Google Gemini 4 Launch Barrage: Four Models in Seven Weeks Top the Chat Rankings — and Lose the Agent War to Claude Fable 5.2
Google shipped four Gemini 4 models in seven weeks and swept the chat-preference top three, yet Claude Fable 5.2 laps it where real agent work is scored. The launch war's real ledger is cost per task and distribution, not points.
On November 18, Google released Gemini 4 Neon — its fourth Gemini 4 model in seven weeks. Argon landed on September 30, the flagship Carbon on October 28, the speed tier Hydrogen on November 12 and the mid-tier Neon on November 18, with the Genie 4 world model entering the Gemini API on November 19 in between. The independent verdicts split on the spot. On LMArena's text preference leaderboard, Carbon leads at 1526, with Argon and Neon right behind — Google swept the top three. But Artificial Analysis's Intelligence Index v5.0.1 ranks Fable 5.2 first at 65, GPT-6.2 Astra second at 62, Neon third at 61 and Carbon fourth at 60. And on Agent Arena, which scores real task outcomes, the gap becomes a cliff: Fable 5.2 scores 27.4, Neon just 18.4, Carbon 16.1.
This was not a launch in isolation; it was a three-way collision inside one week
This was not a launch in isolation; it was a three-way collision inside one week. Anthropic moved first on November 9, releasing Fable 5.2 with a 20% price cut to $8 per million input tokens and $40 output, retaking the top spot the day before Carbon opened to everyone. OpenAI followed on November 17 with GPT-6.2 Astra, also cut 20% to $8/$40. The deeper backdrop is share: ChatGPT's global visit share has fallen from 76.5% in May 2025 to 53.9% by mid-2026, while the Gemini app passed 950 million monthly users (Alphabet's Q2 earnings figure) and climbed to 27.9%. Google also holds what nobody else has: the default entry points of Search, Android and the app store. Anthropic, meanwhile, spent the same week taking the capital-markets headline with the largest IPO on record — priced at $83, raising about $81 billion — and closed its first Nasdaq session as ANT up 17% on November 19.
Google's play is not the highest score; it is the value frontier plus distribution. Neon lists at $3 per million input tokens and $15 output, and its cost per task is $6.44 — about a quarter of Fable 5.2's $26.54. But that math only works against Anthropic: Astra's list price is 2.7 times Neon's, yet its cost per task is just $4.00, because Google's models burn more tokens — Neon generated roughly 283 million tokens across the index evaluation, against a median of 65.1 million. The top thinking tier is a trap of its own: moving from high to xhigh costs 72% more for exactly one extra point. The real weapon is distribution. Hydrogen, at $1.50/$7.50, is the family's only model with a free API tier, and it is now the default model inside Search's AI Mode. Antigravity's Orchestrator lets a big model plan and review while a small model edits code and runs tests — Google claims 24% less quota and 19% faster on average. The unit of competition has shifted from a model to a system. The costs are on the record too: scores depend heavily on the harness — the same Neon manages 12.8% in mini-SWE but 39% inside Google's own Antigravity; developers measured Antigravity's five-hour quota melting down in 15 minutes; and the element-themed naming confused the community so much it became the loudest topic on Hacker News.
Layer one: the price anchor has moved
Layer one: the price anchor has moved. Flagship list prices have fallen from $10/$50 per million tokens to $8/$40, the contested variable is now cost per task rather than list price, and Carbon's December 1 rise from its $2/$10 promotional price back to $4/$20 is the biggest pricing event on the calendar. Layer two: the default mindshare in coding agents has not changed hands. Antigravity plus Neon matches Claude Code on Artificial Analysis's coding-agent index at a lower cost, but real paid traffic and developer habit still sit with Claude Code; the fight has moved to the orchestration and routing layer. Layer three: capital markets are not paying for models. On October 29 Alphabet fell 12.6% in a day on its third-quarter earnings — roughly $475 billion of market value gone, its worst single day since listing — and the model launches were completely drowned out; on Neon's launch day the stock rose nearly 4% pre-market and still closed down 3.46%. The company that actually converted the week's technical noise into money was Anthropic. Layer four: China's labs did not respond — they ran a parallel race. Qwen4-Max went GA on November 16 with cost per task down to $1.67; DeepSeek updated its low-price Flash on November 13; Doubao Work shipped 21 workbench scenarios on November 16 and had already abolished its five-hour quota cap — a pointed contrast with Antigravity's meltdown. The round's biggest hidden loser is xAI: Grok 4.7 scores just 34–35 on the Artificial Analysis index, more than 26 points behind the frontier, and as of November 19 Grok 4.8 exists only as Elon Musk's verbal preview — no API, no price card, no benchmarks.
iHolyshit review
These seven weeks proved one thing: release cadence plus default distribution can turn a preference leaderboard into home turf, but the leaderboard that scores real work does not play along — the agent gap is a systems-engineering gap, not a model gap. Two numbers will settle it: whether Neon's value story survives Carbon's December 1 price rise, and whether that nine-point Agent Arena gap narrows within six months. If it does not, no amount of launching will change what the week is remembered for: Fable 5.2 went public, and did the best work too.
November 20, 2026
Share this article
What do you think of this story?
Anonymous feedback, up to 5 reactions per article.
Reader comments
Comment without signing in, or sign in to display your member name.
Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.
No comments yet. Be the first to share.