Is the AI boom really ending? New laws, new models, and what’s coming next
The AI world isn’t slowing down—if anything, it’s getting more chaotic. On one side, US election results have triggered loud calls to pause or even ban advanced AI. On the other, OpenAI, Anthropic, Google and several up‑and‑coming labs are quietly (and not so quietly) preparing their next wave of powerful models, tools, and infrastructure.
Is the AI boom really ending?
Recent US midterm elections saw Democrats take both the House and the Senate, and AI policy quickly jumped near the top of the agenda. Exit polls suggested that more than half of voters saw AI and data centers as a crucial or major issue. That political momentum has already turned into aggressive regulatory proposals.
One of the most high‑profile ideas is the Ban Artificial Superintelligence Act, backed by Bernie Sanders and Greg Casar. It calls for:
• A legal ban on “superintelligent” AI systems
• A pause on the most advanced AI development until a new federal regulator sets rules
• A cabinet‑level Department of AI
• Harsh penalties, including up to 20 years in prison for individuals and even a corporate “death penalty” for violating companies
Some newly elected lawmakers are also openly advocating for a “full pause” on frontier AI research in the US, arguing that if America slows down, rivals like China will slow down too. Critics point out that this view ignores how aggressively Chinese labs are pushing forward, often open‑sourcing their work and rapidly closing performance gaps with US models.
The real battleground: data centers
While a blanket pause on advanced AI is unlikely to get enough votes (or survive a presidential veto), local fights over data centers are already having real impact. On election day, dozens of communities across 13 states voted on whether to allow or block new data center projects.
In many cases, residents voted to ban or heavily restrict them, sometimes even recalling local officials who supported big AI infrastructure deals. That’s a serious problem for US tech companies, because high‑end AI training and inference require enormous amounts of compute, power, and land.
Big tech is already responding by looking abroad. Google, for example, announced a $15 billion investment in Malaysia to build new data center campuses, clean energy capacity, and tens of thousands of construction jobs. If US regions push AI infrastructure away, countries in Asia and elsewhere are ready to welcome it.
Economists estimate that AI‑related investment was already responsible for the vast majority of US GDP growth in early 2025. Pushing that investment offshore doesn’t stop AI progress—it just shifts where the benefits land.
Leaked: GPT‑6.2 Astra, Soul and a cyber model
Regulation aside, OpenAI appears to be preparing one of its biggest model waves yet. According to reporting from The Information and developer leaks, three new models are on the way:
• GPT‑6.2 Astra – a high‑end multimodal model optimized for advanced reasoning and complex web and app development tasks
• GPT‑6.2 Soul – a slightly lighter sibling that aims to approach Astra‑level performance while offering much higher usage limits
• A dedicated GPT‑6.2 cyber security model – focused on finding and exploiting vulnerabilities at scale
References to GPT‑6.2 Astra and Soul have already appeared in Google Cloud’s console, suggesting that launch is close. OpenAI also has a pattern of announcing major updates on Thursdays, adding to speculation that these models could appear any day.
Early GPT‑6.2 Astra checkpoints look strong
Developers who have access to Astra‑like checkpoints are already stress‑testing them. Early results show:
• Much stronger web development capabilities, including complex animated sites and SVG‑based graphics
• Fewer rendering and structural errors compared to earlier GPT‑6.x checkpoints
• Still somewhat basic UI design sensibilities compared to top design‑focused workflows
In one test, the model generated an entire hand‑drawn anime‑style website: a child flying a red kite that tears open storm clouds to reveal a sunrise, complete with transitions, lighting effects, and componentized code. In another, it recreated a detailed Pokémon‑style SVG scene purely from code, without relying on reference images—something many current models still struggle with.
It’s not yet clear whether this is the final Astra release quality, but it already looks competitive with leading frontier models like Anthropic’s upcoming Fable 5.5 in many coding and graphics tasks. For more background on the GPT‑6 line and how it fits into the broader race, you can check our earlier piece on GPT‑6 Soul leaks and Gemini 4 pressure.
GPT‑6.2 Soul: the potential workhorse model
While Astra will likely grab the headlines, GPT‑6.2 Soul may be the model that most users actually live in. The goal appears to be:
• Near‑Astra performance on everyday reasoning and coding
• Much higher rate limits than the current $200‑tier plans
• A better balance between power and affordability for heavy users and teams
If Soul can really deliver close to Astra‑level quality with generous usage caps, it could become the default choice for many developers, startups, and agencies who constantly hit today’s limits.
OpenAI’s new “family basic” plan and talent grab
On the product side, OpenAI seems to be testing a new “family basic” subscription tier. Hidden references in the ChatGPT checkout flow suggest a plan designed for multiple users sharing one subscription—ideal for households or small teams that don’t need full enterprise features.
At the same time, OpenAI is aggressively hiring from Anthropic. One notable move: the product leader who helped turn Claude Code from a small internal experiment into a powerful coding agent has now joined OpenAI to lead core product management, likely focused on its coding tools. Another senior Anthropic alum has reportedly made the jump as well.
Given how strong Claude Code has become, these hires signal that OpenAI is serious about catching up—or leapfrogging—in AI‑assisted software development.
Images 2.6, Flare and Sunburst: OpenAI’s visual upgrades
OpenAI is also rolling out a major upgrade to its image stack under the GPT Image 2.6 umbrella. Key improvements include:
• Native GIF generation from text prompts or existing images, with control over speed and frame‑level editing
• A reference image system that lets you save characters, faces, pets, or styles and reuse them across generations without re‑uploading
• Faster generation—OpenAI claims up to 50% speed improvements
• New markup and SVG export options for more precise editing and integration into web workflows
Two specialized developer models are also launching via API:
• Flare – optimized for speed, with image generation in as little as ~6 seconds
• Sunburst – focused on higher fidelity and consistent edits across iterations
These models are already available in ChatGPT and Codex on web, desktop, and mobile, and Flare and Sunburst can be accessed through the API for custom pipelines.
Fable 5.5: Anthropic’s wild new creative and 3D abilities
Anthropic’s upcoming Fable 5.5 model keeps showing off increasingly impressive internal checkpoints, even though it hasn’t been publicly released yet. Recent demos include:
• A fully animated scene called “The Birth of Life,” where the model generated the entire animation logic in code (no text‑to‑video model involved)
• Converting a 2D painting of a creepy face into a fully interactive 3D object inside Blender that users can manipulate in real time
• A short film built entirely in JavaScript and Python, with the model writing the code that drives every shot and transition
These examples hint at a model that’s not just good at text and code, but at orchestrating complex multi‑tool creative pipelines—exactly where many believe next‑generation AI agents are headed.
Claude Code gets cheaper and more efficient
Anthropic has also shipped a practical but important upgrade to Claude Code. Previously, background “sub‑agents” used to periodically reread their entire conversation every 30 seconds to update status, consuming unnecessary tokens and eating into user limits.
Now, Anthropic routes those status updates through a lightweight Haiku 5.5 process that summarizes changes using short notes instead. The result:
• Status updates are about 99% cheaper in token usage
• Overall session usage drops by ~5% for affected workflows
• Users effectively get around 6% more prompts under the same limits
This change doesn’t downgrade the main coding model used by the agent; it simply optimizes the bookkeeping. The update is rolling out across Pro, Max, Ultra, Team, and Enterprise plans for Claude Code (but not the desktop app yet).
Anthropic’s invisible text watermarking
Starting November 16, Anthropic is expanding its AI text watermarking to Claude Opus 4.6 and the entire 4.5 family. Unlike visible watermarks, this system uses subtle statistical patterns in word choice and phrasing so that:
• AI‑generated text can be identified with specialized detectors
• Readability and style are minimally affected for human readers
The big open question is robustness. If someone paraphrases the text, edits it heavily, or runs it through another model, will the watermark survive? Anthropic’s rollout will be an early test of how effective this approach can be at scale.
Claude helps crack a 40‑year biology mystery
Beyond coding and chat, Anthropic shared one of the most compelling scientific use cases for AI so far. Researchers used Claude to help uncover the first 3D structure of a vault‑like particle found in bacteria—related to mysterious barrel‑shaped protein complexes first discovered in 1986.
Claude reportedly helped to:
• Search through 15 billion protein sequences to find promising candidates
• Cut a purification process from 3 days down to about 90 minutes
• Generate an initial structural model in roughly 34 hours instead of weeks or months
Human scientists still had to validate and correct the results, but the collaboration shows how frontier models can accelerate real scientific discovery, not just write code or marketing copy.
OpenAI’s cyber model: 948 Linux kernel exploits
One of the most striking security demos comes from OpenAI’s upcoming cyber security model. In a controlled experiment, an autonomous pipeline powered by this model:
• Discovered and successfully exploited 948 unique Linux kernel privilege escalation vulnerabilities
• Identified at least 10 potentially novel exploitation techniques that Linux security experts hadn’t documented
• Reproduced more than 1,700 known vulnerabilities, generating working exploits and even proposing patches
The campaign cost about $1.21 million in API credits—roughly $1,279 per successful exploit. OpenAI’s takeaway is stark: on unpatched systems, low‑privilege access may effectively be equivalent to root access for a well‑resourced attacker armed with such a model.
Labs are framing these systems as defensive—tools for finding and fixing vulnerabilities before attackers do. But they also underline why governments are so focused on AI safety and cyber risk.
OpenAI buys Pierre Computer and aims beyond GitHub
OpenAI is also moving deeper into developer tooling by acquiring Pierre Computer, a startup focused on code storage and open‑source developer infrastructure. Pierre’s technology already powers parts of Codex, OpenAI’s coding environment.
The plan reportedly is to turn Codex into more than an AI coding assistant. Instead, OpenAI wants a full development platform where:
• Developers can store, review, and collaborate on code
• AI agents can help manage Git workflows, refactors, and deployments
• Entire software projects can move from idea to production inside a single AI‑augmented environment
If successful, this would position Codex as a direct competitor not just to GitHub Copilot, but to GitHub itself. For more context on how OpenAI has been evolving its agent and coding stack, see our earlier coverage of OpenAI’s leaked GPT‑5.5 and agent platforms.
Moonshot’s Kimi K 3.1 Pro: a 2.8T‑parameter openweight giant
Chinese lab Moonshot AI has released the model weights for Kimi K 3.1 Pro, one of the largest openweight models to date. Specs include:
• An enormous 2.8 trillion parameters
• Model weights around 1.56 TB in size
• Hardware requirements on the order of eight NVIDIA B300 GPUs with ~2.3 TB of GPU memory
A server capable of running this model could cost in the range of $550,000 to $650,000. That makes Kimi K 3.1 Pro more of a research and enterprise tool than something you’d run on a desktop—but it’s a strong signal that openweight models are scaling rapidly alongside closed ones.
Ripple 65B: a new openweight coding contender
Recursive Intelligence has introduced Ripple 65B preview, a dense 65‑billion‑parameter model that’s already punching above its weight in coding benchmarks. Highlights:
• Trained from scratch on 12 trillion tokens with a compute budget around $32 million
• Achieves ~80.5% on TerminalBench 2.1, edging out some much larger American openweight models in agentic coding tasks
• Uses all 65B parameters each step (unlike mixture‑of‑experts models that only activate a subset)
• Supports a 1 million token context window
• Released under an Apache 2.0 license, making it attractive for commercial use
Ripple can run on heavier local setups, such as NVIDIA DGX‑class hardware, and is positioned as a state‑of‑the‑art US openweight option for building coding agents.
Gemini 4 “Hydrogen” and Google’s latest upgrades
Google is quietly testing a new flash‑tier model internally known as Gemini 4 Hydrogen (sometimes referred to as Gemini 4 Flash). Early checkpoints have appeared in Google AI Studio and developer tools, showing:
• Extremely fast generation speeds around 300–500 tokens per second
• Surprisingly strong visual quality for a speed‑optimized model
• A design focus on lightweight, high‑throughput applications
Alongside Hydrogen, Google has several Gemini updates:
• Computer Use tool GA: Now generally available in the Gemini API, letting agents interact with browsers, desktop apps, and mobile interfaces (clicking buttons, filling forms, navigating UIs) without breaking changes for existing users.
• Live translation in Google Meet: Powered by Gemini 3.5 Live Translate, Meet can now translate between 63 languages in real time, automatically detecting the speaker’s language and converting it into the listener’s choice.
• Voice remixing for Gemini 3.8 Flash TTS: Developers can choose from 2,000 preset voices or design custom ones, then tweak pitch, tone, speed, and accent using simple text prompts.
• Gemini Omni 1.2 Flash: The model previously known as Omni Flash 1.5 is now generally available via API, offering incremental improvements as part of Google’s growing lineup.
Sway 2 Flash: faster, cheaper coding agents
Cognition has launched Sway 2 Flash, a coding model tuned specifically for speed and efficiency in agentic workflows. According to the team, it:
• Outperforms every American openweight model on several coding benchmarks, including models up to 3x larger
• Achieves performance close to the full Sway 2 model at about one‑fifth the cost
• Generates tokens roughly 3x faster than the larger Sway 2
Sway 2 Flash is reportedly based on Inkling Small from Mira Murati’s Thinking Machines Lab. Due to cyber security concerns, Cognition isn’t broadly releasing weights yet; instead, they plan a controlled sharing program for selected researchers and partners.
In head‑to‑head tests against Anthropic’s Opus 5.5 on complex UI coding tasks, Sway 2 Flash holds its own functionally, even if Opus still wins on polish and design. The trade‑off: Sway 2 Flash is significantly faster and cheaper to run, making it ideal for high‑volume coding agents. Through November 30, it’s available for free in Devon desktop and the Devon CLI.
Claude Desktop storage controls
For heavy Claude Code users, Anthropic has added a small but welcome quality‑of‑life feature: a /sto command in the Claude desktop app. It lets you:
• See how much disk space Claude is using
• Clean up inactive tasks and unused Git work trees directly from the interface
As coding agents work across large repos, this kind of housekeeping becomes increasingly important to avoid silent storage bloat.
Rockbot gets phone call capabilities
On the agent side, Rockbot has added the ability to make phone calls in the US. Instead of only chatting with you, Rockbot can now:
• Place calls to third parties
• Speak on your behalf, following scripted or dynamic instructions
• Potentially handle tasks like booking appointments or resolving simple support issues
This is an early example of AI agents stepping beyond text and browser automation into real‑world communication channels.
Quen 3.8 27B: blazing local speeds on consumer GPUs
Finally, someone has managed to run an uncensored version of Quen 3.8 27B locally on a single AMD Radeon RX 7900 XTX with 24 GB of VRAM, achieving eye‑popping performance:
• Around 110 tokens per second during decoding
• Up to 8,000 tokens per second during prefill in some tests
• Stable performance on real agentic workloads
Because this build is uncensored, it avoids many of the typical safety refusals seen in hosted models and gives users more control over what they can do locally (with all the responsibility that entails). It’s a strong showcase of how far optimized inference and modern consumer GPUs have come.
The bottom line
Despite louder calls for pauses and bans, the AI boom is nowhere near over. If anything, it’s entering a new phase: governments are scrambling to regulate, local communities are pushing back on data centers, and labs are racing to ship more powerful models, agents, and infrastructure.
From GPT‑6.2 Astra and Soul to Fable 5.5, Gemini 4 Hydrogen, Ripple 65B, Sway 2 Flash, and giant openweight releases like Kimi K 3.1 Pro, the technical frontier keeps moving forward. The real question isn’t whether AI progress will slow—it’s which countries, companies, and developers will be best positioned to harness it safely and productively.
Comments
No comments yet. Be the first to share your thoughts!