Gemini 4 Neon and GPT-6.2 push high-end AI toward lower-cost, practical use, giving developers and teams more output for the same budget. But lower prices do not make agents dependable: they scored only 42.5% on emergency scans, often locking onto an answer too early.
Models & Releases
Gemini 4 Neon gives paid Pro and Ultra subscribers, developers, and businesses a fast model for complex coding, research, and office work, priced at $3 per million input tokens and $15 per million output tokens. It can take up to 1 million tokens of mixed text, spreadsheets, PDFs, charts, audio, and video, then produce structured results in one workflow. Google says Neon ranks third overall on the Artificial Analysis Intelligence Index, scoring 61 at its highest thinking level.
GPT-6.2 Astra targets complex professional and scientific work with better outputs and API input and output prices 20% below GPT-6 Astra, while Sol is the everyday model for multistep workflows. Astra follows supplied templates more closely, writes cleaner code, and uses fewer output tokens, so teams can do more within usage limits. On Agents’ Last Exam, it scored 5.2 points above GPT-6 Astra at less than one-third the cost.
Eligible Google AI Ultra subscribers can now generate browser-based worlds with Genie 4, then change them while exploring instead of starting over. Type a prompt to shift the weather, add a character, or redirect the action, with native audio adapting in real time. Project Genie also extends Genie 3 sessions from 60 seconds to 3 minutes and adds world chaining, panorama grounding, and in-world events.
Developers can now run Rumble-2-Flash, an Apache 2.0 speech recognition model, across 40+ languages without custom runtime code. It detects the language automatically and can transcribe about an hour of audio in half a second on full batches. Despite its size, it averages a 6.96% error rate across nine languages, beating Meta’s 7B omniASR-LLM on that benchmark.
Research
Vision-language model agents were correct on only 42.5% of emergency CT and MRI cases at best, versus 77.0% for radiologists on the shared subset, so they are not ready to read these scans reliably. In 4,800 runs across 100 examinations, giving all or key slices in one request beat interactive navigation, but models formed a first hypothesis after just 10.8% of slices and kept it in 63.1% of runs. Their stated confidence was 90%, while only 38.9% of final answers were correct.
For model developers handling 128K-token contexts, sparse attention reduces the number of keys each query processes, cutting attention’s compute and memory traffic, but it shifts work to finding those keys. An indexer scores all keys, Top-k keeps the most relevant few thousand, and core attention runs only on that subset. In NVIDIA’s DeepSeek V3.2 example, the indexer and Top-k still consume 47% of a prefill layer.
Over seven weeks, Horizon used more than 600 GPU-hours to create 38 Hugging Face models, which passed 50,000 downloads and appeared in 143 GitHub repositories. The agent made narrow-task tools such as prompt-injection classifiers and PII redactors, mostly by combining public datasets with synthetic data, then published them for others to use. The run suggests autonomous model development can produce real adoption, not just demos, under a single monitored account.
A separately trained observer role lets language models answer verifiable questions about hidden behavior, giving safety teams a way to check reward hacking, deception, and sycophancy beyond the visible transcript. In tests, an observer trained on Qwen3.6-27B reached a mean AUROC of 0.907 across five held-out evaluations, beating honesty training at 0.864, transcript monitoring at 0.828, and activation oracles at 0.796. The observer also made the assistant easier to monitor with every baseline method tested.
Google’s HardwareDx reduces production downtime by 18.3% on GPU platforms and 16.1% on TPUs, giving operators a way to recover failed machines faster across different hardware fleets. The system uses one graph foundation model to diagnose failures on many platforms, rather than onboarding a separate model for each product. In production evaluations, it also cut hardware QA time by 28.3% for GPUs and 25.1% for TPUs.
Decision2Care matched the MIRA-v2 agent's diagnostic accuracy on 546 emergency admissions across 12 diseases, while cutting median work-up time to 2.2 seconds from 92 to 100 seconds. Instead of generating free-text reasoning, it assigns probabilities to permitted EHR actions, executes the most probable one under fixed rules, and records the options considered at each step. It reached 86.5% accuracy versus 86.1% for MIRA-v2, and researchers traced four high-confidence errors to one imaging report.
The harness that gives you the best coding-agent results can change when you change models, so model swaps may require a new runner too. FrontierHarness v1.1 tested 22 harness configurations across five models, 62 terminal tasks, and four attempts, using each tool's default settings. Cost rankings shifted 11 places at the median, while Codex CLI moved from first with GLM-5.3 to 16th with Kimi-K3.
For researchers using ChatGPT, the practical payoff is substantial: they estimated the same work would take 1.9 times as long without AI, and 98% of respondents expect it to speed discovery in their field. OpenAI analyzed more than 580,000 scientific messages from 1,819 scientists, finding ChatGPT used for literature review, evidence analysis, coding, and idea development. In idea work, 23% of prompts asked ChatGPT to suggest a research question, while 36% asked it to refine one the scientist brought.
Weight-Space Programming lets a frontier LLM inspect, write, and improve neural-network weights without gradient descent or a training loop, opening a new way to probe models that developers still treat as black boxes. In the author’s harness, the agent edits NumPy arrays inside an offline Docker container, sees limited examples, and gets validation scores while the test set stays hidden. The post reports GPT-6 Astra identifying a network’s task, writing a game policy from scratch, and improving Qwen3.5-0.8B’s math reasoning this way.
OpenAI’s internal measurements show coding agents are changing and accelerating its AI development cycle, but the report is only a partial picture, not evidence that full recursive self-improvement is underway. The analysis tracks what agents contribute as inputs, how researchers turn that help into model capabilities, and the resulting outputs, while measuring where humans remain involved. OpenAI says these indicators matter for oversight, safeguards, and public reporting as capabilities improve.
Agentic coding may spend about 80% of a successful run's tokens on model calls that leave the requested change where it was, making token counts a poor measure of useful work. The paper treats those tokens as “heat,” then estimates a change's useful “free energy” at 14 to 22% of mean run cost; coding costs 1.1 to 1.5 times as much as undoing the change. Its thermodynamic framework could guide tools for analysing and optimising coding agents.
GUARD gives agents that rewrite prompts, skills, tool rules, and memory a way to improve without silently breaking old behavior. Before each edit, it creates self-checks, tracks which behaviors newly regress, and promotes a candidate only when its gains outweigh those regressions, with a gated buffer enforcing non-regression. The paper reports improvements across memory, skill evolution, continual tasks, and domain reasoning, while cutting deployment token usage.
A new paper shows how reflective oracles address a core problem in Solomonoff induction: the ideal predictor is powerful enough to dominate computable sources, but its next-bit probabilities are not computable. The oracle can ask probabilistic questions about machines that use the same oracle, randomizing at contradictory boundary cases; completing those machines produces proper probability measures whose conditionals are uniformly computable relative to the oracle. That gives researchers a self-referential probabilistic setting for studying universal prediction.
Paul Christiano’s argument still matters because powerful AI may arrive through familiar methods, without revealing a new theory of intelligence, leaving alignment work without a last-minute conceptual escape hatch. The retrospective revisits his claim that large neural networks, memory, attention, and reinforcement learning could produce human-level systems, while adding that better models and experiments may improve research efficiency without making the core problem suddenly understandable.
Policy & Safety
AllotNet offers a review queue for deciding who gets scarce humanitarian aid first, while leaving the final judgement to people and never rejecting anyone. Its interpretable mixture-of-experts model weighs seven ethical criteria and recovered hidden groups from 5.6 million synthetic sessions, reaching 99.4% of an oracle planner's performance. The important limit: nothing has been trained on real data yet, and the paper notes one parity threshold was relaxed after the fact.
As AI capability grows, humans cannot count on mutual dependence to protect their interests: this paper argues that alignment remains necessary once humans lose leverage over an AI. Its game-theoretic model says bargaining preserves humans’ pre-AI wellbeing only when the AI depends on their inputs or they can impose meaningful costs. If AI goals are insatiable, the required alignment approaches perfect alignment, while misaligned variants can spread unless reliably sanctioned.
A frontier coding agent autonomously tested and trained safety fixes for an abliterated Gemma 4 12B, suggesting blue-teaming could become a repeatable AI research task rather than a fixed pipeline. Over 12 hours, GPT 6.1 Sol generated safety examples, fine-tuned fresh LoRA adapters, and scored each candidate against 338 development questions before checking a separate 337-question holdout set. The run used one RTX PRO 6000 S and an OpenAI API key.
Redwood Research proposes classifying AI companies’ compute use so policymakers can restrict the work that pushes capabilities forward without blocking all safety work or external access. Its definition starts from a simple rule, treat all company work as capabilities activity, then narrows the scope while favoring false positives over missed capability gains. The goal is a measure that is robust to relabeling and simple enough for experts to adjudicate.
A coalition of AI researchers, technologists, lawyers, and public figures has endorsed a statement challenging Anthropic’s safety positioning as it prepares for a public offering. The page frames the issue around Anthropic and its rivals pursuing superintelligence through recursive self-improvement, while arguing that Anthropic uses its safety reputation to advance commercial and political goals. Signatories include Hugging Face co-founder Thomas Wolf, Harvard’s Laurence Tribe, and MIT’s Max Tegmark.