# Why the AI Kids Hate Astra 6.2
*The AI Daily Brief — Wednesday, 2026-11-18 · https://aidailybrief.ai/e/2026-11-18*

**The Astra 6.2 backlash may say more about OpenAI's direction than about its models.**

GPT-6.2 Astra and Sol landed to near-universal early-tester hate: unusable, insufferable, a step back. But the same release cut cost per task by up to 35%, opened a big price gap with Anthropic, and was pitched around agents and "the real world" rather than coding benchmarks. The models regressed on exactly what early adopters care about most, coding, while doubling down on efficiency and everyday professional work. That may be a deliberate attempt to carve out a different place in the enterprise market. Perception is in the eye of the beholder, so test it yourself.

---

## By the numbers
- **35%** — Cost-per-task drop for GPT-6.2 Astra ($4 vs. $6.24) on Artificial Analysis
- **2×+** — How much more Fable 5.2 costs than GPT-6.2 Astra on max settings
- **62** — Astra 6.2's Intelligence Index score, up 5 points and second only to Fable 5.2
- **−3 pts** — Sol 6.2's drop on the AA coding agent index, slipping behind Kimi 3.1
- **0 of 30** — Every team members who switched back to Astra 6.2 after trying it
- **103%** — Nvidia's annualized revenue growth in its first post-Jensen quarter
- **$130B** — Nvidia's raised Q4 revenue projection
- **70→75%** — Nvidia's raised growth forecast out to 2028
- **+2%** — Nvidia stock move in the five minutes after Jensen's post

## Headlines

### The outage wave reaches the NSA `[01:00]`
After Microsoft Outlook, the FAA and several big US banks, the NSA's classified internal network went down Tuesday, and the campus went to essential personnel only. Apple and United Airlines also had outages the same day. The NSA blamed "scheduled routine maintenance", and once again sources said there was no evidence of a cyberattack.
*For: Eng, Ops*
Link: https://aidailybrief.ai/e/2026-11-18#nsa-joins-outage-wave

### What we're seeing, and frankly I'd expect to see more of it, are service disruptions coming from updating and modernizing legacy systems. `[02:00]`
*— Sean Cairncross, ONCD*
Asked by reporter Ashley Gold whether this could be a cyberattack, the ONCD chief said the industry has known about this risk for a while and "it's just that it's manifesting." Martin Casado drew the same parallel to Y2K, when most disruption came from the upgrades themselves.
*For: Eng, Ops*
Link: https://aidailybrief.ai/e/2026-11-18#cairncross-modernizing-legacy

### The outages may be downstream of AI, just not the scary way `[02:00]`
Mackenzie Arnold's theory: real fears about automated cyberattacks are driving a surge in major system updates, and more updates mean more failures. AI is also making many of these upgrades affordable and feasible for the first time, so the timing may not be a coincidence even if no attack is involved.
*For: Eng, Ops*
Link: https://aidailybrief.ai/e/2026-11-18#surge-in-updates-surge-in-failures

### We've been on a ticking clock since Mythos/Fable-level models `[03:00]`
Companies are quietly racing to harden systems that have been held together with popsicle sticks and bubblegum for 30 years. Everyone is doing it at once, so the specialists who can fix the breakage are stretched thin. On that view the outages look predictable rather than mysterious, though a real wave of cyberattacks could still come later.
*For: Eng, Ops, Exec*
Link: https://aidailybrief.ai/e/2026-11-18#popsicle-sticks-and-bubblegum

### If the board decides that my return is in the best interest of… the company, and America, I would be honored to answer these calls and return as CEO. `[04:00]`
*— Jensen Huang, in an X post*
In his first statement since resigning last Thursday, Huang said he "made serious mistakes" in his personal life, that he owns them completely, and that he will cooperate with the board's independent investigation. He also cited an outpouring of calls from employees, investors, partners and customers for him to come back.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-11-18#jensen-would-return

### The market votes to bring Jensen back `[05:00]`
Nvidia stock rose 2% in the five minutes after Huang posted. Paul Graham said he didn't envy board members who would have to argue "with a straight face" that any other candidate would do a better job.
*For: Finance*
Link: https://aidailybrief.ai/e/2026-11-18#market-votes-for-jensen

### Nvidia's board: the gears are not turning `[05:00]`
New board chair Harvey Jones opened the earnings call by saying the board "unanimously accepted Jensen's resignation and does not intend to reinstate Jensen as CEO." Interim CEO Colette Kress had already said the same on CNBC: the board has decided, and the focus is on growing the business.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-11-18#nvidia-board-says-no

### But a board faction reportedly wants him back `[06:00]`
Semafor reports that some board members are pushing for Huang's return, and that he spoke with at least one director before his Tuesday post. Three members are said to be vehemently opposed, and it isn't clear where the other seven stand. The story doesn't feel over.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-11-18#semafor-board-faction

### Nvidia's first post-Jensen quarter is another blowout `[06:00]`
Nvidia beat revenue estimates with 103% annualized growth, raised its Q4 projection to $130 billion, and lifted growth forecasts out to 2028 from 70% to 75%. Kress closed with "We've got this," and the stock rose 4.5% after hours to a post-Jensen high. Wedbush's Matt Bryson wasn't sure it changes skeptics' minds.
*For: Finance*
Link: https://aidailybrief.ai/e/2026-11-18#nvidia-strong-quarter

### Nvidia stops shipping even government-approved chips to China `[07:00]`
Kress said guidance includes no licensed Hopper data center products and that Nvidia is no longer shipping them to China. The report doesn't yet reflect revenue lost from shutting down smuggling routes. Jones acknowledged that "national security dynamics are making major changes in our marketplace."
*For: Exec, Legal*
Link: https://aidailybrief.ai/e/2026-11-18#nvidia-exits-china

## Main episode

### GPT-6.2 Astra and Sol get some of the worst first impressions ever `[11:00]`
The week's question was whether OpenAI would drop a spoiler model ahead of Anthropic's IPO. It shipped two, and early reactions were brutal: Matt Shumer said he wouldn't even write a full review, calling them "essentially unusable." This isn't GPT-5-era anger over deprecations. It's about the models themselves.
Link: https://aidailybrief.ai/e/2026-11-18#most-hated-model-release

### OpenAI's pitch leads with agents and cost, not coding `[12:00]`
The announcement promised "super efficient and built to perform where it counts, the real world." It led with Agents' Last Exam and a science variant of Terminal Bench rather than the coding benchmarks that have topped nearly every release for a year. Even the coding section's main selling point was cost.
*For: Product*
Link: https://aidailybrief.ai/e/2026-11-18#built-for-the-real-world

### Small intelligence gains, a strong hold on the middle of the market `[13:00]`
On the Artificial Analysis Intelligence Index, Sol rose from 55 to 57 and Astra gained five points to 62, second only to Fable 5.2. AA says OpenAI now controls most of the Pareto frontier at mid-range prices. Anthropic offers a bit more intelligence at the premium end, and Meta wins on the cheapest models.
Link: https://aidailybrief.ai/e/2026-11-18#artificial-analysis-index

### OpenAI opens a big cost gap with Anthropic `[13:00]`
Astra 6.2 finished AA's benchmark at $4 per task versus $6.24 for Astra 6, a 35% cut. Sol fell to $1.01 from $1.27, about 20%. Fable 5.2 on high settings still costs more than twice what Astra costs on max, so models in the same size class now sit at very different price points.
*For: Finance, Eng*
Link: https://aidailybrief.ai/e/2026-11-18#cost-per-task-gap

### The red flag that set the narrative: a coding regression `[14:00]`
Both models regressed on AA's coding agent index. Astra held its score, while Sol dropped three points to fall behind Kimi 3.1 and 6.1 Terra, driven by SWE Atlas results where it tends to leave out key facts. The open question is whether testers thought the model was bad, or just worse at the one thing they cared about.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-11-18#coding-regression-red-flag

### Every's 30-person team: nobody switched `[15:00]`
Dan Shipper's team tested the models on real coding, writing, design and knowledge work, and no one adopted them. They found writing so concise it was hard to follow, with short answers where long ones were needed. Coding runs often stopped too early on end-to-end builds, and designs fell into a noticeable house style.
*For: Eng, Marketing, Product*
Link: https://aidailybrief.ai/e/2026-11-18#every-team-verdict

### "Clearly super smart," but insufferable to work with `[16:00]`
Theo's long critique says the model "speaks in tongues" and runs in loops, and seems to live in its own world. Ethan Mollick called it uneven and weird, the first frontier OpenAI release since GPT-4.5 that wasn't a clear step forward.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-11-18#smart-but-insufferable

### The other camp: strong on agents, computer use and efficiency `[17:00]`
Some users reported a very different model. Ash praised it for computer use and agentic video editing, Claire Vo for image generation and animation code, and Josh Purtell called both models "absolute monsters" on an auto-research benchmark. Alex Finn said it finished his benchmark in a third of the time, at a third of the price, using a quarter of the tokens.
*For: Eng, Product*
Link: https://aidailybrief.ai/e/2026-11-18#the-defenders

### If quality is all you care about, it flopped. Price-to-output is another story. `[18:00]`
A recurring defense: prices can't keep rising just to do the job a bit better, so efficiency should matter more than raw quality. Others argue efficiency has been OpenAI's real edge for a while, and that this release doubles down on it deliberately.
*For: Finance, Eng*
Link: https://aidailybrief.ai/e/2026-11-18#efficiency-over-quality

### An old split in interaction styles, now more visible `[19:00]`
Anthropic's Fable-class models are the kind you hand a whole assignment and walk away from. OpenAI's are still tuned more for close back-and-forth collaboration. 6.2 may simply extend that pattern, and look worse by comparison because Opus 5.5 is so good at both styles.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-11-18#collaborator-vs-delegate

### Changed days before launch, after the reviews were written `[19:00]`
Some testers blamed the harness more than the model. OpenAI's Tibo said the team made many changes in the final three to four days, after early testers had already recorded their reviews. Critics called that self-sabotage. Theo retested and found it better, but "still directionally very bad."
*For: Product, Marketing*
Link: https://aidailybrief.ai/e/2026-11-18#early-access-self-sabotage

### Not a slam dunk, not era-defining, so test it yourself `[20:00]`
6.2 reads as a very incremental release with some specific tuning choices. Either they don't work in general, or they made the model less useful for some people's use cases. NLW is switching his own OpenAI-based workflows over to find out.
Link: https://aidailybrief.ai/e/2026-11-18#test-it-yourself

### This wasn't an attempt to spoil Anthropic's IPO `[21:00]`
There's no way OpenAI expected this model to steal Anthropic's thunder, especially after Fable 5.2. The more likely motive was defensive: with everyone talking about models because of the IPO, having a recent release beats looking scared after a long silence.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-11-18#not-ipo-sabotage

### The launch ad: "We help your incompetent people fake it" `[22:00]`
OpenAI's two-minute film shows white-collar workers using Astra to finish forgotten tasks for their bosses in under a minute. Critics called it a mockery of humanity and a profoundly sad vision of the future. Swyx said the biggest failure of the launch was that nobody with decision-making power vetoed the ad.
*For: Marketing*
Link: https://aidailybrief.ai/e/2026-11-18#launch-ad-backlash

### Maybe OpenAI is carving out its own lane in enterprise AI `[23:00]`
The ad may just have been Office Space-style fun, complete with 1990s computers and a Lumbergh boss. Or it may be strategy: instead of fighting Anthropic head-to-head on coding, OpenAI offers faster busywork for employees and cost efficiency for their bosses, who are increasingly focused on token caps.
*For: Marketing, Exec*
Link: https://aidailybrief.ai/e/2026-11-18#something-for-you-and-your-boss

*Today's sponsors: KPMG, Granola, Gartner, Blitzy — offers at https://aidailybrief.ai/sponsors*

---
Transcript: https://aidailybrief.ai/e/2026-11-18/transcript.md
Listen: https://pod.link/1680633614 · Ad-free: https://patreon.com/aidailybrief
© 2026 The AI Daily Brief — Until next time, peace ✌