ZeroNoise Logo zeronoise
Post
Gemini 4 Neon Undercuts Rivals on Price as GPT-6.2 Reviews Sour and Anthropic Prices IPO
•
5 min read
• 1082 docs
Google's mid-tier Gemini 4 Neon is landing near the top of independent leaderboards at a fraction of rival costs, while GPT-6.2 sees its first reverts. Also: OpenAI publishes research-acceleration and chip-design data, Anthropic prices its IPO above range, and export-control pressure spreads through the compute supply chain.

Gemini 4 Neon: near-frontier scores at roughly half the cost

Google released Gemini 4 Neon, which sits between Carbon and Hydrogen in its lineup. It costs $3 per million input tokens and $15 per million output tokens, with a 1M-token context window. It is rolling out in the API, AI Studio, Gemini for business, Antigravity and the Gemini app . Arena says this is Google's fastest release cycle yet: 28 days from Argon to Carbon, 15 from Carbon to Hydrogen, and 6 from Hydrogen to Neon .

Independent evaluators largely support Google's positioning:

  • Artificial Analysis: Neon scores 61 at xhigh, 3rd overall behind GPT-6.2 Astra and Claude Fable 5.2. It costs $6.44 per task versus Carbon's $7.76 and takes about 21 minutes versus 36. On the Coding Agent Index it scores 59, in line with Claude Code running Fable 5.2 or Opus 5.5. Its average cost per task is $6.86, against $12.61 for Opus 5.5 and $15.47 for Fable 5.2 .
  • Vals Index: Neon is #2 at 70.9%, 0.4 points behind Fable 5.2 at less than half the cost per task. It is #1 on Vals' finance and legal indexes .
  • Agent Arena: Neon is #2 at 1,336 points from 4.2k real sessions. It beats Opus 5.5 at 53% lower median cost per task and costs 69% less than #1 Fable 5.2 .
  • Cursor: Neon scores 63.7% on CursorBench 4.1, up from 57.3% for Hydrogen .

Google also claims its lowest fabrication rate yet in internal testing . A cyber variant reaches 76% on the hard tier of AutoBaxBench and 51.26% on WebExploitBench .

Early reactions are positive, including from people who criticized GPT-6.2. Theo called Neon "so much more pleasant to work with" than Astra . Jeremy Howard is reverting Solveit's Astra and Sol to their pre-6.2 versions because "6.2 isn't working for me, or for our users." He says this is the first model rollback in two years of building the product .

GPT-6.2: an original proof, but a style regression

The strongest new evidence for Astra came from ValsAI. It reports that Astra produced what it believes is the first Lean proof of Smoktunowicz's Gap Theorem . Astra did not use the three deep results the published proof relies on, none of which are in Mathlib. Instead it defined a new quantity it calls "capacity" . ValsAI says no human touched the run and the proof passed independent kernel replay . On the same benchmark, Sol and Fable 5.2 tied for second, and Fable cost nearly three times as much .

Arena ranks Astra #6 on its new Agent Arena Index, OpenAI's best model there .

Artificial Analysis measured how the models' style changed. Astra's responses are 40% shorter than its predecessor's . Its score on the "Admits Uncertainty" trait fell from 96 to 24 . That shift may help explain the complaints about how the model behaves in practice.

OpenAI publishes research-acceleration data and calls for RSI reporting

OpenAI released new data on how AI is changing its own research. It says researchers' AI use is rising exponentially and human effort on safety and alignment is growing fastest of any research area. It argues labs should be required to track and publicly report indicators of recursive self-improvement (RSI) .

Figures quoted from the report:

  • Agents now complete 46% of 2–4 hour research tasks with no human help, up from 23% in March .
  • Median researcher agent use is $2,400 a day at API prices, and the top 10% spend $23,000 .
  • A 35% autonomous success rate for launching training runs drew skepticism from practitioners .

Readers disagree on what the data shows. Theo Jaffee argues it does not look like imminent superintelligence: agent workdays have grown jaggedly and success on short tasks is flat . @scaling01 counters that agent workdays grew 9–16.4x since May, and that tasks using swarms of 8+ agents rose from 1.9% to 35.8% .

Separately, OpenAI says its models found up to 14% PPA (power, performance, area) improvement on AMD's flagship GPU IP in under five months. It also says they improved its own Jalapeño inference chip, with gains verified on silicon . Greg Brockman says some of the AMD results are in a design that has recently taped out .

Anthropic prices IPO above range amid a safety backlash

Anthropic priced its IPO at $83 per share, above the $70–78 range. It trades on Nasdaq as ANT .

On the same day, the Center for AI Safety launched a statement calling Anthropic "not a safe AI company." Signatories include Steven Pinker, Thomas Wolf and Igor Babuschkin . Dan Hendrycks said many safety groups privately agreed but feared losing funding . He also stopped advising AI companies .

Epoch's analysis of the S-1:

  • Anthropic spent $10B on compute in Q3 alone .
  • Only about 26% of compute went to customer inference in the first nine months of 2026 .

Export controls spread through the compute supply chain

  • Nvidia has stopped shipping licensed Hopper products to China. It sold $2.4B of them there last quarter and now guides to zero .
  • Datasection: The Japanese neocloud cut revenue guidance 89% and hit limit-down . A major add-on order was cancelled on Nov 16 . The FT identified the customer as Tencent . No cause has been disclosed. SemiAnalysis notes that this kind of offshore rental is what remote-access rules target .
  • Samsung has reportedly stopped foundry work on Chinese AI chips under the new BIS rules .
  • Disclosure proposal: The administration is considering requiring US companies to publicly disclose any use of Chinese models such as Qwen .

Safety research and other launches

  • Hiding reasoning from monitors: OpenAI researchers trained a model only to control the style of its chain of thought. Its monitorability on covert side tasks still fell from 0.94 to 0.68. Prompts asking it to reason transparently recovered much of the loss .
  • Self-improvement without instructions: A paper from MIT CSAIL and Tübingen found that frozen models given simple tools built their own learning procedures, without being told how to learn .
  • Genie 4: Google's world model is now in Project Genie, with prompt-based editing mid-session and native audio . It is rolling out to eligible Ultra subscribers .
Gemini 4 Neon Undercuts Rivals on Price as GPT-6.2 Reviews Sour and Anthropic Prices IPO
AI High Signal
  • Theo Jaffee argued that OpenAI’s research-acceleration data did not establish that ASI was imminent: he cited jagged agent-workday growth, little swarm-size growth, token spending still below human-worker spending, little growth in experiments per researcher or internal ECI, limited improvement on tasks under 15 minutes, and low, mostly flat agent task shares. He allowed that Anthropic might be much further ahead or capabilities could accelerate suddenly, while treating imminent full automation of AI R&D and whether it would produce broad real-world gains as open questions.
  • In rebuttal, @scaling01 reported 9–16.4× agent-workday growth across domains since May, and 19× growth in tasks using swarms of 8+ agents (1.9% to 35.8%) and 23× for swarms of 16+ (0.5% to 11.7%). They said token spending was already roughly level with researcher compensation at API prices and could exceed it by early–mid 2027 even after a 5–10× API-margin discount; experiments per researcher were 2.3× higher since January, though not accelerating in curve. They also reported a 23% fall in failures on tasks under 15 minutes and growth since March in agent shares of longer tasks: 1.9× for 2–4-hour tasks, 3.1× for 4–8-hour tasks, and 5× for tasks over 8 hours. They cited new ECI/OECI trends and argued that public ECI can lag acceleration by 2–4 months.
Seems like the OpenAI research acceleration blogpost is not what you would expect if we were definitely on the cusp of artificial superin… I don't agree with this at all and I'm not sure if Theo is even looking at the same plots. let's say Theo's expectations of what the cusp…
AI High Signal

A post claimed DeepMind would release “Gemini 4 Sodium” the next day at $11.75/M introductory pricing, rising to $16.50. In response, MParakhin said DM’s smaller distillation is cheaper to run and permits a higher reasoning budget; he said it performs well on benchmarks, while Carbon is better in real-world use. He argued Google should release its best models rather than build a marketing-driven lineup. The launch and pricing are claims in the quoted post, not confirmed by MParakhin’s reply.

tomorrow DeepMind will be dropping Gemini 4 Sodium, a fast yet powerful frontier model which costs just $11.75/M (introductory pricing, r… That is unfair. Carbon is absolutely great, DM does a smaller distillation (Fable->Opus style), cheaper to run - so it gives it a higher …
AI High Signal
  • Simon Willison compared Gemini 4 Carbon, the newly noted Gemini 4 Neon, and Gemini 4 Hydrogen on a pelican-riding-a-bicycle prompt at low, medium, and high thinking levels, using the test to show differences between models and effort settings in the same family.
  • The comparison image’s high-effort labels report 20,305 output tokens, 20.31¢, and 4m 11s for Carbon; 9,691 tokens, 14.55¢, and 1m 11s for Neon; and 5,311 tokens, 3.99¢, and 19s for Hydrogen.
  • Willison also tried Gemini 4 Neon with a prompt for a Three.js animation of a giant mech version of Yonkers stomping around Manhattan, Queens, and Brooklyn using OpenStreetMap data, and linked the result.
The most useful thing about the pelican riding a bicycle test continues to be illustrating the differences between models in the same fam… I also tried "Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stomps around manhattan and quee…
AI High Signal

@rishdotblog says newer Gemini models are strong at agentic work but misses the world knowledge of older Flash and Flash-Lite models; the author describes Luna as uniquely positioned across knowledge, latency, and cost, without providing comparative metrics.

new gemini models are great at agentic work. but i miss the old flash + flash-lite models for their world knowledge luna is on an island …
AI High Signal

@teortaxesTex disputed an unspecified claim by Jensen about China’s compute capacity, asserting that China would not exceed the rest of the world excluding the US and might not exceed Southeast Asia’s combined capacity; the post gives no timeframe or supporting figures .

I hope we can all be adults and admit that Jensen was bullshitting here China won't even have more compute than the rest of the world mod…
AI High Signal

@sun_hanchi argues that RSI rhetoric favors large AI companies: the post says iteration speed depends on existing model capability and GPU scale, and that incumbents use this argument to suggest startups cannot overtake them through technical ingenuity. It warns that smaller startups and VCs may mistake RSI for an opportunity, while saying data vendors have benefited from the trend.

Crazy how many startups talk about RSI. It's so stupid. RSI is a business story tech giants tell to explain why the big get bigger and th…
AI High Signal

Gemini’s gaming and 3D efforts were highlighted alongside a demo in which Neon turned a raw Smithsonian woolly mammoth scan into an anatomically arranged and rigged 3D skeleton, then generated a museum walkthrough; the team presented the work as bringing together spatial understanding, multimodal reasoning, and coding.

A frontier model that masters symbols, space and time is all you need. It's been pleasure collaborating and working with Gaming and 3D ef… To see how its 3D spatial understanding, multimodal reasoning, and coding come together, the team gave Neon a raw Smithsonian 3D scan of …
AI High Signal

A Zhihu contributor’s review says GPT-6.2’s Sol approaches the previous Astra’s reasoning performance, while Astra is more efficient but struggles to make its work fit together; the review frames GPT-6.2 as improving problem-solving while still struggling with human intent.

1/n GPT-6.2 raises the ceiling on problem-solving while stumbling over human intent. 🧠 That’s the tension in Zhihu contributor toyama nao…
AI High Signal

Kling Image 4.0 is now available on Morphic ; it generates film-like stills in 2K or 4K using up to 10 references, which can then be turned into a Kling 4.0 shot .

Kling Image 4.0 is now available on Morphic. [#KlingAI](https://x.com/hashtag/KlingAI) [#KlingImage4](https://x.com/hashtag/KlingImage4) … Kling Image 4.0 is now available on Morphic. Make stills that look like frames from a film, in 2K or 4K, guided by up to 10 references. T…
AI High Signal

OpenAI put discounted processing in Responses API background mode into public beta: for single- or multi-turn workflows that can wait, GPT-6.2 Astra or Sol can save up to 40% per token versus Standard, with a 2-hour or 8-hour target for the first response.

Not every workflow requires an immediate response. For single- or multi-turn workflows that can wait, save up to 40% per token than Stand…
AI High Signal

The post argues that context or knowledge is the hardest part of building agents and recommends building a “knowledge factory,” drawing an analogy to SaaS companies building software factories.

saas runs on software, so everyone is building a software factory the hardest part of building agents is the context (or knowledge that t…
AI High Signal

@scaling01 alleges that someone downloaded all their tweets, identified their best-performing formats (greentexts and tier lists), and is now generating AI posts in those styles; the account frames AI-generated tweets as a way to gain followers.

btw this person downloaded ALL of my tweets, threatened to doxx me and found that my most successful tweets were greentexts and tier list… they are literally using my tweets to generate as template to generate their tweets you can just lie and generate your tweets with AI to get followers btw ![](https://pbs.twimg.com/media/HXdTWQ4XMAAps_j.jpg)
AI High Signal
  • A post quoting a Chinese report says Samsung has stopped foundry production of compute chips for China under new U.S. BIS rules. The report describes Cambricon’s 690/580 chips as mainly domestic, says the 690 also involved TSMC and Samsung and some 590 chips were made at Samsung, and expects the affected supply routes to harm next year’s revenue. @teortaxesTex characterizes the restrictions as a gradual embargo and warns that mainland Chinese actors cannot assume long-term access to compute or production capacity outside mainland China.
  • A separate post explicitly labels as rumor a claim that TSMC is restricting tape-outs below 12nm; it says Xiaomi’s Xuanjie chip and Horizon/Black Sesame automotive chips could be blocked, with Xuanjie potentially requiring a process change. Treat this claim as unverified.
受美国BIS新规影响,三星Samsung已经停止中国China算力芯片代工 寒武纪:690/580主要国内,690还有TSMC和三星,后面基本没戏了,还有一些590在三星,应该也不行了,对明年收入有一定负面影响 ![](https://pbs.twimg.com/media… No compute or means of producing compute outside Mainland China can be presumed to be long-term available to Mainland Chinese actors. It'… 小道消息: 台积电也限制12nm以下的芯片流片了, ————— 这次连小米的玄戒芯片都要凉,或者改制程。地平线黑芝麻的车载芯片也都封掉,有点离谱了。 PS:传闻消息,供参考,不做投资意见。 ![](https://pbs.twimg.com/media/HXcwgEsbIA…
AI High Signal

Vladimir Putin claimed that Russia’s two flagship language models rank among the world’s best, but the cited statement names neither models nor supporting evidence.

Putin: Russia is at the forefront of digital transformation. Our two flagship language models rank among the best in the world. [![Video]…
AI High Signal

An Opus-translated interpretation of Zartbot's post suggests a classified training project involving a 1-million-GPU cluster that would not use HBM; this is a hint, not a confirmed announcement. The commentator also says claims of 500B active and 300T tokens are plausible with modern clusters, despite being extraordinary compared with the open-source world.

more interestingly (a translation from a serious model for serious people, ie Opus): Zartbot hints at a classified research project for 1… numbers here are bonkers compared to the OSS world. And Zartbot is not some influencer, he's a careful technical thinker. 500B active, 30…
AI High Signal

Theo describes Neon as a counterbalance to GPT-6.2 Astra and says it is more pleasant to work with, offering a subjective user-experience comparison rather than a benchmark. He frames differences between model families as capability (what a model can do at its best) versus coherency (whether it understands what it is doing), and says models have never felt so different.

Neon feels like a weirdly perfect counter balance to GPT-6.2 Astra. It's just so much more pleasant to work with. I'm getting whiplash. [… Capability = "at it's best, what can it do" Coherency = "does it understand what it's doing" Best method I've come up with to explain the…
AI High Signal

A proposed test for recursive self-improvement (RSI) is whether an AI gets better at producing subsequent improvements, not merely whether it trains a better model. PostTrainBench, RSIBench-Data, and AutoRLBench respectively assess autonomous post-training, improving data strategies from experimental feedback, and diagnosing and repairing RL failures. In AutoRLBench’s initial evaluation, agents detected unhealthy training in 82.6% of cases but identified the correct root cause in only 28.2%; a cited analysis found training-strategy changes in just 74 of 3,557 consecutive experiment pairs (about 2.1%). The post argues that RSI evaluation should test the whole improvement process, including reliable diagnosis, evidence-driven strategy changes, transfer to new tasks, and improved research in the next round.

🧠 What would actually count as recursive self-improvement (RSI)? An AI training a better model is only part of the story. Zhihu contribut…
AI High Signal

Prime Intellect introduced Model Factory, a single stack for training, deploying, and continuously improving users’ own models, with options to run on demand, on dedicated clusters, or on customer-provided infrastructure.

Introducing Model Factory. Train, deploy, and continuously improve your own models on one stack. Run on-demand, on dedicated clusters, or…
AI High Signal

Following another model release that he described as unexpectedly good, Theo said T3 Code uses Fable for implementation, GPT-6.2 Astra for playtesting, and Gemini 4 Neon exclusively for copywriting, with each model running through its native harness and subscription. The post does not identify which release prompted the comment.

With yet another unexpectedly good model drop, I'm so thankful to be in T3 Code. - Fable leading implementation - GPT-6.2 Astra as a play…
AI High Signal

During two years of developing Solveit, Jeremy Howard says he had not previously reverted a model to an older version; he is moving Astra and Sol back to their pre-6.2 versions because 6.2 is not working for him or Solveit users.

In the last 2 years of development of Solveit, I'm doing something I haven't done before: reverting a model to an older version. I'm swit…