7 min read · 16 sources

OpenAI’s image model now ships faster than most teams can finish a design review. Images 2.6 is rolling out globally across ChatGPT and Codex on desktop, mobile and web, on every subscription tier, generating images 50% faster than the previous version. OpenAI says its products now produce 4 billion images a week.

The headline feature is motion: users can turn characters, mascots, product shots, sketches and storyboards into short GIF loops, then adjust playback speed, edit individual frames and share them. The quieter feature is consistency - keeping a character, product or brand detail stable across successive editing rounds, which is the part that actually decides whether this is a toy or a production tool.

For developers there are two new API models. GPT-Image-2.6 Flare targets speed and professional layouts like posters and product imagery, and can return an image in as little as six seconds at low quality. Sunburst is the detail model, aimed at medium and high quality settings where edits have to hold.

Ricursive trained Ripple-65B for $32 million, one-tenth the estimated cost of the previous US open-weight state of the art on agentic coding.

Pangram ships a detector with a heatmap

Source: pangram.com ↗

The same week OpenAI makes synthetic images cheaper to produce, Pangram Image 1.0 arrives to find them. It’s the production release of a research preview from a few months back, and Pangram claims state of the art in both AI image and AI video detection.

The interesting part isn’t the classifier score, it’s the output format. Pangram returns a heatmap that localizes exactly where AI content appears inside an image - including the case of a real photograph with a synthetic element composited into it, which is the scenario that breaks whole-image “is this AI?” checks. Pangram says the model is robust to edits, compression and low resolutions, and ships a Chrome extension plus API support for both images and video. A Flash 1.0 variant sits alongside the flagship.

If you run any kind of upload pipeline - marketplace listings, KYC documents, newsroom submissions - a per-pixel verdict with an API is a different integration than a single confidence float.

Ripple-65B hits US open-weight SOTA for $32 million

Source: ricursive.com ↗

Ricursive Intelligence is a full-stack lab building self-improving systems by co-optimizing hardware, software and model, and it trained a language model from scratch to power its chip design platform. Ripple-65B Preview is the general-purpose variant of that proprietary model, pretrained on 12 trillion tokens with support for up to 1M context.

The number that matters: 80.52 on TerminalBench 2.1, which Ricursive calls US open-weight state of the art on agentic coding. It beats models several times its size - Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B and Nvidia Nemotron 3 Ultra 550B. The training run cost $32 million, which Ricursive says is one-tenth the estimated cost of the previous open-weight SOTA on agentic coding.

Weights are on Hugging Face, Together AI, GMI Cloud and OpenRouter. A 65B model that beats 500B-plus rivals on terminal-style agent tasks is the sort of thing that changes what you can self-host on a couple of nodes rather than a dedicated rack.

OpenAI's pipeline found and patched 948 Linux privilege escalations

Source: openai.com ↗

OpenAI built an autonomous pipeline for discovering, exploiting and patching vulnerabilities, and in a single run its cybersecurity model validated and generated patches for 948 Linux local privilege escalation flaws.

LPE bugs are the boring, high-value class: a process running as an unprivileged user finds a way to become root, and every container, CI runner and multi-tenant host is exposed. Finding them is expensive because each one needs a working exploit to prove it’s real, which is exactly the step most automated scanners skip. OpenAI’s claim is that its model both validated the bugs and produced the patches.

What the post does not say is how many patches were merged upstream, or how many survived maintainer review. Treat 948 as the size of the candidate set that passed automated validation, not as 948 fixes in your kernel.

The Stacks Project is now 16 million lines of Lean

Source: axiommath.ai ↗

Axiom released a Lean formalization of the entire Stacks Project, the roughly 7,500-page open-source reference underpinning modern algebraic geometry. The result spans about 16 million lines of code and Axiom calls it the largest integrated, reusable library of formalized mathematics ever built.

Two details make it more than a stunt. Permanent Stacks tags link every Lean declaration back to the canonical text, so a proof assistant user can trace a machine-checked theorem to the paragraph a mathematician already cites. And because it’s one importable library, researchers and AI agents can pull its definitions and theorems directly into new formal proofs instead of re-deriving them.

The Stacks Project started with Columbia’s Johan de Jong in 2005 and grew through hundreds of contributors. Turning that into something a proof checker accepts is a different scale of effort from formalizing a single famous theorem.

Cognition's SWE-2 Flash adds a mid-training stage and jumps 20 points

Source: cognition.com ↗

Cognition introduced SWE-2 Flash, a coding model it describes as fully American, and the recipe change is the story: for the first time the lab inserted a mid-training stage before reinforcement learning. Starting from Thinking Machines’ Inkling-Small, that took Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%.

Cognition says it outperforms all American open-weight models on leading evals, including models 3x larger. Note the licence posture: this is not an open-weights release. Cognition is building a program to share weights securely with industry partners and researchers, citing cyber risk - a middle path that keeps the artifact out of public download while still letting people evaluate it.

Tencent's Hy4 adds vision, weights promised in days

Source: tencent.com ↗

Tencent announced Hy4, its latest flagship, and says it now ranks among the world’s leading open-source models. Weights and system cards are promised “in the coming days,” so treat the ranking as a claim until you can run it.

The new capability is visual understanding - fine-grained image processing, complex chart analysis, mathematical visual reasoning - alongside improvements in instruction following, long-context comprehension and long-horizon agentic work. Tencent says the model can drive both command-line interfaces and GUIs, which is the same territory the terminal-bench numbers above are measuring.

AlphaEvolve goes public, matrix multiplication bound included

Source: github.com ↗

DeepMind published an AlphaEvolve repository: a Gemini-powered coding agent that pairs model-generated algorithm designs with automated evaluators that check the answers. Google has used it on data center efficiency, chip design and AI training.

The repo has a minimal implementation with local evaluation under /src, plus an /omega directory covering the new upper bound on the matrix multiplication exponent. That includes a Colab notebook for rigorous verification using rational arithmetic, and the evolved optimization program in JAX/Haiku that runs gradient descent over 32k steps with 32 random seeds, with a NumPy/SciPy verifier alongside. Reproducible math results in a public repo is a rarity worth the clone.

The week's smaller items

Source: x.com ↗

Cartesia’s H-Nets learn directly from raw bytes and match benchmark performance with substantially less training compute - if that holds, tokenizers stop being a design decision.

OpenAI is acquiring Pierre Computer Company to bolster Codex’s developer tools, including Code Storage and other open-source projects. Meta is acquiring Conductor, the startup that lets developers run multiple AI coding agents in parallel workspaces. And Cat Wu joined OpenAI to lead product management for Core Products, a post that pulled 4.1 million views.

Together Compiler is a continual optimization loop for inference. Together’s justification is concrete: none of the top-10 open-source LLMs on OpenRouter share an architecture, it serves trillions of tokens a day across thousands of traffic shapes, and H100s, B200s, B300s, NVL72, Groq and friends each need different strategies - too many regimes to hand-tune.

Perplexity’s Decisions API can now search the web before answering. Kimi-K3.1-Pro lands as a 3T-class multimodal model with a million-token context, pitched at coding, 3D creation and factual accuracy.

A paper argues you don't need long context

Source: arxiv.org ↗

A paper submitted to arXiv measures what your bill already tells you: later tokens cost more to generate than earlier ones, by as much as 7x on frontier open-weight models. The authors show that with frequent compaction, models lose little to no accuracy across long-horizon and long-input tasks when kept on short contexts - and that performance doesn’t collapse on long-horizon software engineering even when the model takes at most one action per context. They sketch “semantic state space models” from that result and question whether long-context models are needed at all.

Worth reading against the Kimi release above. The industry is shipping million-token windows and publishing evidence that compaction gets you most of the way there for a fraction of the cost.