Google DeepMind Moves Gemini 4 Into Post-Training as Frontier Gap Widens

โดย Yahoo Finance·US·Read original
Summary · why it matters

Google DeepMind has transitioned its next-generation model, Gemini 4, into the early post-training phase, a milestone confirmed by Koray Kavukcuoglu, elevated to Senior Vice President of Google DeepMind on August 12, 2026, during his first public appearance at The Information AI Agenda Live Summit on September 23-24, 2026. Kavukcuoglu said the intention is to roll out an early post-training version as soon as possible, with a target much earlier than the end of the year, though no specific release date was provided. The compressed timeline follows Google's announcement on July 21, 2026 of its most ambitious pre-training run for Gemini 4, while Gemini 3.5 Pro has remained stuck in partner testing since June 2026. The urgency reflects a capability gap: on the Intelligence Index v4.3.2, Google's current frontier model Gemini 3.6 Flash scores 34.0, trailing Anthropic's Opus 5.5 at 57.6, GPT-6 Astra at 52.7, and OpenAI's GPT-6 Sol at 47.5. The announcement came immediately after a wave between September 21 and September 22 in which the market absorbed Grok 4.7 at $2/$6, Opus 5.5 at $4/$20, and GPT-6 Sol and Luna at a permanent 50 percent price cut, collapsing the coordinated pricing structure that had held through the Amodei pacing framework. Google has disclosed no architecture, parameter count, pricing, or benchmarks for Gemini 4, saying only that it is significantly larger than prior models and focused on coding, autonomous agents, and long-horizon agentic workflows.

Impact on assets 1

Artificial Intelligence▲ · 1 stocks
Alphabet Inc Class C
GOOG
▲ PositiveTechnologyrelevance

Google DeepMind moved Gemini 4 into post-training with a target release much earlier than year-end, advancing Alphabet's frontier model.

Theme Impact 2

Off-coverage companies 3

DeepMindPrivate▲ Positive
Technologyrelevance

DeepMind's Gemini 4 entered the early post-training phase, a milestone confirmed by its newly elevated SVP Koray Kavukcuoglu.

OpenAIPrivate▼ Negative
Competitionrelevance

OpenAI's GPT-6 models are named as rivals Google is racing to catch, and the article frames Google's compressed timeline as a response to the widening frontier gap.

AnthropicPrivate± Mixed
Competitionrelevance

Anthropic's Opus 5.5 is cited as the frontier leader Google trails on the Intelligence Index, but the article reports no new Anthropic development.

Related news

8

JPMorgan Says Meta's Muse Could Be Biggest Consumer AI App Since ChatGPT

JPMorgan analysts say Meta Platforms' new Muse AI agent could become the most widely used consumer artificial intelligence app since OpenAI's ChatGPT, after the app hit No. 1 on Apple's U.S. App Store following its Sept. 8 launch. Analyst Doug Anmuth credited Muse's user-friendly interface, generous free tier and Meta's distribution across social platforms serving billions of users daily, and the bank reiterated its Overweight rating on Meta with a price target of $820. Muse reached about 2.5 million U.S. downloads through Sept. 19 with roughly 557,000 daily active users, according to Sensor Tower data cited by Bank of America, and on iOS alone its first 12 days produced 1.43 million downloads, edging out ChatGPT's 1.37 million over the same post-launch window. BofA Global Research said early reviews show the agent's ability to execute multi-step tasks and workflows with limited user intervention is resonating, but cautioned that revenue from ads, subscriptions or commissions within Muse is unlikely to become material before 2028. Muse's basic version is free, with subscription tiers priced at $20 and $100 a month for heavier use, and Meta shares have climbed nearly 20% since the launch.
META ▲
Artificial Intelligence ▲
Yahoo Finance·1hRead more →
▲

Meta's Muse AI App Tops 3.4M Downloads in US and Canada Since Launch

Meta's Muse AI app has gained more than 3.4M downloads in the U.S. and Canada since its launch on Sept. 8, according to Sensor Tower data, as personalized AI apps gain traction with consumers. Sensor Tower estimates Muse had 2.8M downloads globally in its first two weeks, compared with 3.7M for ChatGPT, 400,000 for Grok and 300,000 for Claude during their respective first two weeks. Muse reached No. 1 on the U.S. Apple App Store on Sept. 18 and also topped the U.S. Google Play Store the following day, and it remained No. 1 on both stores as of Thursday. Sensor Tower said Meta's ability to promote Muse across Facebook, Instagram, and WhatsApp may have helped drive its early growth, and that Muse accounted for the largest share of Meta's in-house advertising impressions within 10 days of its launch, ahead of those platforms. Elon Musk's Grok Bot has also continued to grow despite Muse's launch, recording 2.2M global downloads since its August launch and averaging about 300,000 daily active users as of Sept. 22, with daily active users up 43% in the week ended Sept. 21 and downloads and daily active users up 134% and 164%, respectively, in the two weeks through Sept. 21 from the previous two-week period. Sensor Tower said Grok Bot has generated more than $1M in subscription revenue, suggesting some users are willing to pay for AI services that offer more personalized features.
METASPCX
Artificial Intelligence ▲
Seeking Alpha·1hRead more →
4impact 4

DeepSeek Revenue Run Rate Tops $1 Billion Ahead of $7.45 Billion Funding Round

DeepSeek's annualized revenue run rate has topped $1 billion, more than doubling in recent months as the Chinese AI startup converts surging model adoption into stronger monetization. Chief Executive Liang Wenfeng disclosed the figure during a recent investor meeting, according to The Information, which Reuters said it could not independently verify. The growth was partly driven by substantial price increases, with DeepSeek reportedly raising model pricing by 2.3 to 4.5 times. DeepSeek is now targeting a second funding round of 50 billion yuan, or about $7.45 billion, at a 500 billion yuan valuation by the end of October, and is also preparing for a potential listing on Shanghai's STAR Market. The company is still prioritizing model development over near-term inference revenue, with more than 70% of its computing capacity used to train new models while less than 30% supports existing-model inference, and earlier this month it launched V4.1-Flash, a new model built for faster inference, higher throughput and lower infrastructure requirements.
Semiconductor…MSFT
Artificial Intelligence ▲
GuruFocus·2hRead more →
© 2026 MarketDXNews