Anthropic targets pre-Thanksgiving IPO at up to $1.75tn

Anthropic seeks $80-100bn in a pre-Thanksgiving IPO, METR says its models improve twice as fast, and GPT-6 Astra fully solves 26% of real coding tasks.

GENERAL Newsletter
H
Hugenode

Anthropic is pressing ahead with a stock market listing before Thanksgiving. Because of a tech selloff, it is accepting a lower valuation than its most bullish investors wanted. A METR report says AI is now driving most of the acceleration in Anthropic’s own model development, though Anthropic redacted much of the evidence. And an Alibaba-led benchmark shows that coding agents with near-perfect scores on older tests still leave most real user requests unfinished.

Anthropic keeps its pre-Thanksgiving IPO on track at a lower valuation

Anthropic wants to raise $80-100 billion at a valuation of $1.25-1.75 trillion, according to investors cited by the Financial Times and reported by The Decoder. That is above Anthropic’s current $965 billion valuation but below the $2 trillion the most optimistic investors had hoped for. The Decoder puts Anthropic’s annualized revenue at about $81 billion, up from roughly $65 billion at the end of July.

The Decoder attributes this week’s tech selloff to rising borrowing costs and heavy AI spending plans. Democratic midterm wins accelerated it, because the party had promised to slow data center construction. The Nasdaq 100 fell 6.4 percent on Thursday and recovered somewhat on Friday. We covered the market rout and the data center ballot measures earlier this week.

Latent Space is more cautious about the price. It says some valuation estimates have fallen as low as $1 trillion. It puts Anthropic’s run rate at $80 billion against a year-end target of $100 billion. It also suggests the rout might be leveling off at around a 10 percent Nasdaq decline, half the 20 percent drop that would mean a bear market.

OpenAI is raising money in the same market. The Decoder reports that OpenAI cut its funding-round target from $1.4 trillion to $1.25 trillion and raised only about a third of the $30 billion it originally sought, as we reported.

What it means: for teams building on Claude, nothing changes operationally yet. In our view, a listing this large would put Anthropic’s revenue growth and spending under much closer public scrutiny. If the deal prices near the low end of the range, it would suggest investors still fund frontier labs, but at prices anchored to revenue rather than expectations.

METR report says AI is accelerating Anthropic’s own model development

The AI safety research group METR published a report on Thursday after an 11-week investigation inside Anthropic. According to The Verge, METR found that Anthropic’s models are improving roughly twice as fast as a year ago. Benchmark scores suggest about two years of historical progress in the last year. METR says it is confident this is not simply the result of more training compute. It attributes the majority of the acceleration, as of October, to AI itself.

Part of the evidence is self-reported. In informal Slack polls and interviews, employees using Anthropic’s most capable internal model said they were on average about 12 times as productive as they would be without AI. That is roughly double what staff reported a few months earlier while using Mythos 5. Much of the acceleration still appears to come from AI assisting human researchers rather than working on its own. METR does see signs that fully autonomous contributions will soon play a larger role, and it found no clear signs of diminishing returns.

METR also says Anthropic finished training three unreleased models over the past six months. They are codenamed Model 2, Model 3 and Model 4, and they completed training in July, August-September and September-October respectively. METR notes that these are the codenames shared with it, and that Anthropic uses different names internally.

The main caveat is the redactions. METR says it submitted a roughly 100-page report and that Anthropic redacted a majority of the evidence METR found most informative. A small number of its data requests were also denied. METR acknowledges that Anthropic shared more than originally scoped. It also says most of the redacted material was very high-level. As a result, the report says, “the public has to rely on METR’s judgment much more than would be ideal”.

Anthropic cofounder Sam McCandlish responded on X. He said that revealing what drives the speedup would be commercially sensitive and would risk handing others a roadmap to go faster. He added that Anthropic will publish its own report on the pace of AI progress, along with new safety steps, in the coming weeks.

The Verge reports a fierce backlash from safety researchers and commentators. They point to the gap between Anthropic’s safety advocacy and the acceleration the report describes. CEO Dario Amodei has publicly called on labs to slow down, and the company is preparing to go public at the same time.

What it means, in our reading: the Claude models customers use today trail what Anthropic runs internally, and three finished models have not been released. If you plan around current Claude capabilities, allow for them to shift sooner than an annual planning cycle assumes. Treat the 12x figure as self-reported, not measured.

Alibaba’s Co-Design Bench: GPT-6 Astra fully solves only 26.4% of real coding tasks

Researchers from Alibaba and several Asian universities built Co-Design Bench from real user requests to change existing web apps. Some of those requests include images, audio, video or documents. The Decoder reports that GPT-6 Astra, the best of ten models tested, met about 70 percent of a task’s requirements on average. It produced a fully correct result in only 26.4 percent of tasks. The authors built the test because older benchmarks are saturated. On them, agents that can run, check and fix their own code score above 99 percent, and scores on tests such as WebCompass-edit are near-perfect.

The researchers favored requests that users had to rephrase several times or roll back, so the set leans toward tasks that had already caused agents trouble. One AI judge tests each finished app in a browser. A second judge re-checks every positive verdict and looks for mock-ups or hardcoded results.

Claude Opus 5 and Qwen3.8-Max ranked behind GPT-6 Astra, and Grok 4.6 came last. GPT-6 Astra fully solved 41 percent of multimodal tasks and just under 18 percent of repository tasks. Every model did worse once a task demanded three or four types of change across layout, design, interaction and content.

Two findings matter when choosing a tool. First, spending more tokens did not fix the weakness. GLM-5.2 used nearly six times as many tokens as GPT-6 Astra on smaller changes and still scored significantly worse. Second, the harness mattered. Qwen3.8-27B scored 75 or 56 percent on larger overhauls depending on the harness, and the harness could flip which of two models came out ahead. That second finding is not settled: The Decoder cites another recent study that found barely any measurable benefit from more elaborate harnesses, even though they used up to three times as many tokens.

The benchmark has limits. It tests one fixed request per task with no follow-up corrections, and the AI judges can make mistakes. Alibaba’s own Qwen models place well, and the smaller Qwen3.8-27B ranked second on larger overhauls, though the authors do not claim a general advantage for it.

What it means: test the model and its harness together, on tasks from your own backlog, and budget for human cleanup on most agent output. The result fits an earlier METR study cited by The Decoder. In that study, project leads accepted AI-generated code at a rate roughly 24 percentage points below SWE-bench Verified scores. We have covered this pattern before in what agent benchmarks miss.

What to watch

  • Anthropic’s IPO pricing. Does it land inside the $1.25-1.75 trillion range before Thanksgiving, or drift toward the $1 trillion estimates if the selloff resumes? A related question is whether the run rate reaches the $100 billion year-end target.
  • Anthropic’s own report. McCandlish promised a report on the pace of AI progress and new safety measures in the coming weeks. Watch whether it discloses any of the evidence redacted from METR’s version.
  • Model 2, 3 and 4. METR says all three have finished training. No release dates have been reported.
  • Harness reporting. Watch whether coding-agent leaderboards start listing the harness alongside the model, as the Co-Design Bench authors recommend.

Sources