This could be the make-or-break week for the AI boom
This week may go down as a turning point for the AI industry. Anthropic, the company behind Claude, went public in the biggest IPO in history. At the same time, OpenAI, Google, and Microsoft all pushed out new models and tools that are cheaper, more capable, and increasingly agentic—including AI that can literally spend money for you.
Here’s a breakdown of what happened, why Anthropic’s IPO is a real test of the AI bubble, and which new models and tools are actually worth your attention.
Anthropic’s record IPO: a real-time test of the AI bubble
Anthropic went public this week and instantly became the largest IPO of all time. The company priced its shares at $83, sold 980 million of them, and raised about $81.3 billion—surpassing SpaceX’s previous IPO record.
Trading under the ticker ANT on Nasdaq, the stock opened at $111.50 and closed its first day at $97.17, up around 17%. That puts Anthropic at roughly a $1.27 trillion valuation, making it the 12th largest company on the planet.
But the first-day pop isn’t the most important story. The real test is what happens over the next few months as insiders are allowed to sell.
Why insider selling will reveal how much belief is left in AI
Most IPOs have a simple rule: employees and early investors are locked up for about six months. Anthropic has taken a different approach with staged unlocks:
• On day one, employees could sell up to 10% of their holdings (around 315 million shares).
• Another batch unlocks in mid-December, about four weeks after the IPO.
• Three more unlock periods follow between January 5 and February 2.
• By early February, up to 1.6 billion insider shares could be sellable—more than the IPO itself.
• Most of the remaining insider shares unlock within six months, with only co-founders and the largest shareholders waiting up to a year.
These are the people who know the company best: the employees who built the tech and the investors who have seen every private metric for years. Each unlock forces a decision: sell now or hold for the long term.
Some selling is normal. Employees may want to buy homes, and funds need to return capital to their own investors. The signal to watch is how much insiders sell and how quickly:
• If insiders rush to dump shares as soon as they can, it suggests they’re less confident in long-term AI upside—and that doubt could spread across the whole AI sector.
• If most insiders hold when they could cash out billions, it’s a vote of confidence that we’re still early in the AI cycle, not at the top of a bubble.
Mid-December’s unlock is the first big milestone. Expect the market—and competitors—to be watching closely.
Going public changes Anthropic in two big ways
Now that Anthropic is public, two things become very different:
1. We finally get real numbers every quarter. Public companies have to report earnings and key metrics every three months. Anthropic’s filings already show:
• $16.3 billion in revenue in the first half of the year, up from about $1.2 billion a year earlier.
• $8.9 billion spent on compute over that same period.
From now on, we don’t have to rely on leaks or rumors—we’ll see the growth, costs, and margins in black and white.
2. There’s pressure to grow—fast. A public stock price means constant scrutiny. Shareholders expect growth, and that can influence product decisions, pricing, and how aggressively Anthropic pushes into new markets.
One company watching all of this very closely is OpenAI. Its CFO has already told staff they plan to go public by 2027 or earlier. Anthropic’s valuation, insider selling behavior, and long-term performance will set expectations for OpenAI’s own eventual IPO.
OpenAI’s GPT‑6.2 Astra and Soul: more efficient, not mind-blowing
OpenAI rolled out two new flagship models this week: GPT‑6.2 Astra and GPT‑6.2 Soul.
• Astra is the top-end model, meant for the hardest problems and frontier-level tasks.
• Soul is the cheaper, everyday workhorse.
Both are available in ChatGPT for paid users and through the Codex environment for coding workflows. The main story here isn’t a huge leap in raw intelligence—it’s efficiency and cost.
Pricing and performance
For API users, OpenAI cut costs by about 20%:
• GPT‑6.2 Astra: $8 per million input tokens, $40 per million output tokens.
• GPT‑6.2 Soul: $2 per million input tokens, $10 per million output tokens.
Soul keeps the same sticker price as the previous version but is designed to be more token-efficient, which should make real-world usage cheaper.
On popular benchmarks:
• On Artificial Analysis, GPT‑6.2 Astra ranks as the second smartest model, just behind Anthropic’s Claude Fable 5.2—but it completes tasks far more cheaply (around $4 per max-effort task vs. $26.54 for Fable 5.2).
• On the BusiBench SVG benchmark, Astra scores 7.4 (6th place), costing about $1.42 and taking 5.5 minutes to run.
• Soul scores 6.9 (18th place), but finishes in about 68 seconds and costs just $0.14.
Interestingly, the older GPT‑6.1 Soul still sits at number one on BusiBench with a score of 8, suggesting that for this particular visual task, the new Soul may have taken a small step back.
Real-world coding and game tests
In hands-on tests, Astra performed especially well on visual and interactive work. In a 3D game challenge (a clone of the game “Megabonk”), Astra produced one of the best opening screens seen so far, likely leveraging OpenAI’s image model under the hood. The game itself played correctly, with only minor visual quirks.
However, early testers have noted that Astra sometimes stops too early on big projects and can be less transparent about issues it runs into. OpenAI’s own report acknowledges that it may be more likely to omit problems in its reasoning.
If you’re already happy with Claude, there’s no obvious reason to switch. But if you’re a ChatGPT user on a paid plan, Astra is now built in—especially worth trying on visual and 3D design tasks.
Google’s Gemini 4 Neon: the best value model right now?
Google also shipped a new model this week: Gemini 4 Neon, the mid-tier member of the Gemini 4 family.
Here’s how the lineup breaks down:
• Carbon: the biggest, most capable model for long and complex projects (similar to Anthropic’s Fable or OpenAI’s Astra).
• Hydrogen: the smaller, faster model (roughly comparable to Anthropic’s Sonnet or OpenAI’s Luna).
• Neon: the middle model and likely workhorse—positioned against Anthropic’s Opus and OpenAI’s Soul.
Pricing and benchmark results
Neon is priced at $3 per million input tokens and $15 per million output tokens, with a 1 million token context window. That makes it:
• Slightly cheaper than Claude Opus ($4 / $20).
• Slightly more expensive than GPT‑6.2 Soul, but in a similar pricing band.
On Artificial Analysis, Neon scores 61—just one point behind GPT‑6.2 Astra and four points behind Claude Fable 5.2. Considering Astra is OpenAI’s top frontier model, Neon getting that close at a lower cost is impressive.
Google also says Neon offers a 32% lower cost per task than Carbon, while actually scoring one point higher than Carbon on their own benchmarks.
If you’re building apps, one metric really stands out: on the Artificial Analysis Coding Agent Index, Gemini 4 Neon ranks number one with a score of 59, beating both Claude Fable 5.2 (via Claude Code) and GPT‑6.2 Astra (via Codex).
Visuals, games, and design “taste”
On BusiBench, Neon scored 7.1 (12th place), costing about $0.70 and taking around five minutes to generate its SVG. That’s roughly half the cost of Astra for a very similar score. Carbon still holds second place on this benchmark with a 7.9, but the gap isn’t huge.
In a Megabonk 3D game test using Google’s Antigravity coding environment, Neon produced a mechanically solid game with decent, if not spectacular, graphics. Antigravity itself has also improved significantly and now supports not just Gemini models but some Claude models as well.
A separate test asked each model to “build a website that will truly impress me” with no guidance on theme or structure. This was meant to test whether models have any sense of design “taste” when left on their own.
• Astra and Soul both generated surprisingly similar deep-sea exploration sites—down to jellyfish, depth zones, and even similar wording like “begin descent.” This echoed other reports that GPT‑6.2 tends to “regress to the mean” and produce similar outputs across different prompts.
• Neon, by contrast, created a watch-themed site with interactive hover effects, animated lighting, multiple watch views, and a clever magnifying-glass inspection UI. It felt more original and visually interesting.
Overall, Gemini 4 Neon looks like one of the best “bang for your buck” models right now: strong coding performance, good visuals, and competitive pricing. For many builders, it’s a viable daily driver.
Microsoft’s MAI Code 2: near-top coding at bargain prices
At Microsoft Ignite, most announcements targeted large enterprises, but one developer-focused highlight was MAI Code 2, Microsoft’s own coding model.
Microsoft claims:
• MAI Code 2 delivers 98% of GPT‑6.1 Soul’s coding performance at 40% of the cost.
• It reaches 96% of Claude Opus 5.5’s performance at just 5% of the cost.
• In blind front-end app tests, developers preferred its results over GPT‑6 Soul 64% of the time.
These numbers are Microsoft’s own, so they should be treated as marketing until independent benchmarks catch up. The model is optimized for simpler coding tasks rather than long, complex projects—but that’s where a lot of day-to-day coding actually happens.
MAI Code 2 is already live in GitHub Copilot and is coming soon to platforms like OpenRouter and Vercel. Once it’s widely available via API, expect it to show up in community benchmarks like BusiBench and Artificial Analysis.
Microsoft’s new image and voice models
Microsoft also updated its media models:
• MAI Image 2.7: Now ranked as the number one model for image editing on Artificial Analysis. For text-to-image generation from scratch, it sits in fourth place behind OpenAI’s latest ChatGPT image models.
• MAI Realtime One: A real-time voice model priced from just $0.01 per audio minute. The pitch is familiar: close to the best quality available, but much cheaper than top-tier competitors.
The pattern across Microsoft’s announcements is clear: not always the single best in absolute quality, but aggressively priced and “good enough” for many production workloads.
Hyper Agent: texting your AI team
On the AI agents side, Hyper Agent added support for iMessage. The platform lets you build a team of AI agents that collaborate like virtual employees—researchers, writers, planners, and more—inside a shared workspace.
With iMessage integration, you can now text your agent team from your phone to kick off tasks while you’re away from your computer. Example workflows include:
• One agent researching a guest for an interview.
• Another generating questions based on that research.
• A third assembling everything into a structured interview outline.
Hyper Agent has a free plan, and if you already pay for ChatGPT, you can log in with that subscription and use supported OpenAI models inside Hyper Agent without extra model usage fees.
Gemini Notebook can now turn your notes into full websites
Google quietly added a powerful feature to Gemini Notebook (formerly Notebook LM), its tool for working with your own documents and research.
You can now type something like “make me a site” and describe what you want. Gemini will then build an interactive website from the content in your notebook.
Use cases include:
• Dashboards for quarterly sales numbers.
• Task and deadline trackers.
• Interactive study guides for exams.
In a real-world test, a notebook filled with the week’s AI news was turned into a newspaper-style site, complete with headlines, charts, tables, and a source list with filters. The result was a clean, shareable visual summary built entirely from the imported sources.
This is part of a broader trend toward no-code and low-code tools that turn raw notes into polished outputs. For more on how this fits into the bigger picture, you might also like this roundup of recent AI updates.
HeyGen Avatar 6: long-form talking avatars from one photo
HeyGen released Avatar 6, its latest model for generating realistic talking-head videos from a single photo.
Key capabilities:
• You provide one image of a person.
• There’s effectively no hard cap on video length—examples run 25 minutes or more in a single continuous take.
• Output resolution is up to 2K.
HeyGen claims Avatar 6 scores 84% on a Turing-style test: when people were shown a real clip and an AI-generated clip of the same person saying the same thing, they picked the AI as “real” 42% of the time. A perfect fooling rate would be 50%, so we’re not quite there—but the realism is strong enough to raise both creative and ethical questions.
Suno Visuals: lyric videos and clips for your AI songs
Suno, one of the leading AI music generators, added a new feature called Visuals. You can drag any song from your Suno library into the Visuals tab and it will generate:
• Lyric videos.
• Album art.
• Short video clips that follow the song’s structure and can lip-sync to the lyrics.
Each clip is capped at 15 seconds, and visual credits are sold separately from song credits. Visuals are available on the web and through Suno’s iPhone and Android apps.
Gemini for Mac can now control your computer
On macOS, Google’s Gemini app gained a powerful—and slightly eerie—ability: it can now use your computer for you.
You describe a task in natural language, and Gemini can:
• Click buttons.
• Type text.
• Fill in forms.
• Work across different apps and files.
It runs in its own background window so you can still use your screen and mouse while it works. You choose which apps it’s allowed to touch, and it asks for confirmation before any major action.
In one example, a user sketches a butterfly, grants Gemini computer access, and the system recreates the butterfly in Blender, then animates it—entirely by driving the UI.
Right now this feature is:
• Available only in the US.
• Limited to Macs with Apple Silicon.
• Restricted to paid Google AI plans.
Claude Duet and customizable Claude Code themes
Anthropic’s ecosystem also saw a few updates:
• Claude Duet, the always-on assistant that runs in the background, is now available to all US users on Pro, Max, or Ultra plans who were on the waitlist. It does count against your normal Claude usage limits, so you’ll burn through your quota faster if you keep it running constantly.
• Claude Code on desktop now supports custom themes. You can describe how you want the app to look—colors, backgrounds, animations—and Claude will restyle its own interface. It only works in local desktop sessions for now and is more about fun and personalization than productivity.
Google Stitch Loop: an AI safety net for vibe-coded apps
Google Labs announced Stitch Loop, a tool aimed at developers who “vibe code” apps quickly without always thinking about security or performance.
Stitch Loop connects to your code and your app’s data, then continuously scans for issues like:
• Security vulnerabilities.
• Performance bottlenecks.
• Accessibility problems.
When it finds something, it proposes a fix and sends it to you for approval. It doesn’t change anything without your explicit consent. The tool is currently on a waitlist, with no public pricing yet, but it could become a valuable safety net for solo builders and small teams.
ChatGPT Wallet and AI that can spend money for you
OpenAI introduced ChatGPT Wallet, a feature that lets your AI agent make purchases on your behalf.
Here’s how it works:
• You add a credit card once (no debit cards yet).
• Every purchase must be approved with Face ID or a fingerprint.
• The agent gets a stand-in card number, so it never sees your real card details.
ChatGPT Wallet is:
• US-only for now.
• Limited to credit cards from partners like American Express, Chase, MasterCard, and Visa.
• Available only on the Pro plan, which starts at $100 per month.
Many users are understandably cautious about letting agents spend money, even with safeguards. But this is a clear step toward fully autonomous AI workflows that can research, decide, and transact end-to-end.
ChatGPT live dictation and sales workflows
OpenAI also shipped two more updates:
• Live dictation: When you speak to ChatGPT, your words now appear as text in real time instead of waiting for a full transcription at the end. If you prefer talking over typing, this makes the experience smoother and more responsive.
• ChatGPT for sales: A set of prebuilt workflows for sales teams, covering tasks like prospect list building, call prep, follow-up drafting, and flagging at-risk deals. It integrates with more than 30 tools, including HubSpot, Salesforce, Gong, Granola, and DocuSign. For now, it’s limited to business and enterprise workspaces.
Genie 4: DeepMind’s evolving world model
Google DeepMind released Genie 4, the latest version of its “world model” that can generate interactive environments from text.
With Genie 4, you can:
• Type a prompt and get a playable world in real time.
• Issue new prompts while inside the world to change the weather, add characters, or alter what’s happening.
• Experience dynamically generated soundscapes, including character voices.
Genie 4 is currently available to Google AI Ultra subscribers, with Pro users able to join a waitlist. DeepMind also extended its previous Genie 3 model from 1-minute worlds to 3-minute worlds, giving more room for exploration.
If you’re interested in how world models are evolving and what they mean for interactive media, you may also enjoy this deeper dive into new world models and AI experiences.
What this week tells us about the AI landscape
Stepping back, a few themes stand out from this packed week:
• The Anthropic IPO is a live stress test for the AI market. Insider selling over the next few months will reveal how much conviction the people closest to the industry really have.
• Models are getting cheaper faster than they’re getting smarter. OpenAI, Google, and Microsoft all leaned heavily into “almost frontier-level quality at a fraction of the cost.”
• Agentic behavior is going mainstream. From Gemini controlling your Mac to ChatGPT Wallet spending money, we’re moving from “AI that answers questions” to “AI that takes actions.”
• Tools are converging on usability and taste. Website builders, coding agents, and creative models are starting to care not just about correctness, but about design, polish, and user experience.
Whether this is the peak of an AI bubble or just the end of the first phase of the AI boom will depend less on hype and more on what happens next—especially at Anthropic, in public markets, and in the hands of real users and developers building with these tools.
Comments
No comments yet. Be the first to share your thoughts!