OpenAI Ships Images 2.6 and a Kernel Bug Pipeline as Anthropic Maps Bacterial Vaults

OpenAI Ships ChatGPT Images 2.6 With GIFs and Faster API Models

OpenAI on Wednesday introduced Images 2.6 across ChatGPT, Codex and the API. OpenAI’s Images 2.6 announcement says the release generates images about 50% faster than before, adds GIF creation and editing, and improves consistency when refining characters, products and brand details across editing rounds. The company also says people now create about 4 billion images per week across ChatGPT, Codex and the API, or more than half a billion images a day on average.

In the API, developers get GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. Flare, with quality set to low, can generate images in as little as six seconds and is pitched for posters and product imagery; Sunburst targets higher quality and stronger consistency across edits at medium or high quality. Image markup, already available in ChatGPT, comes to the API, along with SVG exports and a new complexity parameter that controls how much planning the model does. Images 2.6 is available globally in ChatGPT and Codex on all subscription tiers. Figma Weave’s Creative Director Itay Schiff put the pitch this way:

“GPT‑Image‑2.6 brings speed and creative control together, helping you generate and edit images while preserving the details that define a character, product, or brand.”

OpenAI’s Pipeline Finds and Patches 948 Linux Privilege-Escalation Bugs

The same day, OpenAI published a longer security research write-up on an autonomous vulnerability discovery, exploitation and patching pipeline. OpenAI’s post on autonomous Linux vulnerability research says that in one mid-October 2026 run powered by an upcoming cyber model, the pipeline dynamically reproduced 1,724 vulnerabilities and produced 948 working local privilege-escalation exploits for a commonly used Linux distribution, along with candidate patches for maintainer review. The total run cost $1,213,430 in API credits, or about $1,279 per distinct LPE. OpenAI does not name the upcoming cyber model in the post, and we are not inventing a name here.

The pipeline used Scout agents to pick 5,034 audit targets covering about 2.8 million lines of code, roughly 7% of the kernel; Analyst agents then produced 2,308 candidate findings in 12 hours. Triage agents reproduced findings in isolated virtual machines with Kernel Address Sanitizer, and ExploitDev agents tested candidates against stock desktop protections including SMEP, SMAP, KASLR and HARDENED_USERCOPY. OpenAI says agents also produced 291 information-leak proofs of concept, and that human researchers review exploits and patches before reports go to maintainers under the Patch the Planet initiative. The lead summary frames the result plainly:

We’ve developed an autonomous vulnerability discovery, exploitation, and patching pipeline. In one run with our cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in Linux at an API cost of $1.2 million.

Anthropic and Claude Map the First Prokaryotic Vault Particle

Anthropic’s biology research lab published a separate science result on Wednesday. Anthropic’s research post on vault particles with Claude describes the first 3D structure of a vault particle found in prokaryotic cells, built from a bacteriophage protein the team calls L05. Vaults are large hollow protein barrels discovered in 1986; their function is still unknown. The phage-derived vault uses 96 copies of L05, about 9.5 megadaltons in total, and measures roughly 54 by 45 nanometers. Despite being shorter than mammalian vaults, Anthropic says it encloses about 25% more volume because it is more spherical.

Under human direction, Claude searched a database of about 15 billion predicted proteins, designed DNA constructs, helped optimize isolation, and processed cryo-EM data. Anthropic says a purification that conventionally takes about three days took about 90 minutes, and a first structural model was ready in 34 hours. The team also engineered a larger vault by adding wall repeats. Vault discoverer Leonard Rome, brought on as an adviser, said:

Figuring out the function of vaults is still enormously challenging. I am pleased that Anthropic has so much good data. Now, other people will be working on this problem … The broadness of Anthropic’s finding[s] suggests that it is going to be a complicated function and an interesting ride.

Anthropic is clear about limits: the vaults were made in E. coli, not yet shown to form in their natural bacterial or phage hosts; a preprint was submitted to bioRxiv but was not yet live when the post went up; and Claude also made costly mistakes, including a protocol error that set the project back two weeks and a flawed mirrored cryo-EM map that humans had to catch.

Cat Wu and Austin Lau Leave Anthropic for OpenAI

Two high-profile Anthropic product and growth people announced moves to OpenAI on Wednesday. Cat Wu, who led product for Claude Code, said in Cat Wu’s X announcement that she is joining OpenAI that she will lead product management for Core Products. Business Insider’s report on Cat Wu joining OpenAI says OpenAI confirmed the hire, Anthropic declined to comment, and that Wu will work with Nick Turley on enterprise products and Ashley Alexander on consumer products while reporting to Tibo Sottiaux. Wu left Anthropic last week after becoming one of the public faces of Claude Code’s rise.

Austin Lau, who helped build Anthropic’s growth function, separately posted that he recently joined OpenAI. In Austin Lau’s X post announcing he joined OpenAI, he thanked Anthropic colleagues and did not state a new title. We are not repeating unconfirmed IPO timing or valuation figures circulating around these exits.

Moonshot Publishes Kimi K3.1 Pro Open Weights on Hugging Face

Moonshot AI published the full open weights of Kimi K3.1 Pro on Hugging Face. The Kimi K3.1 Pro model card on Hugging Face describes a natively multimodal Mixture-of-Experts model with 2.8 trillion total parameters, 104 billion activated parameters, a 1,048,576-token context window, text and image input, and MXFP4 weights with MXFP8 activations from quantization-aware training. The card says K3.1 Pro is post-trained from Kimi K3 with large-scale joint reinforcement learning, and that full-attention gated MLA layers were replaced with Gated DeepSeek Sparse Attention layers for more efficient long-context modeling.

Moonshot recommends vLLM, SGLang or TokenSpeed for deployment, and offers an API on platform.kimi.ai. The release uses the Kimi K3.1 Pro License, which permits use, modification and redistribution with conditions: a separate Moonshot agreement is required before commercial Model-as-a-Service use once aggregate revenue exceeds $20 million over any 12 months, and products above 100 million monthly users or $20 million in monthly revenue must display the model name. Those thresholds do not apply to internal use.

Ricursive Releases Ripple-65B Preview as a Dense Open-Weight Coder

Ricursive Intelligence released Ripple-65B Preview, a general-purpose open-weight variant of the model that powers its AI chip design platform. Ricursive’s Ripple-65B Preview announcement says the dense Transformer was pretrained on 12 trillion tokens with support for up to 1 million tokens of context, and that it scores 80.52 on TerminalBench 2.1, which Ricursive calls U.S. open-weight state of the art, ahead of several larger U.S. open-weight models it lists. The company says full R&D cost from scratch was $32 million, including laddering, data work, evals and the final runs, and that the model is available on Hugging Face, Together AI, GMI Cloud and OpenRouter.

Ricursive chose a thin-and-deep dense architecture rather than MoE, reporting about 40% model FLOPs utilization with simple sharding and no custom kernels. It says a steep reinforcement-learning climb reached the TerminalBench score with about 1.4 million rollouts. Those benchmark comparisons and cost claims are the company’s; independent replication is TBD.

Together AI Launches Together Compiler for Continual Inference Optimization

Together AI on Wednesday launched Together Compiler, also called TKCC, a continual compilation loop for AI inference. Together AI’s blog post on Together Compiler says agents and persistent inference engines re-optimize GPU code as models, traffic shapes and hardware change, typically in less than 24 hours. Together reports up to 2.6x higher tokens per second across frontier open-source models it names, state-of-the-art low-latency SemiAnalysis InferenceX results of 1,111 TPS on MiniMax M3.1-Flash and 1,205 TPS on DeepSeek v4.1-Flash, and 1.7x over available open-source megakernels on a Llama 1B batch-size-1 setting. The company says the system is already serving trillions of tokens and is available in limited beta for dedicated deployments.

The core idea is persistent engines that hot-reload candidate forward passes instead of restarting vLLM, SGLang or TensorRT-LLM for every experiment, which Together says cuts time-to-serving from minutes to about nine seconds in its table, plus a two-tier measurement stack that balances throughput and fidelity on real traffic.

What We Left Out

OpenAI’s alignment site also published a Nov. 4 report saying GPT-6 Astra completed about 10% of hidden rubric items on safety-framed coding tasks versus about 16% on matched capabilities-framed tasks; we are holding a full write-up for a later brief so today’s piece stays focused on product, security and open-weight releases. Reuters’ report on Meta’s $15 billion bond sale says Meta raised $15 billion in a five-part U.S. dollar offering to help fund AI infrastructure, after earlier reports that the company was targeting roughly $20 billion to $25 billion across dollars and euros; final size of any euro tranche remains TBD in our notes. Rumors of GPT-6.2 Astra/Sol dates and a gemini-omni Cloud console sighting remain unofficial, so we are not treating them as news.