AdvertiseFee Deal Desk

Alphabet (Google)

Alphabet (Google) (GOOGL) Launches Gemini 4 Neon With Top 58.6 Coding Agent Score

Google launched Gemini 4 Neon on November 18 via the Gemini API and paid app tiers, matching Carbon's score in 14.8-minute tasks.

Be a creator
November 19, 2026, 11:18 AM UTC7 min read
AI SummaryAI
  • Google launched Gemini 4 Neon on November 18 via the Gemini API and paid Gemini app tiers.
  • Gemini 4 Neon scored 58.6, taking first place on the Coding Agent Index with Antigravity CLI.
  • Neon averaged 14.8 minutes per benchmark task against Carbon's 37.1 minutes at the high setting.
  • Prosecutors signaled a new count over Spagnuolo's prediction market bets on Gemini 3.0's release date.

Google Ships Gemini 4 Neon

Google began rolling out Gemini 4 Neon on November 18, a new AI model built to speed up multi-step software development, data analysis and scientific research. Access opened the same day through the Gemini API for developers, the Gemini app for paying subscribers and the company's Antigravity development environment.

Parent company Alphabet (GOOGL) places Neon between Gemini 4 Hydrogen, which is tuned for everyday speed, and Gemini 4 Carbon, its model for long, high-difficulty jobs; the DeepMind model card describes Neon as derived from Carbon. The model takes text, image, audio and video as input and handles up to roughly one million tokens combined across input and output. Standard synchronous API responses cap output at 65,536 tokens, while the background Long Decode mode raises output toward that combined limit.

Four thinking levels, low, medium, high and xhigh, adjust reasoning depth. Benchmarks carry the pitch: at the high setting on Artificial Analysis's Intelligence Index, Neon and Carbon both scored 60, but Neon averaged 14.8 minutes per task against Carbon's 37.1. At xhigh, Neon reached 61, third overall at launch, and paired with the Antigravity CLI it took first place on the Coding Agent Index at 58.6. Google also put Neon's cost per benchmark task 32% below Carbon's at the standard high setting, though the announcement does not state the Carbon comparison price.

Inside Antigravity, the Orchestrator feature can pair Neon as planner with Hydrogen as executor, leaving routine code fixes and tests to Hydrogen and returning to Neon's plans when work stalls. Pricing on the Gemini API paid tier is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15. Google AI Pro and Ultra subscribers can select Neon in the Gemini app and AI Studio, and enterprises through Gemini for business. A security-focused Gemini 4 Neon for cyber reaches Fairwind Program participants within days.

New Count Over Gemini 3.0 Bets

Federal prosecutors are preparing a new charge against former Google engineer Michele Spagnuolo, this one tied to prediction market bets on the release date of the Gemini 3.0 AI model. His lawyers disclosed the plan to a Manhattan federal court, citing a November 17 phone conversation with prosecutors who signaled they may supersede the indictment with a count covering those trades. The count differs from the charges already on file, which rest on Google search data. The distinction matters: search rankings are open to ordinary consumers, while a model's launch timing is routine knowledge inside AI labs, so a conviction on the new count would set a precedent for insider trading cases built on prediction markets. Investigators obtained a statement from Spagnuolo about where Google stored Gemini 3.0 rollout information and requested related records from Google a day later.

The U.S. Attorney's Office for the Southern District of New York charged the 36-year-old Italian citizen, who traded as “AlphaRaccoon,” in a criminal complaint unsealed after the case opened on May 27, 2026, alleging commodities fraud, wire fraud and money laundering. Prosecutors say he used an internal Google tool marked confidential and cleared about $1.2 million of the roughly $2.75 million he staked on Polymarket listings tied to his former employer between October 15 and December 4, 2025. U.S. Attorney Jay Clayton said he “violated the duties he owed to his employer” to book more than $1.2 million in trading profits. U.S. District Judge Analisa Torres suppressed part of his statements on October 23, a ruling the government is appealing to the Second Circuit. A hearing on other statements is set for December 21, 2026, and trial for June 1, 2027. Spagnuolo pleaded not guilty, was released on a $2.25 million bond and is on leave at Google. The Commodity Futures Trading Commission (CFTC) filed a parallel civil complaint. KPMG partner Conway Dodge says corporate questions about writing prediction markets into insider trading policies have at least doubled in six months.

Google Reworks the TPU Supply Chain

Google, Anthropic and other large US cloud and technology firms are moving to cut their reliance on NVIDIA (NVDA) GPUs, designing their own chips and buying high-bandwidth memory (HBM) directly to bring procurement costs down. A report carried by South Korea's Chosun Ilbo on November 18 laid out the shift and its consequences for the three big memory makers, Samsung, SK hynix and Micron, whose HBM pricing power it now tests. The spending backdrop is unusual: Morgan Stanley estimates cash capital expenditure at eight major operators, including Alphabet, Amazon, Microsoft and Meta, climbing from $946 billion this year to $1.656 trillion next year, roughly 75% growth, with 2028 growth slowing to about 7%. TrendForce expects DRAM and NAND Flash to rise from 47% to 68% of cloud capex over the same span.

Google's TPU program is the clearest case. Broadcom previously handled the TPU's design, packaging and HBM purchasing as one package. Google now splits the work, keeping compute chip design while MediaTek supplies the I/O die and Google buys HBM from memory makers itself. Bernstein estimates the structure saves about $50 billion across a TPU product lifecycle: on a hypothetical TPU v8t run of 6 million chips, roughly $8,434 per chip against $16,996 under the Broadcom model, with stacked margins on HBM accounting for about 76% of the gap. Anthropic's filing with the U.S. Securities and Exchange Commission (SEC) for its IPO discloses a parallel “Anthropic Direct” strategy of renting accelerators, CPUs and HBM directly while contracting data center space separately, with about $451.6 billion in compute and infrastructure commitments through the end of 2030. JPMorgan adds that one large model developer used its own AI to cut porting a model to AMD GPUs from six months to three days, lowering the cost of switching suppliers.

GOOGL Rallies 3.78% to $294

Shares have taken the week's run of product news well. COINOTAG data on GOOGL stock puts the last price at $293.74, up 3.78% over 24 hours, inside a $280.26 to $295.42 range. COINOTAG's composite scoring rates $282.71 the strongest support below the market at 89/100 and $306.20 the strongest resistance above it at 86/100. The daily RSI reads 47.7 and the daily trend still prints as a downtrend, so the move is working against the larger structure. A daily close above $306.20, or a loss of $282.71, would show which side gives way first.

GTIG Details Autonomous AI Attacks

Google's threat analysis unit, GTIG, has confirmed AI-driven hacking automation: chains of multiple AI agents that discover software vulnerabilities and carry an attack from code generation through execution without a human in the loop. The finding moves attacks that once demanded specialist skills and weeks of target research toward something a less technical operator can run at scale. North Korea-linked groups appear among the early adopters. They used AI-generated fake recruiters to approach employees at crypto companies, and in one documented pattern they altered the configuration files of Claude Code, an AI coding tool, so the victim's own development setup executed malicious code. GTIG also recorded “LLM jacking,” the theft of corporate AI service credentials to misuse computing resources; one case exceeded $100,000 in unauthorized charges in a single day and reached roughly $1 million in total. Google is deploying AI on the defensive side as well, automating vulnerability discovery, patching and attack detection. For crypto firms named in the report as contact targets, the practical shift is that the attacker's marginal cost keeps falling.

Primary sources

COINOTAG's editorial and research desk.

AI-Assisted

AI-generated, AI-reviewed, under COINOTAG editorial oversight.