5avAouhgWA

today, we introduce a little mischief to the world. for four years, we’ve been brewing real brain cells on electrodes to understand how they compute. now, we’ve turned those discoveries into a video model anyone can use. from 0.5¢ a second, mischief 1 is at least 5x faster than its base and the cheapest image-to-video model on @fal and @OpenRouter. soon to be on @awscloud marketplace. it’s not a delusion anymore. biocompute is real, commercial and this is just the beginning. happy halloween.

We cancelled our Figma enterprise plan This is how we design at HeroUI now 🧵 https://t.co/5v5ARROXvi

So many SO-ARMs! 💪 Here is the robotics workshop organized at @NVIDIAGTC Berlin that uses SO-ARM101 and LeRobot to control real and simulated robotic arms. The light box makes it easier to get policies working in any location and light conditions. https://t.co/s2QE6UWQVg

I built bbtree, the foundation of databases and filesystems, and the fastest concurrent btree. Code: https://t.co/CWpHJ7c5Bj Blog post & demo: https://t.co/xRrdiPcgbz https://t.co/w2snig2zny

https://t.co/ZcNtg74dVS

Today we are releasing JST-2, the world's first full-stack physics-native world foundation model. Many world models learn in pixel/video space. We went back to first principles and built the entire model stack in 3D. Generate a 3D scene in seconds and expand it in real time. Create characters with motion built in, all through one general architecture.

Papers with Code is now open-source! Apache 2.0-license. Let's collaboratively make the best paper interface! https://t.co/hwh6CEApyK

People assume that in an AI doing formal verification world, they can take their existing codebase, ask AI to prove it correct, and get either a proof or a bug. Sadly reality has a few bones to pick with this, and one of them is the "existing codebase" part. And that's due to the good ol' halting problem. The short of it is that there is no general purpose algorithm that takes any program and input and correctly determines whether it will terminate or run forever. With a few mathematical flourishes we can extend that to "Rice's theorem" and prove this is true for any nontrivial semantic property, meaning we can't automatically prove arbitrary things about arbitrary programs. Now obviously we can prove some things about some programs, or else formal verification wouldn't exist as a field. The trick is that we write the code in a style suitable for verification. One example of this is totality: in Lean, all functions must be guaranteed to return a value for all inputs. If we want some code that runs forever (like f.ex a server), instead of `while true` we do `while (i < 10^60)` so that it's guaranteed to terminate and return a value (after the end of the universe). (Totality is also why, in Lean, `1/0 == 0`. If you convert your code to Lean without accounting for this, you could have correct lean but incorrect code!) Another "style" we use is state machines. Even if your code doesn't quite fit an SM architecture, if you want to prove it at scale, you're probably going to go with a state machine. State machines are the carcinisation of provable code. Hopefully by now the problem is clearer: even if your codebase is objectively correct, it might not be in the right shape to be proven correct! Okay, just have the AI rewrite your codebase into another form, right? Two problems. One: you don't know if you've introduced errors in the rewrite because you can't prove them equivalent (since one of them is not in a provable shape). Two: making code more verifiable can make it worse on other metrics, like modifiability, performance, and complexity. And now you've got the old bugbear of software engineers, the dreaded tradeoff. Now, don't get me wrong, formal verification has a huge role in the future of software, but it's not going to be a breezy "prove my Ruby thx" Claude prompt. It's still going to take work and expertise on our parts to prove our code, and to strike the right balance between formal verification and other forms of correctness. This is one reason I'm so excited to be working at @AntithesisHQ , which tackles a lot of the same properties as proof methods, but verifies them with "deterministic simulation testing" and fuzzing instead of with a proofs. So it's not as thorough and doesn't give you confidence in all inputs, but it automatically works on pretty much any code shape. I've literally grabbed open source projects off Github, told Claude "check properties XYZ in Antithesis thx", and gotten bug reports back. No state machine rewrites needed.

Intel and AMD are working together to simplify AI development on x86. By bringing @PyTorch CPU optimizations together through oneDNN, we’re creating a more consistent software path for developers across Intel and AMD processors. Learn more: https://t.co/ZWjNXxW5SH

Hermes now lets you spawn subagents with different models or profiles to fit the job I made a video showing how to install, configure, and use the Advanced Delegation plugin There are two modes: model routes and profile routes In settings, you can select which models fit which roles, so you can use a Luna for cheap, fast tasks and then Astra when you need deep reasoning For profile routing, you can set which profiles to use when. If the subagent is used for research, it can use the researcher profile. If it is a writing task, it picks the writer profile Now you can ensure that your agents are using the right model, tools, and skills for the job you give it Check it out!

1/ JEPA selects which features its world model keeps. But how? Will it keep a feature it can't predict, like a traffic light? Judging each feature alone, earlier work says no. We show evidence that this is wrong and introduce a new characterization that explains why. 👇🧵 https://t.co/URWx0Xoq87

We launched the Gemini API CLI. A CLI for calling the Gemini API. Give it to the agent you already use (Claude Code, Codex, Antigravity, Cursor, OpenClaw, Hermes...). It's NOT a new coding agent you can use it too: 1. Generate images, video, speech, music. 2. Analyze PDFs, images, audio, video, YouTube URLs, 3. Transcribe recordings to md / json / srt, 4. Run Deep Research in the background. Stateless. No memory, no chat mode. Your agent owns the context and the plan, gemini-api does the API call. Build for Agents, it's self-describing. `--help`, `--dry-run`... Start here: https://t.co/D812RdijJO

Today we're releasing Carbon-X, a 14B model that brings DNA and natural language together for the first time 🧬 1. Beats every open language model we tested across DNA benchmarks. Within a point of GPT-6.1 Sol at >1,000× lower inference cost. 2. Mid-trained on 300B tokens of interleaved DNA and text, followed by SFT (without distillation!) and RL. 3. A new hybrid tokenizer encodes DNA using ~3× fewer tokens than BPE. 4. Gives other models a DNA specialist subagent: Kimi K3 gains 11.1 points when it consults Carbon-X. 5. Runs locally on a laptop with llama.cpp. Apache 2.0. Your DNA stays on your hardware. Model: https://t.co/7wBE7WUVp1 Blog: https://t.co/A3OjalTpIJ

+1 more