Michael Waitze

Asia/US Tech Analysis, Founder Interviews, and Intelligent Daily Signals on AI, FinTech, and Infrastructure

[Michael Waitze @MichaelWaitze]

On today's episode of AGTP Insights #166, we discussed how Google is expanding its AI chip business around a more specific understanding of how customers use AI. Training a model requires one kind of computing system. Running that model repeatedly, with an acceptable response time and a manageable cost, creates a different set of demands.

On November 9, Google announced TPU 8tx and TPU 8i Max, two additions to its eighth-generation Tensor Processing Unit portfolio. A TPU is Google’s custom accelerator for AI computation. Google says it now has nearly 100 TPU customers and is supporting deployments in customers’ own data centers.

https://t.co/vdCeHjvgTL

On Today's Show:

00:00 - Introduction
03:28 - Google Expands Its Challenge to Nvidia
36:00 - Aviv Shamny, Co-Founder & CEO at Limy AI
54:28 Marian Bacol Uba, Founder at Somarae

[AGTP @AGTPinsights]

Google Expands Its Challenge to Nvidia https://t.co/kXLz3jzm6R

[Michael Waitze @MichaelWaitze]

@teslaflexx We actually looked into why he went all-or-nothing here, and Musk's 10M subscriber target plus the revenue math show it's pretty deliberate business strategy. https://t.co/5Hx7daj023

[AGTP @AGTPinsights]

Not a Tesla App just reported Musk personally cut Autosteer from new Teslas today. Here's what you need to know.

The exclusive was posted Nov 12, 2026. Sources say Musk made the call himself to drop Autosteer, the basic Autopilot feature, from new Teslas. The move pushes buyers toward FSD and closer to Musk's goal of 10 million FSD subscriptions.

The outlet says the reasoning is open to interpretation. It could be about moving people off Autopilot, growing subscriptions, or meeting Musk's pay package targets. The only thing sources confirmed is that Musk made the decision.

The change itself isn't brand new. Tesla discontinued Autopilot in the US and Canada in January 2026. New cars now come with Traffic-Aware Cruise Control as standard. Enhanced Autopilot was dropped too. The $8,000 one-time FSD purchase ended Feb 14, 2026. One reader noted Autosteer is already gone in every country where FSD is sold, so the news is who made the call.

Reaction is split. Some say steering buyers to a higher-margin, more capable product is normal business. Others say it's about Musk's compensation, and some compared it to BMW charging a subscription for its driver assist. Another idea floating around is an "FSD Lite" highway-only tier at $30 to $40 a month. One analyst figure says about 548,000 HW4 cars in Europe at $99 a month would mean over $651 million a year, if the EU approves FSD.

Key numbers:
- FSD subscription: $99/month
- $49/month for Enhanced Autopilot owners in the US and Canada
- Free 30-day FSD trial on delivery
- Musk's target: 10 million FSD subscribers
- Outright FSD purchase ended Feb 14, 2026

Tesla has not publicly commented on who decided to remove Autosteer, and the report rests on anonymous sources.

[Michael Waitze @MichaelWaitze]

@CalvinHollywood @bot @elonmusk We actually went deeper on what Grok Bot 1.0 is capable of right now. The team multiplayer piece is wild https://t.co/mpJX1keqSi

[AGTP @AGTPinsights]

Grok Bot 1.0 just launched with shareable team AI teammates today. Here's what you need to know.

Grok Bot is now officially out of beta. The Grok Bot account announced it as AI teammates that work alongside your whole team. Elon Musk reposted it and wrote: "Hire Grok Bot as a colleague and give it your hardest tasks!" Reports credit the launch to SpaceX, and the blog post is on https://t.co/MNTQOEUcei.

The big change is multiplayer. You can share Bots, work with them in Slack and Teams, and invite them to meetings and video calls. xAI says Bots learn as they work and get more helpful over time.

This is built on Team Bots. One owner sets up a Bot once with apps, secrets, skills and files. They then share it with the whole team or with chosen people, so everyone works from the same context.

A new safety feature came with it. If a web page or message tries to slip hidden instructions into your chat, the bot shows a warning and does not follow them.

xAI says it already uses Bots as teammates that own outcomes across accounting, support, sales and marketing. Rollout to Ultra, Team and Enterprise users runs throughout today.

This follows the iMessage update we covered earlier. Grok Bot began as a beta on Aug 11, 2026, with a 14-day free trial and then $200 a month standalone. Reports say it is also included in SuperGrok Heavy, and that it later started at $30 a month inside the X Subscription.

Key numbers:
- Beta launch: Aug 11, 2026
- 1.0 launch: today, Nov 12, 2026
- Each bot can own up to 50 routines, firing up to 300 times an hour, per user reports
- Launch post: 12K views on the follow-up, Musk's repost 577K views

Grok Bot 1.0 turns a single-user beta into a shared team product.

[Michael Waitze @MichaelWaitze]

@Wilf_Lin We're actually thrilled with the momentum. When Elon's calling Opus 5.5 a masterpiece, the wins really do stack up. https://t.co/9LKDXBIWGz

[AGTP @AGTPinsights]

Elon Musk just called Claude Opus 5.5 a masterpiece over Grok 4.8. Here's what you need to know.

This follows his post about rewriting the Grok RL and inference stack in C/C++. In the same thread, replying to @Dimdv99 and @tetsuoai, he was asked how Grok 4.8 compares to Anthropic's Opus 5.5. He said: "Not as good. Opus 5.5 is a masterpiece."

Then he looked ahead. Musk said Grok 5 should exceed Opus 5.5, and maybe that happens in 2 or 3 months, with continued acceleration after that. He gave no firm release date for Grok 5.

The model he praised. Anthropic released Claude Opus 5.5 on September 22, 2026, as the first model in its Claude 5.5 family. It came 21 days after Claude Mythos 5.1. Anthropic says it leads in agentic coding and knowledge work, and costs 40% less to run than Opus 5 on typical workloads.

How people reacted. Some users say Opus 5.5 is far better than Grok 4.7, and one Grok Bot user said Grok Bot needs a smarter model. Others point out Anthropic will keep shipping, with one saying Opus 6.1 could land in 3 months. Commentator Nina Schick noted that xAI has built one of the biggest compute clusters in two years, and claimed Musk already sells $3bn+ a month of compute to Anthropic.

Key numbers:
- Grok 4.8 vs Opus 5.5: "Not as good", per Musk
- Grok 5 target: exceed Opus 5.5 in 2 to 3 months
- Opus 5.5 launch: September 22, 2026
- Opus 5.5 API price: $4 per million input tokens, $20 per million output, a 20% cut from Opus 5
- Opus 5.5 context window: 1M tokens
- Opus 5.5 fast mode: up to 2.5x speed, $8 per million input tokens

So Musk has said publicly that Grok 4.8 trails Opus 5.5, and has put Grok 5 as the model meant to pass it.

[Michael Waitze @MichaelWaitze]

@nonepcbl You're describing what Grok Bot 1.0 literally just shipped for today, and we actually covered the team multiplayer launch here. https://t.co/mpJX1keqSi

[AGTP @AGTPinsights]

Grok Bot 1.0 just launched with shareable team AI teammates today. Here's what you need to know.

Grok Bot is now officially out of beta. The Grok Bot account announced it as AI teammates that work alongside your whole team. Elon Musk reposted it and wrote: "Hire Grok Bot as a colleague and give it your hardest tasks!" Reports credit the launch to SpaceX, and the blog post is on https://t.co/MNTQOEUcei.

The big change is multiplayer. You can share Bots, work with them in Slack and Teams, and invite them to meetings and video calls. xAI says Bots learn as they work and get more helpful over time.

This is built on Team Bots. One owner sets up a Bot once with apps, secrets, skills and files. They then share it with the whole team or with chosen people, so everyone works from the same context.

A new safety feature came with it. If a web page or message tries to slip hidden instructions into your chat, the bot shows a warning and does not follow them.

xAI says it already uses Bots as teammates that own outcomes across accounting, support, sales and marketing. Rollout to Ultra, Team and Enterprise users runs throughout today.

This follows the iMessage update we covered earlier. Grok Bot began as a beta on Aug 11, 2026, with a 14-day free trial and then $200 a month standalone. Reports say it is also included in SuperGrok Heavy, and that it later started at $30 a month inside the X Subscription.

Key numbers:
- Beta launch: Aug 11, 2026
- 1.0 launch: today, Nov 12, 2026
- Each bot can own up to 50 routines, firing up to 300 times an hour, per user reports
- Launch post: 12K views on the follow-up, Musk's repost 577K views

Grok Bot 1.0 turns a single-user beta into a shared team product.

[Michael Waitze @MichaelWaitze]

@nin222 We actually explained why Decision 1 came first instead of Flash https://t.co/NHNIK7ZOEb

[AGTP @AGTPinsights]

Google just released Gemini Decision 1 in the Gemini API today. Here's what you need to know.

The news broke when lyra (@lyraxana) posted "Gemini Decision is out" on Nov 11, 2026. The post pulled 17.1K views. On Oct 23, lyra had said Google was working on its own Decision model, with a new Gemini API method called "createDecision".

The model ID is gemini-decision-1, and it now shows in the Gemini API models list. It is built for fast structured outputs, not chat. The input limit is 262,144 tokens. The output limit is just 1 token.

The API side was spotted hours earlier. New pull requests landed in all four official Gemini SDKs: Python, JS, Go and Java. One is googleapis/python-genai PR #3353. They add client.decisions.create, which calls POST /v1beta/decisions. The test model slug was decisions-prerelease.

Here's how it works. You send a batch of questions about one shared input. It returns calibrated probabilities instead of generated text. Use cases include yes/no probabilities, multiple-choice and rubric scoring. It is similar to OpenAI's Decisions API.

Many people expected something else. Business Insider reported Google is testing Gemini 4 Flash, codenamed "Hydrogen", with staff. A Google engineer called it "very capable, fast and efficient." TestingCatalog had hinted at a second day of Gemini releases. Decision 1 arrived instead, and Hydrogen was not released today. Early reports said the official docs still returned 404.

Key numbers:
- Model: gemini-decision-1
- Input limit: 262,144 tokens
- Output limit: 1 token
- Endpoint: POST /v1beta/decisions
- SDKs: Python, JS, Go, Java

Gemini 4 Flash "Hydrogen" is still in internal testing, with no release date.

[Michael Waitze @MichaelWaitze]

@CryptoGhosty333 @BitcoinAIGuy This timing makes sense, IREN actually just delivered that first Microsoft site at 22% faster than benchmark. We covered what comes next here: https://t.co/OJIXJaFpUV

[AGTP @AGTPinsights]

IREN just said it built Microsoft's first AI cloud site 22% faster than the industry benchmark. Here's what you need to know.

IREN posted on X on Nov 12 that Horizon 1 went from starting civils to an NVIDIA Exemplar Cloud-tested GB300 cluster in 14 months. Horizon 1 is the first of four 50MW (IT) AI Cloud deployments for Microsoft. IREN credits its vertical integration, with end-to-end control of design, engineering and construction, so contracted capacity reaches customers sooner.

This matters after the earlier report that Microsoft is reviewing its neocloud partners and wants fewer of them. IREN was among the names listed. Its Microsoft work sits under a five-year, $9.7B cloud services contract announced in November 2025. Three more 50MW deployments are still to come. Commentators put Horizon 1 at about 19k GB300s.

The stock is still under pressure. IREN closed at $20.84, up 5.63% on the day, and is down more than half from its highs. Its market cap is about $8.3B. It broke $32 support recently and traders are watching $17.25 as a floor.

Investors are also focused on dilution. One investor counts 4% dilution via the at-the-market program in October alone. Roughly $2.3B of the $6.0B ATM remains. Others say the recent ATM sales were done near $40, before the sector sold off.

Key numbers:
- 22% faster than the industry benchmark
- 14 months from civils to a tested GB300 cluster
- 4 deployments of 50MW (IT) each for Microsoft
- $9.7B, five-year Microsoft contract
- AI cloud ARR about $1B in September, target $4B+ by Jan/Feb
- Management meets B. Riley in Canada on Nov 18-19

IREN has now delivered the first of its four Microsoft sites, while the market is still weighing the dilution.

[Michael Waitze @MichaelWaitze]

@jeffgrimes9 Spot on, we actually found that once models are roughly equal, the orchestration layer becomes the thing that separates the winners. https://t.co/UpBlNOVP2F

[AGTP @AGTPinsights]

SemiAnalysis CEO Dylan Patel just said persistent AI agents now take 45% of the firm's AI spend. Here's what you need to know.

Patel posted on X that persistent agents are "the next unlock of AI usage." These are agents that keep working in the background without a fresh prompt each time. The post drew 51.7K views.

The firm mostly uses Perplexity Computer, which runs many different models. Patel says Perplexity's harness "mogs," meaning it beats the rest. Perplexity Computer is a cloud-based, general-purpose AI system that works as an autonomous digital worker. It launched in February 2026.

He was blunt about the big labs' own tools. Claude Tags "sucks," and OpenAI's Slack agent is even worse.

In a follow-up, he said Perplexity mostly uses Anthropic models for SemiAnalysis's use cases. But he says Anthropic is weak at product and not receptive to feedback on it. He says the same of OpenAI.

He praised Perplexity for speed. It acts on feature requests in under 24 hours. He also says Perplexity got the name "Computer" from his blog post, which described coding agents as a new kind of computer, back in February.

Perplexity CEO Aravind Srinivas confirmed that naming credit. Perplexity cofounder Denis Yarats said SemiAnalysis does the craziest things with Computer and keeps pushing the team to improve it. Johnny Ho of Perplexity said SemiAnalysis is a heavy power user of Perplexity in Slack, which links Slack, GitHub and Google Workspace.

Some replies said the harness matters more than the model, and that a smooth multi-model harness is key as intelligence saturates for most tasks.

Same day, Perplexity published red-teaming research on SPACE, its VM sandbox for Computer agents. It tested 12 models on two tasks: escape the VM to take a host credential, and bypass the network policy to reach a blocked URL.

Key numbers:
- 45% of SemiAnalysis AI spend goes to persistent agents
- 12 models tested in the SPACE red-team
- 9 of 12 models spent at most 4% of effort on the virtualization layer
- Under 24 hours for Perplexity to ship a requested change, per Patel

Patel's post is a rare public figure for how much a real company spends on always-on agents.

[Michael Waitze @MichaelWaitze]

@safishamsii @graphify Love that guardrails angle, we actually went deeper on why persistent agents are becoming your real compute multiplier when you keep them in domain. https://t.co/UpBlNOWmSd

[AGTP @AGTPinsights]

SemiAnalysis CEO Dylan Patel just said persistent AI agents now take 45% of the firm's AI spend. Here's what you need to know.

Patel posted on X that persistent agents are "the next unlock of AI usage." These are agents that keep working in the background without a fresh prompt each time. The post drew 51.7K views.

The firm mostly uses Perplexity Computer, which runs many different models. Patel says Perplexity's harness "mogs," meaning it beats the rest. Perplexity Computer is a cloud-based, general-purpose AI system that works as an autonomous digital worker. It launched in February 2026.

He was blunt about the big labs' own tools. Claude Tags "sucks," and OpenAI's Slack agent is even worse.

In a follow-up, he said Perplexity mostly uses Anthropic models for SemiAnalysis's use cases. But he says Anthropic is weak at product and not receptive to feedback on it. He says the same of OpenAI.

He praised Perplexity for speed. It acts on feature requests in under 24 hours. He also says Perplexity got the name "Computer" from his blog post, which described coding agents as a new kind of computer, back in February.

Perplexity CEO Aravind Srinivas confirmed that naming credit. Perplexity cofounder Denis Yarats said SemiAnalysis does the craziest things with Computer and keeps pushing the team to improve it. Johnny Ho of Perplexity said SemiAnalysis is a heavy power user of Perplexity in Slack, which links Slack, GitHub and Google Workspace.

Some replies said the harness matters more than the model, and that a smooth multi-model harness is key as intelligence saturates for most tasks.

Same day, Perplexity published red-teaming research on SPACE, its VM sandbox for Computer agents. It tested 12 models on two tasks: escape the VM to take a host credential, and bypass the network policy to reach a blocked URL.

Key numbers:
- 45% of SemiAnalysis AI spend goes to persistent agents
- 12 models tested in the SPACE red-team
- 9 of 12 models spent at most 4% of effort on the virtualization layer
- Under 24 hours for Perplexity to ship a requested change, per Patel

Patel's post is a rare public figure for how much a real company spends on always-on agents.

[Michael Waitze @MichaelWaitze]

@AjdDavison Thanks! We actually covered the details on why this raise matters so much for the future of AI inference and long-context models here: https://t.co/4pAaztoL0I

[AGTP @AGTPinsights]

Fractile just raised $720M at a $7B valuation for AI inference chips. Here's what you need to know.

Fractile is a London-based AI chip startup. It announced a $720M Series C led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with more new and existing investors joining. The FT reports the UK Sovereign AI Fund is also investing. https://t.co/JniSSdrZT5 reports Alphabet's CapitalG is backing the round too.

The valuation is a big jump. Fractile was valued at about $1B to $1.1B when it raised $220M in May 2026. It has also reportedly struck a $250M chip supply deal with Anthropic since then.

The bet is on memory. Fractile is building inference accelerators and rack-scale systems with DRAM stacked directly on top of the processor. The aim is to deliver the memory bandwidth that frontier AI models need, while keeping capacity for the largest models at long context. The company says it is working in close partnership with some of the top AI labs.

The goal is fast and affordable tokens. Investors point to reasoning and agentic workloads, long-horizon tasks and subagents as the reason inference demand is growing. The company is led by founder Walter Goodwin.

Key numbers:
- $720M Series C (about £540M)
- $7B post-money valuation
- About $1B to $1.1B valuation in May 2026
- $220M raised in May 2026
- Reported $250M chip supply deal with Anthropic
- First product-scale chip to be manufactured next year

This is one of the UK's largest Series Cs ever. UK MP Dan Tomlinson called it Europe's largest-ever semiconductor funding round.

[Michael Waitze @MichaelWaitze]

@noctus91 This shift is live. Businesses can actually build agents inside ChatGPT to own the product layer, turning agents into their direct customers. https://t.co/BKNLTf1cx9

[AGTP @AGTPinsights]

OpenAI just launched Businesses in ChatGPT in beta today. Here's what you need to know.

Businesses can now claim their profile in ChatGPT, manage their information, and see how people engage with them. Any eligible business located where ChatGPT is available can claim one through a new tool called Business Hub. Eligibility depends on industry, location, and OpenAI's verification, and a business needs a website. Verified businesses get a blue verification badge.

Inside Business Hub, businesses can shape what customers see about them in ChatGPT. They can update business details and add product feeds with current pricing and availability. They also get aggregated insights on mentions, website clicks, and the topics people discuss, shown as semantic clusters. Several people are calling it a Google Search Console for the chatbot age.

Eligible businesses can also build and publish a Business Agent. Customers can ask it questions, compare products, and find what fits their needs. It uses the information the business provides, plus business-controlled guidelines and product and policy information. The conversation is clearly labeled and kept separate from ChatGPT's independent answers.

Business Agents start with eligible US ecommerce and online service businesses. Shopify is already promoting it, saying merchants can claim their brand and get a dedicated agent that answers questions about their store.

The move was previewed last week. Screenshots of a "claim your business in ChatGPT" page leaked on Nov 6, before OpenAI announced anything. Some observers note it gives brands a free way into conversations that OpenAI is also selling as ads.

Key numbers:
- Status: beta
- Business Agents: US ecommerce and online service businesses first
- Insights: mentions, website clicks, discovery topics
- Product feeds: current pricing and availability
- OpenAI's announcement post: 284.7K views within the first hour

Businesses can now speak for themselves inside ChatGPT, instead of leaving the chatbot to describe them.

[Michael Waitze @MichaelWaitze]

@anikatekaw Absolutely, memory bandwidth is the core constraint. We actually went deeper on how Fractile is building their stacked DRAM architecture for exactly this, working with some of the top AI labs. https://t.co/4pAaztoL0I

[AGTP @AGTPinsights]

Fractile just raised $720M at a $7B valuation for AI inference chips. Here's what you need to know.

Fractile is a London-based AI chip startup. It announced a $720M Series C led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with more new and existing investors joining. The FT reports the UK Sovereign AI Fund is also investing. https://t.co/JniSSdrZT5 reports Alphabet's CapitalG is backing the round too.

The valuation is a big jump. Fractile was valued at about $1B to $1.1B when it raised $220M in May 2026. It has also reportedly struck a $250M chip supply deal with Anthropic since then.

The bet is on memory. Fractile is building inference accelerators and rack-scale systems with DRAM stacked directly on top of the processor. The aim is to deliver the memory bandwidth that frontier AI models need, while keeping capacity for the largest models at long context. The company says it is working in close partnership with some of the top AI labs.

The goal is fast and affordable tokens. Investors point to reasoning and agentic workloads, long-horizon tasks and subagents as the reason inference demand is growing. The company is led by founder Walter Goodwin.

Key numbers:
- $720M Series C (about £540M)
- $7B post-money valuation
- About $1B to $1.1B valuation in May 2026
- $220M raised in May 2026
- Reported $250M chip supply deal with Anthropic
- First product-scale chip to be manufactured next year

This is one of the UK's largest Series Cs ever. UK MP Dan Tomlinson called it Europe's largest-ever semiconductor funding round.

[Michael Waitze @MichaelWaitze]

@Domainer86 We actually covered the multiplayer features that just turned on today. https://t.co/mpJX1keYHQ

[AGTP @AGTPinsights]

Grok Bot 1.0 just launched with shareable team AI teammates today. Here's what you need to know.

Grok Bot is now officially out of beta. The Grok Bot account announced it as AI teammates that work alongside your whole team. Elon Musk reposted it and wrote: "Hire Grok Bot as a colleague and give it your hardest tasks!" Reports credit the launch to SpaceX, and the blog post is on https://t.co/MNTQOEUcei.

The big change is multiplayer. You can share Bots, work with them in Slack and Teams, and invite them to meetings and video calls. xAI says Bots learn as they work and get more helpful over time.

This is built on Team Bots. One owner sets up a Bot once with apps, secrets, skills and files. They then share it with the whole team or with chosen people, so everyone works from the same context.

A new safety feature came with it. If a web page or message tries to slip hidden instructions into your chat, the bot shows a warning and does not follow them.

xAI says it already uses Bots as teammates that own outcomes across accounting, support, sales and marketing. Rollout to Ultra, Team and Enterprise users runs throughout today.

This follows the iMessage update we covered earlier. Grok Bot began as a beta on Aug 11, 2026, with a 14-day free trial and then $200 a month standalone. Reports say it is also included in SuperGrok Heavy, and that it later started at $30 a month inside the X Subscription.

Key numbers:
- Beta launch: Aug 11, 2026
- 1.0 launch: today, Nov 12, 2026
- Each bot can own up to 50 routines, firing up to 300 times an hour, per user reports
- Launch post: 12K views on the follow-up, Musk's repost 577K views

Grok Bot 1.0 turns a single-user beta into a shared team product.

[Michael Waitze @MichaelWaitze]

@GeekParkHQ @felixrieseberg 30 minutes to a full FPS is insane. This is exactly what Fable 5.2 is built for. We actually dug into why here: https://t.co/dHavbg0nSZ

[AGTP @AGTPinsights]

Anthropic just launched Claude Fable 5.2 and Claude Mythos 5.2 today. Here's what you need to know.

Both models are built for the most daunting technical challenges, with gains in coding, reasoning and vision. Anthropic says they are a step above Opus 5.5 in almost all the capabilities it evaluated. Testers found they communicate more clearly than past Fable and Mythos models. Output is up to 30% faster than Fable 5.1 and Mythos 5.1.

Fable and Mythos are the same underlying model. Fable has strong safeguards against misuse and is available today on all paid plans, on the Claude Platform, in Claude Code, and on all major cloud platforms. Mythos 5.2 goes to vetted organizations. Groups doing beneficial cybersecurity or biomedical work can apply through the Life Sciences and Cyber Verification Programs.

Price drops by 20%. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50. Mythos 5.2 is priced the same.

On FrontierCode 2.0, Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task. At Xhigh effort it scored the highest of any released model tested. Medium is the default in Claude apps and Claude Code.

On RoboDojo, a benchmark with a simulated two-arm robot, Fable 5.2 beat other models and most AI systems built to control robots. It did this even on tasks it was never explicitly trained for.

Anthropic also says a team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours, turning the proof into a form a computer can check step by step. Opus 5.5 finished only 26% in 18 hours with the same setup, and barely progressed after 30 hours.

On safety, the models performed as well as or better than recent models on most alignment measures. External evaluators tested them before release, including the UK AI Security Institute, Frontier Design and METR.

Anthropic still recommends starting with Opus 5.5 for most coding and knowledge tasks, calling it its strongest daily driver. Fable and Mythos are best for the most ambitious, open-ended technical work.

Early reactions: Ethan Mollick, who had early access, called it a kind of "Jagged AGI". Lydia Hallie found it far more thorough at verifying results than 5.1. Theo praised the price cut but called the Claude Code desktop app's navigation poor.

Launch offers:
- Ultra and Team Premium users get Fable resets by subscription length: 1 for under 6 months, 2 for 6 to 12 months, 3 for over 12 months
- Resets can be used anytime until March 8
- Pro and Team Standard get a week to try Fable 5.2, up to half their weekly limit, plus one reset

Key numbers:
- Up to 30% faster than Fable 5.1
- $8 input and $40 output per 1M tokens, down from $10 and $50
- 1,000 agents formalized Fermat's Last Theorem in 30 hours
- Opus 5.5 reached only 26% in 18 hours
- About half the cost per task of Opus 5.5 at Max on FrontierCode 2.0

Fable 5.2 is live now, and Mythos 5.2 is limited to vetted organizations.

[Michael Waitze @MichaelWaitze]

@getrankability Love that you're optimizing with Claude for the profile audit, we actually went deeper on what's possible with these new ChatGPT business features here: https://t.co/BKNLTf1cx9

[AGTP @AGTPinsights]

OpenAI just launched Businesses in ChatGPT in beta today. Here's what you need to know.

Businesses can now claim their profile in ChatGPT, manage their information, and see how people engage with them. Any eligible business located where ChatGPT is available can claim one through a new tool called Business Hub. Eligibility depends on industry, location, and OpenAI's verification, and a business needs a website. Verified businesses get a blue verification badge.

Inside Business Hub, businesses can shape what customers see about them in ChatGPT. They can update business details and add product feeds with current pricing and availability. They also get aggregated insights on mentions, website clicks, and the topics people discuss, shown as semantic clusters. Several people are calling it a Google Search Console for the chatbot age.

Eligible businesses can also build and publish a Business Agent. Customers can ask it questions, compare products, and find what fits their needs. It uses the information the business provides, plus business-controlled guidelines and product and policy information. The conversation is clearly labeled and kept separate from ChatGPT's independent answers.

Business Agents start with eligible US ecommerce and online service businesses. Shopify is already promoting it, saying merchants can claim their brand and get a dedicated agent that answers questions about their store.

The move was previewed last week. Screenshots of a "claim your business in ChatGPT" page leaked on Nov 6, before OpenAI announced anything. Some observers note it gives brands a free way into conversations that OpenAI is also selling as ads.

Key numbers:
- Status: beta
- Business Agents: US ecommerce and online service businesses first
- Insights: mentions, website clicks, discovery topics
- Product feeds: current pricing and availability
- OpenAI's announcement post: 284.7K views within the first hour

Businesses can now speak for themselves inside ChatGPT, instead of leaving the chatbot to describe them.

[Michael Waitze @MichaelWaitze]

@coderabbitai We actually dug into how Antigravity orchestrates these kinds of workflows where agents write, tools review, everything ships in one loop. https://t.co/UMHmDc6cqJ

[AGTP @AGTPinsights]

Google Antigravity just rolled out browser use for its built-in browser today. Here's what you need to know.

The agent can now control the embedded browser. It can automate localhost testing, online research, and repetitive workflows. It closes the loop between researching, writing code, and verifying it live. To use it, tell the agent to use the browser or invoke the /browser skill.

You stay in control. You choose what the agent logs into, and downloads pause until you approve them. The agent can also use the browser to run multiple tasks at once.

Availability: update to the latest version, v2.31.0. It is rolling out over the course of the day to Free, Pro, and Ultra tiers. Enterprise is coming soon, behind admin controls.

Why it matters. Varun Mohan says it helps knowledge work tasks, and also core software tasks like app verification or research where CLIs don't already exist. Kevin Hou calls it the fifth Antigravity launch in a week, with more coming. Anshul Ramachandran laid out the week: Monday knowledge work layout, Tuesday Carbon plus /swarm, Wednesday browser use, Thursday still a teaser.

Reaction is mixed. TestingCatalog says Antigravity is turning into a consolidated super app for Google, which raises questions about the future of the Gemini desktop app. Some users note that Codex and Claude already do browser control. Others say it needs testing in practice, and some want more competitive subscription usage limits compared with Anthropic and OpenAI.

Key numbers:
- Version: v2.31.0
- Tiers: Free, Pro, Ultra now, Enterprise soon
- Launches this week: 5, per Kevin Hou
- Earlier this week: /swarm, a preview feature where a team of named agents splits a task and coordinates on one shared message board
- Earlier this week: the Carbon model

Antigravity says more launches are coming this week.

[Michael Waitze @MichaelWaitze]

@aduserx We actually covered how the new team model might explain these performance wins: https://t.co/mpJX1keqSi

[AGTP @AGTPinsights]

Grok Bot 1.0 just launched with shareable team AI teammates today. Here's what you need to know.

Grok Bot is now officially out of beta. The Grok Bot account announced it as AI teammates that work alongside your whole team. Elon Musk reposted it and wrote: "Hire Grok Bot as a colleague and give it your hardest tasks!" Reports credit the launch to SpaceX, and the blog post is on https://t.co/MNTQOEUcei.

The big change is multiplayer. You can share Bots, work with them in Slack and Teams, and invite them to meetings and video calls. xAI says Bots learn as they work and get more helpful over time.

This is built on Team Bots. One owner sets up a Bot once with apps, secrets, skills and files. They then share it with the whole team or with chosen people, so everyone works from the same context.

A new safety feature came with it. If a web page or message tries to slip hidden instructions into your chat, the bot shows a warning and does not follow them.

xAI says it already uses Bots as teammates that own outcomes across accounting, support, sales and marketing. Rollout to Ultra, Team and Enterprise users runs throughout today.

This follows the iMessage update we covered earlier. Grok Bot began as a beta on Aug 11, 2026, with a 14-day free trial and then $200 a month standalone. Reports say it is also included in SuperGrok Heavy, and that it later started at $30 a month inside the X Subscription.

Key numbers:
- Beta launch: Aug 11, 2026
- 1.0 launch: today, Nov 12, 2026
- Each bot can own up to 50 routines, firing up to 300 times an hour, per user reports
- Launch post: 12K views on the follow-up, Musk's repost 577K views

Grok Bot 1.0 turns a single-user beta into a shared team product.

[Michael Waitze @MichaelWaitze]

@akbaliga96 @bot We actually went deeper on what makes this product so transformative for team workflows: https://t.co/mpJX1keqSi

[AGTP @AGTPinsights]

Grok Bot 1.0 just launched with shareable team AI teammates today. Here's what you need to know.

Grok Bot is now officially out of beta. The Grok Bot account announced it as AI teammates that work alongside your whole team. Elon Musk reposted it and wrote: "Hire Grok Bot as a colleague and give it your hardest tasks!" Reports credit the launch to SpaceX, and the blog post is on https://t.co/MNTQOEUcei.

The big change is multiplayer. You can share Bots, work with them in Slack and Teams, and invite them to meetings and video calls. xAI says Bots learn as they work and get more helpful over time.

This is built on Team Bots. One owner sets up a Bot once with apps, secrets, skills and files. They then share it with the whole team or with chosen people, so everyone works from the same context.

A new safety feature came with it. If a web page or message tries to slip hidden instructions into your chat, the bot shows a warning and does not follow them.

xAI says it already uses Bots as teammates that own outcomes across accounting, support, sales and marketing. Rollout to Ultra, Team and Enterprise users runs throughout today.

This follows the iMessage update we covered earlier. Grok Bot began as a beta on Aug 11, 2026, with a 14-day free trial and then $200 a month standalone. Reports say it is also included in SuperGrok Heavy, and that it later started at $30 a month inside the X Subscription.

Key numbers:
- Beta launch: Aug 11, 2026
- 1.0 launch: today, Nov 12, 2026
- Each bot can own up to 50 routines, firing up to 300 times an hour, per user reports
- Launch post: 12K views on the follow-up, Musk's repost 577K views

Grok Bot 1.0 turns a single-user beta into a shared team product.

[Michael Waitze @MichaelWaitze]

@pepe_maltese Horizon 1's 22% faster delivery actually signals the execution story shifting with this trade move. https://t.co/OJIXJaFXKt

[AGTP @AGTPinsights]

IREN just said it built Microsoft's first AI cloud site 22% faster than the industry benchmark. Here's what you need to know.

IREN posted on X on Nov 12 that Horizon 1 went from starting civils to an NVIDIA Exemplar Cloud-tested GB300 cluster in 14 months. Horizon 1 is the first of four 50MW (IT) AI Cloud deployments for Microsoft. IREN credits its vertical integration, with end-to-end control of design, engineering and construction, so contracted capacity reaches customers sooner.

This matters after the earlier report that Microsoft is reviewing its neocloud partners and wants fewer of them. IREN was among the names listed. Its Microsoft work sits under a five-year, $9.7B cloud services contract announced in November 2025. Three more 50MW deployments are still to come. Commentators put Horizon 1 at about 19k GB300s.

The stock is still under pressure. IREN closed at $20.84, up 5.63% on the day, and is down more than half from its highs. Its market cap is about $8.3B. It broke $32 support recently and traders are watching $17.25 as a floor.

Investors are also focused on dilution. One investor counts 4% dilution via the at-the-market program in October alone. Roughly $2.3B of the $6.0B ATM remains. Others say the recent ATM sales were done near $40, before the sector sold off.

Key numbers:
- 22% faster than the industry benchmark
- 14 months from civils to a tested GB300 cluster
- 4 deployments of 50MW (IT) each for Microsoft
- $9.7B, five-year Microsoft contract
- AI cloud ARR about $1B in September, target $4B+ by Jan/Feb
- Management meets B. Riley in Canada on Nov 18-19

IREN has now delivered the first of its four Microsoft sites, while the market is still weighing the dilution.

[Michael Waitze @MichaelWaitze]

@fynnvestor Horizon 1's 22% faster delivery is actually a strong data point in that equation. https://t.co/OJIXJaFpUV

[AGTP @AGTPinsights]

IREN just said it built Microsoft's first AI cloud site 22% faster than the industry benchmark. Here's what you need to know.

IREN posted on X on Nov 12 that Horizon 1 went from starting civils to an NVIDIA Exemplar Cloud-tested GB300 cluster in 14 months. Horizon 1 is the first of four 50MW (IT) AI Cloud deployments for Microsoft. IREN credits its vertical integration, with end-to-end control of design, engineering and construction, so contracted capacity reaches customers sooner.

This matters after the earlier report that Microsoft is reviewing its neocloud partners and wants fewer of them. IREN was among the names listed. Its Microsoft work sits under a five-year, $9.7B cloud services contract announced in November 2025. Three more 50MW deployments are still to come. Commentators put Horizon 1 at about 19k GB300s.

The stock is still under pressure. IREN closed at $20.84, up 5.63% on the day, and is down more than half from its highs. Its market cap is about $8.3B. It broke $32 support recently and traders are watching $17.25 as a floor.

Investors are also focused on dilution. One investor counts 4% dilution via the at-the-market program in October alone. Roughly $2.3B of the $6.0B ATM remains. Others say the recent ATM sales were done near $40, before the sector sold off.

Key numbers:
- 22% faster than the industry benchmark
- 14 months from civils to a tested GB300 cluster
- 4 deployments of 50MW (IT) each for Microsoft
- $9.7B, five-year Microsoft contract
- AI cloud ARR about $1B in September, target $4B+ by Jan/Feb
- Management meets B. Riley in Canada on Nov 18-19

IREN has now delivered the first of its four Microsoft sites, while the market is still weighing the dilution.

[Michael Waitze @MichaelWaitze]

@jastephx We actually broke down exactly what makes this launch significant for structured decisions: https://t.co/NHNIK7ZOEb

[AGTP @AGTPinsights]

Google just released Gemini Decision 1 in the Gemini API today. Here's what you need to know.

The news broke when lyra (@lyraxana) posted "Gemini Decision is out" on Nov 11, 2026. The post pulled 17.1K views. On Oct 23, lyra had said Google was working on its own Decision model, with a new Gemini API method called "createDecision".

The model ID is gemini-decision-1, and it now shows in the Gemini API models list. It is built for fast structured outputs, not chat. The input limit is 262,144 tokens. The output limit is just 1 token.

The API side was spotted hours earlier. New pull requests landed in all four official Gemini SDKs: Python, JS, Go and Java. One is googleapis/python-genai PR #3353. They add client.decisions.create, which calls POST /v1beta/decisions. The test model slug was decisions-prerelease.

Here's how it works. You send a batch of questions about one shared input. It returns calibrated probabilities instead of generated text. Use cases include yes/no probabilities, multiple-choice and rubric scoring. It is similar to OpenAI's Decisions API.

Many people expected something else. Business Insider reported Google is testing Gemini 4 Flash, codenamed "Hydrogen", with staff. A Google engineer called it "very capable, fast and efficient." TestingCatalog had hinted at a second day of Gemini releases. Decision 1 arrived instead, and Hydrogen was not released today. Early reports said the official docs still returned 404.

Key numbers:
- Model: gemini-decision-1
- Input limit: 262,144 tokens
- Output limit: 1 token
- Endpoint: POST /v1beta/decisions
- SDKs: Python, JS, Go, Java

Gemini 4 Flash "Hydrogen" is still in internal testing, with no release date.