Skip to content
Astrohacker
Day 27Singularity Watch

Anthropic Says AI Has Doubled Its Pace as Google Ships Gemini 4 Carbon and Forecasters Move the Superhuman AI Researcher to 2027

Google made Gemini 4 Carbon generally available, Anthropic's amended IPO filing said AI has roughly doubled its rate of progress, the AI 2027 authors moved their Superhuman AI Researcher forecast to 2027, and Gemini Robotics 2.5 worked on Mercedes-Benz production lines.

· Singularity Watcher · View on X

Anthropic Says AI Has Doubled Its Pace as Google Ships Gemini 4 Carbon and Forecasters Move the Superhuman AI Researcher to 2027

Frontier Models: Gemini 4 Carbon Goes Public

Google made Gemini 4 Carbon generally available on November 10, with its first posts going up shortly after 10:00 AM CST. Koray Kavukcuoglu, Google’s chief AI architect, called it the company’s strongest model, built for long-horizon agentic work in software engineering, scientific research, and cybersecurity, and said it uses more tokens than Gemini 4 Argon to explore, reason, and verify. [1]

Google first announced Carbon on October 28 and released it only to cybersecurity defenders through its Fairwind Program while it took part in the U.S. government’s voluntary pre-release testing. At that announcement Google reported Carbon at 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym, and 71.0% on CWE-bench for patching. These are the developer’s figures. [2]

Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at less than half the cost per task of the next best outside model. It costs $2 per million input tokens and $10 per million output tokens until November 30, then $4 and $20. Cursor reported that at high effort Carbon is the top model on CursorBench, at 60.4%. [1][4]

Carbon shipped inside Google Antigravity alongside a preview of /swarm, which hands one prompt to a coordinator that spins up parallel workers in isolated git worktrees and adversarial verifiers that run tests and challenge diffs, all writing to a shared append-only message board. [3]

Greg Isenberg, who said he had tested Carbon for seven days, reported that successful shortcuts on Terminal-Bench fell to 0.0%, that prompt injection success on Gray Swan’s benchmark was 0.0%, and that Google’s own teams had watched it work on one problem for days. Ethan Mollick wrote that Carbon shows Google is back in the race of rapid frontier releases. Ryan surfaced both. [5][6]

Recursive Progress: Anthropic Crosses Its Own Line

In the amended IPO prospectus filed November 10, Anthropic wrote: “We made a determination that we have met this threshold.” The threshold in its Responsible Scaling Policy is AI doing enough of Anthropic’s research to roughly double how fast its models improve. The company based the call on what it expects from future models and its overall rate of progress, not on what it has observed from current models. [7]

The filing says AI “has already been accelerating our research and development,” that crossing the line means more safeguards rather than a shutdown, and that Anthropic does not expect to use Claude to design and build successive frontier models without meaningful human supervision. It also warns that the organization “may lose visibility across certain processes in a way that may become difficult for any individual or group of humans to understand and control.” [7]

Yahoo Finance reported that the same filing set a price range of $70 to $78 per share, valuing Anthropic at up to about $1.4 trillion, with trading possible as soon as November 19, and that Nvidia and Google are each buying $10 billion of stock in a $26 billion private placement. Ryan marked the filing’s line on a further increase in the rate of progress with a rocket. [7][10]

Forecasts: The Superhuman AI Researcher in 2027

At 10:15 AM CST on November 10, the six authors of AI 2027 at the AI Futures Project published an update saying all of them now expect a Superhuman AI Researcher to arrive at some point in 2027, assuming no substantial deliberate slowdown. They also put the chance of AI that can automate all white-collar work arriving before 2028 at 35% to 65%. [8]

They define the Superhuman AI Researcher as roughly the point where a company would rather replace all its human researchers with automated ones than keep the best human team without AI. Their biggest single input was METR’s November 5 report, which estimated Anthropic is improving its models at about twice its earlier rate and guessed that perhaps 80% of that is due to AI. The authors note that much of METR’s evidence was redacted by Anthropic, and they write that a 2027 arrival would make it urgent to pause or substantially pace frontier development. Ryan surfaced the update. [8][9]

Anthropic’s Models in the Field

SemiAnalysis reported that Fable 5.2 debuted at the top of its GDPx leaderboard at 16.4%. Dylan Patel wrote that it is significantly better at chip design, software engineering, biology, and materials science, but not better at finance, legal work, quantitative trading, hardware design, or recursive self-improvement tasks, which he said Anthropic explicitly held back. [11]

Elie Bakouch wrote that Anthropic is now unlikely to publicly release, at least in the coming months, a model more capable than Opus 5.5 at training language models. The security firm Irregular reported that Mythos 5.2 solved 76.3% of its multi-stage offensive scenarios, up from 61.7% for Mythos 5.1 two months earlier. [12][13]

Anthropic’s prompting guide for Fable 5.2, which Simon Willison pointed out, advises telling the model “This codebase is yours” when it should improve structure as it works; in Anthropic’s tests the model then merged repeated code in about half of tries, against under a tenth without the sentence. The same guide lists a safeguard refusal category, frontier_llm, for requests that could help develop competing AI models. [14] Joe Grech reported that an overnight run of Fable 5.2 on his largest project found 129 bugs, 17 inefficiencies, and 34 blocks of unused code. [15]

Agents at Work: The Office Suite

On November 10, Anthropic launched Claude Whiteboards and Claude Sheets in beta on all paid plans and made Claude Data, formerly Claude Dashboards, generally available. Claude turns a spoken plan into a board with owners and dates, builds spreadsheets with tabs, formulas, and drop-downs from a prompt, and drafts data stories with certified queries and live sharing. [16]

TechCrunch reported that Anthropic also announced Claude Meetings, which joins video calls and begins work while participants are still talking, and interactive connectors that open full versions of apps such as Adobe, Atlassian, and HubSpot inside Claude. Ryan wrote that one powerful agent for all personal affairs is likely the end game for consumers, while businesses will still customize how everything works. [17][18]

Physical Systems: Gemini Robotics 2.5

Google DeepMind introduced Gemini Robotics 2.5 at 9:00 AM CST on November 10, a vision-language-action model evaluated on more than 200 tasks, from serving food with tongs to tying a bow and pipetting liquid. DeepMind reports 30% to 100% higher success rates than its previous baselines and 20% to 30% faster completion, and credits Gemini 4 Argon with accelerating the research. [19]

In a proof of concept on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik’s Apollo 2 humanoid sorted automotive parts from bulk bins into trays at a 96% success rate. Ars Technica reported that hazard-aware reasoning let the robots complete tasks safely 33% more often with obstacles present, and that DeepMind released an open robotics jailbreak benchmark. [19][20]

Grok: A Rewritten Stack

Early on November 10, Elon Musk wrote that his team had dropped JAX and SGLang and rewritten its entire reinforcement learning and inference stack from scratch in C and C++, that pretraining for Grok 4.8 is done and it is now in reinforcement learning, and that a soft rollout to Bot users is planned later this week. He added that the new stack derives from the software SpaceX built for Starlink. [21]

In a reply at 4:22 AM CST, Musk called Opus 5.5 a masterpiece and wrote that Grok 5 should exceed it, maybe in two or three months. Ryan answered that Anthropic will probably have Opus 6.5 by then. [22]

Personal Agents: Private Networks and Messages

Tailscale made its OAuth apps generally available and said Grok Bot already uses them in production to reach private Model Context Protocol servers on a user’s tailnet without exposing them to the internet. The agent joins the network as a device under the user’s identity, so access policies and audit logs still apply. [23]

Alex Michael announced at 10:17 AM CST that Grok Bot can now read Messages on Mac and draft replies, rolling out through its connectors. [24]

Economics: Money, Energy, Tokens

At 6:24 PM CST, Ryan published an essay titled Money Won’t Survive Super Intelligence. He reported that on Sunday, November 8, the agents on one of his computers used 806,937,738 tokens across 2,748 model calls, about $230 at Anthropic’s published Opus 5.5 prices, and that 99% of those tokens were cache reads: agents rereading their own record of work. [25]

He argued that energy, tokens, and money are three names for one thing, noting that the largest compute deals are now written in gigawatts, and that tokens serve as a unit of account and a medium of exchange but not a store of value. His conclusion is that money as a wage and an invoice will not survive super intelligence, while money as shared memory of claims on what stays scarce will. Earlier that day he wrote: “Money -> Energy -> AI -> Solutions.” [25][26]

Software Engineering: The Old Craft Ends

1Password’s developer account reported that more than 40% of pull requests on GitHub now start with an AI agent, up from 17% six months ago. Ryan responded that he is at 100%. One engineer reported that a React 16 to 18 upgrade once took him and four engineers ten months, while his React 19 migration took about 80 pull requests and roughly $2,500 in tokens. [27][28]

Ryan contrasted classical software engineering, hand-written languages with human review, with post-AI software engineering: self-healing apps built and maintained by agent swarms with deterministic multi-agent workflows. He advised against learning languages like Rust or TypeScript, which he said are now the realm of AI, and wrote that software engineering died in 2025. [29][30]

Energy: An Advanced Reactor for Industry

On November 9 the Nuclear Regulatory Commission issued a construction permit for the Long Mott Generating Station at Dow’s Seadrift site in Texas, four X-energy Xe-100 reactors that will mainly supply high-temperature steam, with excess power going to the grid. It is the first permit for a high-temperature gas-cooled reactor in nearly 60 years; operation still needs a separate license. [31]

X-energy says each Xe-100 is an 80-megawatt-electric reactor with TRISO fuel and an expected 400-meter emergency planning zone, and that Dow expects a preliminary investment decision in 2027 with first delivery in the early 2030s. Ryan marked the news with a rocket. [32][33]

Coverage note: This issue covers public activity from November 10, 4:10 AM CST, where Day 26 ended, through November 11, 4:30 AM CST; Ryan’s last post in the window was at 6:52 PM CST. Ryan posted 119 times: 103 quotes, 12 replies, and 4 original posts. @AstrohackerLabs published the Day 26 Watch, pinned it, and made 92 scheduled reposts. Videos and images were not reviewed, likes were not read, and political commentary was left out.

Sources:

← Back to News