| View in browser: The Diff No. 048 |
| The Diff No. 048 Friday, 20 November 2026 · AI dev news, only what is verified |
 |
Anthropic’s first day on Nasdaq: opened at $111.50, closed at $97.17, seventeen percent above the record $83 price. And three days into GPT-6.2 Astra and Sol, the price cut is real but the reviews are split. One story is about what public markets will pay for the compute plan: an opening print 30% above the biggest IPO price ever set. The other is about what you will pay per token: OpenAI cut its frontier API prices 20% the same week independent benchmarks started disagreeing with each other. Both land on the same desk - yours, next sprint, in the model line item and the capacity roadmap. The market priced Anthropic’s future in a day. The builders are still pricing GPT-6.2 - one task at a time. |
IN THIS ISSUE  | 01 📈 Anthropic’s first day: opened at $111.50, closed at $97.17 - 17% above the biggest IPO price on record. AP confirms the close and the 30% opening pop over the $83 price. Reuters clocked a $2 trillion market value during the debut. Selling stockholders - founders among them - cashed out nearly $2 billion, per TechCrunch’s read of the share counts. |  | 02 🤖 Three days with GPT-6.2 Astra and Sol: the 20% price cut is real - and the reviews are already split. OpenAI’s own launch post prices Astra at $8 in / $40 out, a fifth below GPT-6 Astra, and holds Sol at $2 / $10. Artificial Analysis ranks Astra #3 overall; SimpleBench’s leaderboard shows it regressing below the human baseline. |
|
🔥 VIRAL The loud ones on Hacker News. Numbers are a morning snapshot. |  01 AP market close report, Nov 19 · Reuters intraday valuation · TechCrunch cash-out arithmetic · priced per Anthropic’s own Nov 18 announcement 📈 Anthropic’s first day: opened at $111.50, closed at $97.17 - 17% above the biggest IPO price on record.AP confirms the close and the 30% opening pop over the $83 price. Reuters clocked a $2 trillion market value during the debut. Selling stockholders - founders among them - cashed out nearly $2 billion, per TechCrunch’s read of the share counts. SCENE At the opening cross Thursday, ANT printed $111.50 - more than 30% above the $83.00 price set Wednesday night - and spent the day fading to a $97.17 close, still up 17%, on a day when the rest of Wall Street fell sharply. Reuters reported Anthropic crossing a $2 trillion market value during the session. No first-day collapse, no moonshot held to the close: a strong, ordinary-looking pop on an extraordinary size. THE THING The verified numbers: the $83.00 price, 980,000,000 shares and the ANT ticker are from Anthropic’s official pricing announcement. AP reports the open at $111.50 and the close at $97.17, up more than 17% on the day. The $2 trillion figure is Reuters’ intraday arithmetic, not a company statement. TechCrunch totals the insider side: the 23,648,646 selling-stockholder shares from the pricing notice work out to nearly $2 billion at $83 - money that went to sellers, including founders, not to Anthropic’s compute fund. The company’s own primary raise stands at roughly $79 billion of fresh capital, with a 147-million-share greenshoe still outstanding for 30 days. ✂️ STEAL IT Steal the tape-reading, not the ticker. A 30% open fading to +17% is what price discovery looks like on a name everyone had an opinion about - the open is positioning, the close is closer to a verdict. Three things now matter for anyone whose stack runs on Claude: whether underwriters exercise the greenshoe (more supply, more confidence), where ANT settles over the next two weeks, and Anthropic’s first quarterly report as a public company - the moment its compute spend becomes a line item analysts grade. Public-market patience is now part of your vendor’s funding model. Watch VERDICT Nothing to migrate. The useful habit from yesterday holds: separate the facts (open $111.50, close $97.17, per AP) from the framing ($2 trillion is Reuters’ intraday math) and from anyone’s AI-bubble thesis, which the first day neither proved nor settled. This fails you if you read one day’s close as a judgment on Claude’s roadmap - it prices demand for the stock, not the model. AP: the $97.17 close and $111.50 open · Reuters: the $2 trillion intraday mark (via Investing.com) · TechCrunch: the founders’ nearly $2 billion cash-out · Anthropic: the official pricing announcement Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew it |
 02 OpenAI’s official launch post, Nov 17 · Artificial Analysis independent evals · SimpleBench public leaderboard · early-tester reactions 🤖 Three days with GPT-6.2 Astra and Sol: the 20% price cut is real - and the reviews are already split.OpenAI’s own launch post prices Astra at $8 in / $40 out, a fifth below GPT-6 Astra, and holds Sol at $2 / $10. Artificial Analysis ranks Astra #3 overall; SimpleBench’s leaderboard shows it regressing below the human baseline. SCENE OpenAI launched two models Tuesday: GPT-6.2 Astra, its frontier model for “your hardest projects,” and GPT-6.2 Sol, “our go-to model for everyday work” - their words. Within 48 hours the discourse split into two stories that barely overlap: a genuine price cut, and a genuine argument about whether the models are actually better. THE THING The verified facts, from OpenAI’s launch post: Astra’s API prices are $8 per million input tokens ($0.40 cached) and $40 per million output - 20% below GPT-6 Astra - and Sol keeps GPT-6.1 Sol’s prices, which third-party trackers list at $2 in and $10 out; context is about 1.05 million tokens, with input rates doubling past 272,000. OpenAI’s own claims, its wording: a new high on Agents’ Last Exam, “5.5 points ahead of Claude Fable 5.2 at just 5% of the cost,” and “the best model in the world for scientists.” The independent picture is mixed: Artificial Analysis scores Astra 62 on its Intelligence Index, #3 of 95, and Sol 57, #13, at about $1.01 per task. The ARC Prize account posted that Astra scored 100% on ARC-AGI-3. And SimpleBench’s leaderboard shows Astra at 77.3% - below its 83.7% human baseline and 6.3 points under GPT-6 Astra - a regression that drove a large Reddit thread. Theo of t3.gg posted a piece titled “I hate GPT-6.2 Astra.” One tester, one opinion - but a widely shared one. ✂️ STEAL IT Steal the arbitrage window. A 20% frontier price cut lands in your cost model today whether or not the model is better, and Sol at $2/$10 undercuts Gemini 4 Neon’s $3/$15 on sticker price. But the eval signals genuinely conflict - top-three on one index, below-human-baseline on another - which is exactly the situation where your own eval set beats everyone’s leaderboard. Before you migrate anything, run your five hardest real prompts on both GPT-6.2 models and your current default. The launch-week discount is only a bargain if your tasks are in the winning column. Try today VERDICT The price cut is a fact you can bank; the quality is a question only your workload can answer. Treat OpenAI’s benchmark claims as its own marketing and the SimpleBench result as one hard test among several. This fails you if you switch production traffic on either the hype or the backlash - both are 72 hours old. OpenAI: the official GPT-6.2 Astra and Sol announcement · Artificial Analysis: Astra intelligence, speed and price · Artificial Analysis: Sol intelligence, speed and price · SimpleBench: the public leaderboard Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperTest it for me |
📊 THE ATTENTION GAP HN points at about 06:30 IST, Nov 20. Anthropic’s stock-market debut outdrew its black-hole physics result more than two to one. The market story is louder; the research result is the one that changes what these systems can do. | Anthropic’s Nasdaq debut thread | | 997 HN | | Claude black-hole physics thread | | 431 HN |
|
ALSO ON THE RADAR Claude Fable 5.2 solved two open problems in black-hole physics. Anthropic’s research post: a side project by researcher Hongbin Chen grew into a background swarm of Claude agents that ran for about six weeks and completed two calculations physicists had left unfinished - one on the Hawking-Page phase transition set up in 2003, one on the modular bootstrap ceiling. Results were formally verified in Lean and checked by a physicist; preprints are linked from the post. Anthropic’s own caveat: neither result describes our universe directly, and neither has practical applications. Anthropic: the research post → California’s attorney general subpoenaed Nvidia. From the official press release: AG Rob Bonta served an investigative subpoena asking whether Nvidia’s advanced chips or confidential information reached China and whether the company misled investors and the public, in the wake of the Jensen Huang affair reporting. Bonta’s statement: companies cannot mislead their investors. AP and Politico match the release. California DOJ: the press release → Genie 4 is live in Project Genie for AI Ultra subscribers. Google’s rollout post: eligible Ultra subscribers get Genie 4 worlds with mid-session prompting - change weather, drop in characters, redirect action while exploring - plus real-time native audio, a Project Genie agent that helps build worlds, and a community gallery. Browser-based; no API yet mentioned in the post. Google: the Project Genie update → |
🧪 ONE THING TO TRY TONIGHT Fifteen minutes: rerun your model cost sheet, then run one hard prompt on the new pair. Open your current model cost assumptions and swap in the new sticker prices: GPT-6.2 Astra at $8 in / $40 out, GPT-6.2 Sol at $2 / $10, Gemini 4 Neon at $3 / $15. Then take one genuinely hard prompt from your own work - not a demo prompt - and run it on Sol and Neon side by side. Note correctness, not vibes. The price war is only useful to you at the task level. |
What this issue could not see. Anthropic’s open, close and percentage moves are AP’s market reporting; the $2 trillion figure is Reuters’ intraday arithmetic, not a company statement; the founders’ cash-out is TechCrunch’s calculation from share counts in Anthropic’s own pricing announcement. GPT-6.2 pricing and benchmark claims are from OpenAI’s official launch post - its wording, its assessments; Sol’s $2/$10 figures are corroborated by independent trackers. Artificial Analysis and SimpleBench numbers are from their public pages at about 06:30 IST; the ARC-AGI-3 perfect score is the ARC Prize account’s post, and Theo’s piece is one tester’s opinion. HN points are a snapshot. Nvidia subpoena facts are from the California DOJ press release. No video this issue: no launch or trading-day video from a verified official channel was found. Reddit and X were not consulted. |
|
| Images: credited in each report; source screenshots and announcement artwork, not our tests. |
|