Anthropic began trading on Nasdaq. OpenAI shipped Astra and Sol, while Google released Gemini 4 Neon. The contrast in Matt Wolfe's November 20 roundup is worth paying attention to: frontier labs need enormous amounts of capital, but developers have more options for getting the same job done at a lower cost.
In Wolfe's tests, Neon made the website he liked most, Astra produced the more polished-looking game, and Sol used the smallest SVG budget. Those results answer different questions. So does a company raising billions versus investors deciding what its shares are worth.
Watch the Weekly Roundup
Source: Matt Wolfe's November 20, 2026 video and transcript. Figures and rollout details were checked against filings, reporting and official release notes. Hyperagent's iMessage update at 10:36 is a paid sponsor segment, not an independent test.
Anthropic's IPO: What It Actually Raised
Wolfe leads with the record $81.34 billion base deal. That number is correct for the initial pricing, but it excludes the extra shares that were sold by closing. The three figures in circulation measure different things.
| Measure | Amount | What it includes |
|---|---|---|
| Base public offering | $81.34B gross | 980 million shares priced at $83 on November 18. |
| Final public offering | About $93.54B gross | 1.127 billion shares at $83 after the underwriters bought an additional 147 million. Includes shares sold by existing stockholders. |
| Public offering plus private placement | About $119.5B gross | The final public offering plus a separate $26.0B private placement, before fees and offering costs. Not the amount Anthropic alone kept. |
Anthropic's Class A stock began trading on November 19 as ANT. It opened at $111.50 and closed at $97.17, about 17% above the offer price but below its opening trade. Those are first-day prices, not the company's proceeds and not a current quote.
Wolfe is watching the staged lock-up releases. The final prospectus lists December 15, 2026 and January 5, 2027 among the scheduled dates. An unlock makes shares eligible to sell; it doesn't establish that anyone sold, or why. The filing offers a better starting point for judging the business: for the nine months through September 30, Anthropic reported $33.2B in revenue and a $1.3B operating loss. Watch how revenue, operating costs and compute commitments move together. One listing can't settle whether the whole sector is in a bubble.
Astra, Sol and Neon: What Wolfe Actually Tested
OpenAI positions Astra for harder coding and professional work, and Sol for routine work that still takes multiple steps. Both launched in ChatGPT Work, Codex and the API on November 17, with eligibility depending on the surface and plan. Google positions Gemini 4 Neon between Hydrogen and Carbon. It lists a one-million-token context window and access via the paid Gemini API and eligible Google AI Pro and Ultra plans.
| Model | Standard API input / output | Wolfe's website result |
|---|---|---|
| GPT-6.2 Astra | $8 / $40 | An ocean atlas with animated sea life, depth controls and sound. Wolfe found it less interactive than it looked. |
| GPT-6.2 Sol | $2 / $10 | Another deep-sea descent, with depth-linked scrolling and sonar. Some of its wording and ideas were strikingly similar to Astra's. |
| Gemini 4 Neon | $3 / $15 | A mechanical-watch showcase with a responsive spotlight and a magnifier. Wolfe preferred its visual direction and interaction. |
Prices are USD per million input or output tokens, before caching and tool charges. OpenAI's listed rates here are for short-context Standard requests; above 272,000 input tokens, a different rate applies to the full request.
The website test gave all three models the same open brief and let them pick the subject. That's useful for seeing what each model does without art direction. It doesn't test accessibility, maintenance, or whether a site delivers on a specified brief.
In Wolfe's SVG test, Astra scored 7.4 at $1.42, Neon 7.1 at $0.70 and Sol 6.9 at $0.14. These are his reported run costs, not API price quotes. In the Megabonk-style game demo, Astra produced the stronger opening screen but got the skateboard orientation and leveling pace wrong. Neon's version had simpler graphics but better-matched controls and mechanics in Wolfe's run. He didn't run Sol on that game.
Watch Wolfe's Full GPT-6.2 Review
Matt Wolfe's standalone GPT-6.2 video covers the SVG and game runs in more depth. Read our separate breakdown for the benchmark and evaluation context.
Microsoft's Ignite Releases: Lower-Cost Options to Test
- MAI Code 2 and Code 2 Flash: Both are listed in GitHub Copilot. Microsoft positions Flash for smaller, repeated changes. Across four coding benchmarks, Microsoft reports Code 2 reached 98% of GPT-6.1 Sol's performance at 40% of its cost. The comparison uses promotional MAI pricing and an older Sol version, not GPT-6.2 Sol.
- MAI Image 2.7 and Flash: Microsoft lists both in public preview on Foundry at about $35 and $10 per 1,000 images respectively. Its highest launch ranking was for image editing, a different task from generating an image from scratch.
- MAI Realtime 1: Microsoft's live speech-to-speech model starts at $0.01 per audio minute. The announcement lists a MAI Labs demo and a limited-access Foundry preview.
Gemini Notebook Builds a Site From Your Sources
In the hands-on demo, Wolfe supplied a notebook containing links for that week's news and asked it to make a site. In about two minutes, it returned a newspaper-like page with charts, tables and filterable source links. Some source sites had blocked ingestion, so the result couldn't replace checking the original reporting. In particular, verify any chart and what a valuation or price figure actually represents before sharing it.
Google announced a global web rollout for Gemini Notebook users aged 18 and older, including Workspace subscribers, with mobile to follow. A good first project is a study guide or internal dashboard from material you are allowed to share. Review the inputs and the finished site before sharing outside your team.
Wallet, Genie and the Other Updates
| Update | What changed |
|---|---|
| HeyGen Avatar 6 26:15 in the episode | HeyGen shows continuous talking-avatar videos from a photo, with an optional identity reference. The vendor demos include clips longer than 25 minutes at 2K. Use likenesses and voices only with permission; Wolfe did not test it in the episode. |
| Suno Visuals 27:10 in the episode | Make cover art, lyric videos or clips from a track. A new clip starts at up to 15 seconds and can be extended. Visual credits are separate from song credits. |
| Gemini computer use on Mac 27:38 in the episode | Gemini can work across selected apps and websites in a separate window. Google says the English-language US rollout requires an Apple Silicon Mac on macOS 15 or later, a personal account aged 18+, and AI Plus, Pro or Ultra. |
| Claude Duet 28:20 in the episode | Anthropic still describes its proactive assistant as a US research preview for Pro, Max and Ultra. Wolfe covers wider access for waitlisted users, not a global launch. |
| Claude Code desktop themes 28:42 in the episode | The Claude developer team shows custom desktop themes, including colors and animated backgrounds. Wolfe notes the feature is for local desktop sessions. |
| Stitch Loop 29:06 in the episode | Google Labs opened a waitlist for a product maintenance experiment. It can use live metrics to suggest improvements and open a pull request, but a human reviews before merge. |
| ChatGPT Wallet 29:47 in the episode | Initially rolling out through dots for eligible US ChatGPT Pro subscribers. Users review and approve each purchase; payment networks supply substitute credentials so the agent does not see the actual card details. |
| ChatGPT live dictation 30:30 in the episode | Words appear as text while you speak rather than after you finish recording. That is prompt dictation, not the same thing as a live voice conversation. |
| ChatGPT for Sales 30:44 in the episode | The Sales plugin and 30+ new or expanded partner plugins bring customer context into call prep, proposals and follow-ups. Business and Enterprise admins can enable it; verify drafts and CRM writes before they go out. |
| Project Genie: Genie 4 and Genie 3 31:10 in the episode | Genie 4 adds mid-session world changes and native audio for eligible Ultra users, with a Pro waitlist. The jump from one to three minutes per generation applies to the updated Genie 3, which also adds world chaining. |
For details, see Karan Ahuja's Duet access update and Wolfe's separate Genie 3 segment. Wallet's per-purchase approval and Stitch Loop's pull-request review are worth looking at together: both tools can do the preparation while leaving the decision with you.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Anthropic goes public |
| 05:33 | OpenAI launches GPT-6.2 Astra and Sol |
| 10:36 | Hyperagent adds iMessage support (sponsor) |
| 11:34 | Google launches Gemini 4 Neon |
| 16:35 | Astra vs Sol vs Neon: Website test |
| 21:55 | Microsoft introduces MAI Code 2 |
| 23:40 | Microsoft releases MAI Image 2.7 |
| 24:00 | Microsoft launches MAI Realtime 1 |
| 24:21 | Gemini Notebook can build websites |
| 26:15 | HeyGen launches Avatar 6 |
| 27:10 | Suno launches Visuals |
| 27:38 | Gemini for Mac gets computer use |
| 28:20 | Claude Duet expands access |
| 28:42 | Claude Code adds custom desktop themes |
| 29:06 | Google Labs introduces Stitch Loop |
| 29:47 | OpenAI launches ChatGPT Wallet |
| 30:30 | ChatGPT adds live dictation |
| 30:44 | OpenAI launches ChatGPT for Sales |
| 31:10 | Genie 4 arrives in Project Genie |
| 31:44 | Google extends Genie 3 sessions |