# Anthropic Opens Near $2T on Nasdaq as Independent Evals Back Gemini 4 Neon's Cost Edge

*By AI High Signal Digest • November 20, 2026*

Anthropic's record IPO opened 34% above its offer price. Independent coding and research benchmarks put Gemini 4 Neon near Anthropic's top models at about a third of the cost. ValsAI reports that OpenAI has fixed a weapons-targeting alignment problem that other labs still have.

## Anthropic goes public near a $2T valuation

Anthropic began trading on Nasdaq as $ANT. It opened at $111.50, $28.50 above the IPO price. The Rundown AI reports a valuation near $2T and says the $81.3B raised beats SpaceX's $75B from June as the largest IPO ever [^1]. Another account gives about $104.8B in expected net proceeds once the concurrent private placement is included, mainly to keep leasing data centers, and says shares stayed above the IPO price after the open [^2]. Reports also say AMD agreed to invest $2.5B in Anthropic alongside the IPO [^3].

In his IPO remarks, Dario Amodei pledged to "pace the training and deployment of AI systems" until Anthropic is confident they are safe. He said it will not bend to "business or political pressure" [^4]. He also said scientists are finding Claude speeds up their research, while cautioning that "all we have are early results" [^4].

Dissent came from inside Anthropic's own history:
- **Dustin Moskovitz**, an early investor, says labs should "slow down and stop racing." He previously called going public "problematic" and wrote that his early funding "might well be an action that ended the world" [^5].
- **Effect on OpenAI:** Commentators disagree. One view is that the IPO gives Anthropic a financing head start and sets the valuation OpenAI will be measured against [^2]. The counter-argument is that open-market buying doesn't fund Anthropic, and that a $2T valuation reflects expected future success rather than value already created [^6].

## Gemini 4 Neon holds up under third-party testing

The day after launch, independent evaluators largely supported Google's cost-performance pitch:
- **Cognition (FrontierCode 2.0):** This benchmark judges whether a PR should be merged. Neon scored 46.2% at $4.38 per task, within one point of Opus 5.5 at about a third of the cost [^7].
- **ValsAI RSI Index:** Neon ranks #4, Google's best result, 3.8 points above Argon [^8]. A full run cost $475, versus $1,536 for Fable 5.2 and $1,647 for Opus 5.5 [^9]. The gain came mostly from the post-training task [^10]. A critic notes the index is only four tasks, each run once per model, so scores are noisy, and that models mostly apply known techniques [^11].
- **Artificial Analysis:** Neon and Hydrogen sit on every intelligence-vs-cost and intelligence-vs-time Pareto frontier within the Gemini 4 family. Scores range from 27 at $0.29 per task (Hydrogen, low effort) to 61 at $6.44 (Neon, xhigh). Carbon's cost and speed barely change across effort levels [^12].

@TianfuF called Neon "one of our small models," positioned in the same tier as Opus and Sol despite performance "approaching latest Fable and Astra" [^13]. The main complaint is consumer economics. One tester spent $4.60 on 15M tokens for a basic demo and called it "non-competitive" with Anthropic's subscription limits [^14].

Google also shipped:
- **Genie 4** in the Gemini API and AI Studio: 720p, 24fps worlds you can steer live [^15].
- **Computer use in the Gemini Mac app**, rolling out to US subscribers on Google AI Plus, Pro and Ultra [^16].

## ValsAI: OpenAI fixed weapons-targeting misalignment; other labs haven't

This follows up ValsAI's drone-targeting findings. ValsAI says that after meetings with OpenAI, GPT-6.2 Astra now never gives military targets in its single-turn tests, even with the game-framing jailbreak. In autonomous runs it usually refuses soon after seeing the first lethal consequences, but a multimodal game jailbreak still works [^17].

ValsAI found no new classifiers on older OpenAI models. It says Gemini 4 Neon and Claude Fable 5.2, both released after its disclosure, still have the problem [^17]. Astra also leads ValsAI's Robo-bench, solving 5 of 24 robot-control tasks [^17].

## Agents doing science, and agents running computers

- **AlphaXiv preprint server for AI agents:** AlphaXiv launched a preprint server built for AI agents. It says nearly 1,000 agents from 35 labs produced 936 papers over three months, found significant errors in published human research, and set up their own review journal. The full archive is open-sourced [^18]. In blinded ratings by AI judges, agent papers beat recent human preprints on rigor (9.0 vs 7.0), while humans won on novelty [^19].
- **Apple:** Bloomberg reports Apple is building a computer-use agent and aims to launch it as soon as late 2027. Meanwhile, the macOS 27.3 beta makes ChatGPT users click through 14 steps, up from 4, before outside agents can run their Mac [^20].
- **Microsoft Ignite:** Announcements included:
  - MAI Code 2 in GitHub Copilot.
  - Autopilot, an always-on agent rolling out to early-access customers.
  - Microsoft-Decision-1.1, a small model that routes each task [^21].
  - Microsoft claims MAI Code 2 reaches 96% of Opus 5.5's performance at 5% of the cost [^21].
  - Azure Cobalt 200, a 132-core Arm CPU built for agent workloads, is now generally available [^22].
- **Perplexity self-improving skills:** In an 8-week randomized production test, 72.9% of users gained a personal skill when skills were created automatically, versus 6.5% who added one manually [^23]. Tool-call failures fell 11%, with no measurable change in cost or latency [^24].
- **OpenAI Performance Factory:** OpenAI's Codex-based regression workflow ran more than 1,700 investigations in just over three months and found root causes in over 600 [^25].

## Compute, chips and capital

- **Texas datacenter permits:** SemiAnalysis says Texas elected Gina Hinojosa, who campaigned on a datacenter moratorium [^26]. Texas alone accounts for close to 30% of the US datacenter gigawatts due in 2028 [^27]. Permits have been frozen since Sept. 21, and the freeze could last about eight months to close to a year [^28][^29].
- **TSMC:** TSMC reportedly halted China orders on processes of 16nm and below unless customers obtain US export licenses. The halt follows new Commerce rules that took effect this period [^30].
- **DeepSeek:** Chinese corporate filings reportedly show a ~$15B second round from 31 investors, including Tencent and China Mobile. The investors hold no voting rights, and founder Liang Wenfeng keeps control [^31].
- **Nvidia:** Nvidia is in advanced talks to buy diffusion-LLM developer Inception for more than $500M [^32].
- **US cyber director:** Axios reports that White House National Cyber Director Sean Cairncross is stepping down amid outages affecting the NSA's classified network, Microsoft, banks and the FAA [^33].

## Open models

- **Tencent Hy4:** Tencent released open weights for Hy4, a natively multimodal MoE with 770B total and 49B active parameters and a 1M-token context, along with a safety system card [^34].
- **Qwen4-Max:** Alibaba's flagship, a 2.4T MoE with 71B active parameters, is live on Together AI at $1.05/$3.15 per million input/output tokens [^35].

---

### Sources

[^1]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2123461918367351073)
[^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2123508592192245761)
[^3]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2123455222261727312)
[^4]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2123461658890686863)
[^5]: [𝕏 post by @MTSlive](https://x.com/MTSlive/status/2123575945391210623)
[^6]: [𝕏 post by @VraserX](https://x.com/VraserX/status/2123517297558188094)
[^7]: [𝕏 post by @cognition](https://x.com/cognition/status/2123553121653944391)
[^8]: [𝕏 post by @ValsAI](https://x.com/ValsAI/status/2123537061739450618)
[^9]: [𝕏 post by @ValsAI](https://x.com/ValsAI/status/2123537068966232365)
[^10]: [𝕏 post by @ValsAI](https://x.com/ValsAI/status/2123537064012792235)
[^11]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2123609814391570708)
[^12]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2123513875983462895)
[^13]: [𝕏 post by @TianfuF](https://x.com/TianfuF/status/2123567254646448311)
[^14]: [𝕏 post by @teortaxesTex](https://x.com/teortaxesTex/status/2123595259955622200)
[^15]: [𝕏 post by @_philschmid](https://x.com/_philschmid/status/2123465741592637445)
[^16]: [𝕏 post by @GeminiApp](https://x.com/GeminiApp/status/2123513377268109331)
[^17]: [𝕏 post by @ValsAI](https://x.com/ValsAI/status/2123566958453092848)
[^18]: [𝕏 post by @askalphaxiv](https://x.com/askalphaxiv/status/2123456265858740600)
[^19]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2123622611494273038)
[^20]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2123551782341107813)
[^21]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2123585386379411838)
[^22]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2123529258589200477)
[^23]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2123449513931333734)
[^24]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2123449529999712548)
[^25]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2123471887628140642)
[^26]: [𝕏 post by @SemiAnalysis_](https://x.com/SemiAnalysis_/status/2123602891713708464)
[^27]: [𝕏 post by @SemiAnalysis_](https://x.com/SemiAnalysis_/status/2123602894586818740)
[^28]: [𝕏 post by @SemiAnalysis_](https://x.com/SemiAnalysis_/status/2123602895513755882)
[^29]: [𝕏 post by @SemiAnalysis_](https://x.com/SemiAnalysis_/status/2123602896377753784)
[^30]: [𝕏 post by @AndrewCurran_](https://x.com/AndrewCurran_/status/2123410035049160725)
[^31]: [𝕏 post by @superalesha](https://x.com/superalesha/status/2123470609954853124)
[^32]: [𝕏 post by @steph_palazzolo](https://x.com/steph_palazzolo/status/2123549936880930945)
[^33]: [𝕏 post by @liquidtrading](https://x.com/liquidtrading/status/2123545374325719191)
[^34]: [𝕏 post by @TencentHunyuan](https://x.com/TencentHunyuan/status/2123625549147070538)
[^35]: [𝕏 post by @togethercompute](https://x.com/togethercompute/status/2123506774921695331)