Google launched Gemini 4 Neon on November 18, 2026: a $3/$15 per million token model that ranks third on independent leaderboards and tests how real Google’s AI comeback is. It is the fourth Gemini 4 model in seven weeks, and it arrived one year to the day after Gemini 3 Pro, the release that first had analysts saying Google was back. It also shipped on the day Anthropic priced its IPO, the rival whose Claude Fable 5.2 still tops the main independent leaderboards. Read with the independent numbers beside it, Google’s announcement describes a real comeback of a specific kind: the machine is back, the crown and the customers are not yet.
TLDR
Gemini 4 Neon is Google’s new middle model, priced at $3 per million input tokens and $15 per million output tokens with a 1 million token context window. As of November 19, it is third on the Artificial Analysis Intelligence Index behind Claude Fable 5.2 and GPT-6.2 Astra, and second on the Vals Index, at well under half the cost per task of Anthropic’s top models. Google’s comeback is cost, speed and release cadence, built on its own chips, its own coding harness, a Cloud business growing 103% and a billion-user app. What is still unproven is the part that pays: enterprise spend, independent verification of most launch numbers, and whether the capex bill turns into free cash flow.
Thoughts
Google’s comeback is a cost comeback, not a return to first place, and Neon’s real pitch is cost per task: thinking levels you can dial down and a spot on the price and performance frontier. Artificial Analysis measured what that buys. As of November 19 its Intelligence Index has Claude Fable 5.2 at 65, GPT-6.2 Astra at 62 and Neon at 61, and running the index cost $6.44 on Neon against $10.04 on Claude Opus 5.5 and $26.54 on Fable 5.2. Google’s comparison table names only Opus 5.5 and GPT-6.2 Sol, leaving out the two models above Neon. The edge also holds against Anthropic, not OpenAI: Astra ran the same index for $4.00, and Sol for $1.01. Nor is this a price cut. Google’s frontier list price has risen each generation, from $1.25/$10 for Gemini 2.5 Pro to $2/$12 for Gemini 3 Pro to $4/$20 for Argon and Carbon from December 1, so the “efficient” middle model costs more per token than last year’s flagship. What Google sells now is capability per dollar, during a compute shortage in which it can charge for it.
Google’s best AI business is partly Anthropic’s success, and Neon is the attempt to win the customer directly. Google Cloud revenue reached $30.7 billion last quarter, up 103%, and TPU system sales (Google’s own AI chips, now sold for customers’ data centers) were described on the call as a meaningful contributor, with Anthropic among the named buyers. Anthropic has committed to up to a million TPUs and to “multiple gigawatts” more from 2027, though Amazon remains its primary training partner. Reports put Alphabet’s Anthropic stake at about $149 billion after the IPO pricing. Meanwhile Menlo Ventures, itself an Anthropic investor, estimates Google’s share of enterprise LLM API spend at 9%, down from 21%, against 55% for Anthropic. Google’s own usage data points the other way (nearly 32 billion tokens per minute through its APIs, more than four times a year ago), and the hedge pays either way. Neon, with its enterprise availability and $0.15 cached input price, is aimed at the spend Google is losing.
The most revealing number in the launch is about software, not the model. On Google’s SWE-sweep test, Neon scored 12.8% in a minimal harness (the software that wraps a model so it can read files, run tools and work in steps) and 39.0% inside Google Antigravity: the same weights, triple the score. Opus 5.5 went from 3.6% to 10.1% in Claude Code, and GPT-6.2 Sol fell from 11.6% to 8.3% in Codex. That is Google arguing the product around the model is now its edge, which is what the roughly $2.4 billion Windsurf licence and hiring deal of July 2025 was for. But Google computed these numbers itself, the test’s task set is not publicly described, and the result does not travel to anyone else’s tools. Adoption is where the comeback is thinnest: developer surveys suggest Antigravity is used by a single-digit share of developers, far behind Claude Code.
Investors paid for the comeback early, then got the bill. Alphabet crossed $4 trillion in January and closed at $4.8 trillion on May 8, up 160% in twelve months. By November 19 it was about 31% below its record, near $3.47 trillion by Blockonomi’s count. In between, third-quarter free cash flow went negative $9.1 billion, the second negative quarter running, on $51.4 billion of quarterly capex, with 2026 guided to $195 to $205 billion and 2027 expected to increase significantly again. Neon fits that squeeze: the launch post sells it in the Gemini app as a way to stretch usage limits further. Nothing shows Neon moved the shares; that week belonged to Anthropic’s IPO. The other side is real too: Cloud operating margin reached 36.2%, and as Pichai put it in his Big Technology interview, “Everyone is short.”
Speed has a safety cost, and it sits in the back half of the announcement. The model card says Neon is adapted from Carbon, that frontier safety results were carried forward from Argon because Neon’s capability gain fell below the threshold for new critical capability evaluations, that critical levels for cyber and biological uplift “cannot be ruled out”, and that the full Safety and Security Report is still to come. Redwood Research argued on November 16 that the Argon and Carbon reports rest on monitoring rather than demonstrated alignment. In fairness, Google’s reported numbers are strong, vetted cyber access is what Anthropic and OpenAI do too, and Redwood wrote before Neon existed and did not call for blocking release. The concern is narrower: four models in seven weeks means safety paperwork is being inherited rather than redone.
Key Takeaways
- Gemini 4 Neon costs $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens, with a 1 million token context window. It is paid only.
- As of November 19, Artificial Analysis ranks Neon third among models on its Intelligence Index at 61, behind Claude Fable 5.2 (65) and GPT-6.2 Astra (62), one point ahead of Gemini 4 Carbon.
- On the Vals Index, Neon is second at 70.87%, in a statistical tie with the models around it, for $7.40 per test against $17.94 for Fable 5.2.
- Google’s post says Neon debuted second on Arena’s Agent Arena Index at 1,336. Arena’s board, updated November 18, showed Neon third at 1,333, three points behind Opus 5.5, at $2.58 per task against $6.04.
- The cost advantage holds against Anthropic, not OpenAI: GPT-6.2 Sol ran the Artificial Analysis index for $1.01 and GPT-6.2 Astra for $4.00, against Neon’s $6.44.
- Google went from 20th on Agent Arena in August (Gemini 3.7 Flash) to third and fourth in November (Neon and Carbon).
- Carbon took first place on Artificial Analysis on October 28 and is fourth today, so Google held the top spot for under a month.
- The same Neon weights scored 12.8% in a minimal harness and 39.0% in Google Antigravity on Google’s SWE-sweep test, a result Google computed itself.
- Google’s frontier list price has risen each generation, so Neon is efficient per task, not cheap per token.
- Google Cloud grew 103% to $30.7 billion in the third quarter, helped by TPU sales to customers including Anthropic.
- By Menlo Ventures’ estimate (Menlo is an Anthropic investor), Google’s share of enterprise LLM API spend fell to 9% from 21%.
- Alphabet’s free cash flow was negative $9.1 billion in the third quarter, its second negative quarter in a row, on about $200 billion of 2026 capex.
- Neon’s model card carries frontier safety results forward from Argon and says critical cyber and biological uplift “cannot be ruled out”. The full safety report is pending.
Detailed Summary
What Gemini 4 Neon Is: Price, Context Window and Availability
Neon was announced by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect. It costs $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens (a 95% discount), with a 1 million token context window and a March 2026 knowledge cutoff. It has four thinking levels with high as the default. It is rolling out in the Gemini API paid tier, Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers. There is no free tier.
Where Neon Sits in the Gemini 4 Family
The Gemini 4 line now has four models named after elements. Argon, announced September 30, was the first. Gemini 4 Carbon is the strongest, sent to vetted cyber defenders on October 28 and generally available November 10, at $2/$10 until November 30 and $4/$20 after. Hydrogen, out November 12 at $1.50/$7.50, handles fast everyday work. Neon sits between Hydrogen and Carbon, and according to 9to5Google it replaces Argon as “Deeper analysis” in the Gemini app’s model picker. Google’s Logan Kilpatrick conceded to VentureBeat that the company had “a little too much fun with naming.”
Gemini 4 Neon Benchmarks: What Independent Leaderboards Show
Independent boards back a strong third place. Artificial Analysis has Neon at 61, and Vals has it second at 70.87%, two hundredths of a point ahead of Claude Sonnet 5.5. Zapier’s AutomationBench ranks Neon first at 56.3%, with Gemini models in nine of the top ten places. Andon Labs’ Vending-Bench 2 also ranks it first at $16,843, though Carbon is within the error bar and the newest Claude and GPT models are not on that board. In Google’s own table, Neon trails Opus 5.5 on Terminal-Bench 4.1 and FrontierCode 2.0, and GPT-6.2 Sol on Chartography.
Gemini 4 Neon Pricing vs Claude Opus 5.5 and GPT-6.2 Sol
By list price, Neon is 25% below Opus 5.5 ($4/$20) and 50% above Claude Sonnet 5.5 and GPT-6.2 Sol ($2/$10), per VentureBeat. Google claims a 32% lower cost per task than Carbon at the default high setting; Artificial Analysis measured about 17% at xhigh and said the saving came from pricing, since Neon used about 15% more output tokens. Until Carbon’s introductory price ends November 30, Neon is the dearer of the two per token. Neon is also verbose: 283 million output tokens on the Artificial Analysis run against a 65.1 million median, at 21.5 minutes per task on xhigh.
Antigravity, the Orchestrator and the SWE-sweep Result
Google positions Neon as a standalone agent and as an Orchestrator directing parallel Hydrogen subagents in Antigravity, the coding tool it launched on November 18, 2025. The SWE-sweep jump from 12.8% to 39.0% inside Antigravity is the post’s biggest coding claim. Two of the four demos show the scale Google is aiming at: a first-person ski simulator on real Mont Blanc terrain with 10,200 lines of custom shaders, and a 100,000-cell Google Sheet that relights a photo using over 500,000 live formulas. Developer Flavio Copes found Neon right on every test but slower and pricier than Opus 5.5 on a short task, and about 42% cheaper on a long-context one.
Cyber Defense and the Fairwind Program
Google reports 76% on the hard tier of AutoBaxBench, 90.5% on SEC-bench Pro, 51.26% on WebExploitBench v1.2 and 70.7% on CWE-bench v2, all Google-reported. A version called Neon for cyber goes to vetted members of the Fairwind Program over the coming days. Fairwind is Google DeepMind’s vetted-access route for defenders, with priority for governments, critical infrastructure and core platforms. Neon is the third Gemini 4 model whose cyber-capable version has gone through it first.
Safety: What the Model Card Says
The launch post reports a 0.09% misaligned action rate, a 99.2% pass rate on Google’s agent honesty evaluation, 0.0% cheating on Terminal-Bench 4.0, and 0.0% attack success on Gray Swan’s prompt injection benchmark with real-time safeguards at the highest thinking level. The model card adds that Neon’s frontier safety results are carried forward from Argon after a capability assessment fell below the threshold for new tests, and that cyber and biological critical levels “cannot be ruled out.” The full Safety and Security Report is due in the coming days.
From Code Red to Gemini 4: The Comeback Timeline
In December 2022, The New York Times reported that ChatGPT had set off a “code red” at Google. In February 2023, a Bard demo error helped wipe about $100 billion off Alphabet in a day. Gemini 3 Pro, on November 18, 2025, brought the first “Google is back” verdicts. Then came about ten months without a new flagship while Anthropic and OpenAI traded releases, until Argon on September 30, 2026, when Artificial Analysis said Google was “back among the top three labs.” This is Google’s second comeback in a year, and the first one faded.
The Business Behind the Comeback: Cloud, TPUs and Capex
Third-quarter Cloud revenue was $30.7 billion, up 103%, with operating margin at 36.2% and a backlog of about $561 billion. The Gemini app passed one billion monthly users (ChatGPT’s billion is weekly, so the two are not like for like). Against that sit negative free cash flow, about $200 billion of capex this year, Search growth slowing from 19% in the first quarter to 12% in the third, and a 9% enterprise API share by Menlo’s estimate. Three things would settle the comeback question: enterprise spend share, independent replication of the launch numbers, and free cash flow turning positive.
Notable Quotes
“Neon delivers frontier-level reasoning for complex, high-frequency workflows faster and more cost efficiently.”
Koray Kavukcuoglu, Google DeepMind, in the launch post
“On SWE-sweep, it already leads on the mini-SWE harness (12.8%) and more than triples to a state-of-the-art 39.0% in Google Antigravity showing what the Antigravity harness unlocks.”
Koray Kavukcuoglu, on the harness result
“While Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, Neon balances depth and efficiency.”
Koray Kavukcuoglu, on where Neon sits in the Gemini 4 family
“It’s cheaper, faster, and better behaved than Carbon.”
Varun Mohan of Google, quoted by VentureBeat
“On short tasks it thinks a lot, and that changes the math.”
Flavio Copes, developer, after testing Neon against Opus 5.5 and Carbon
“Google Cloud revenues grew 103%, powered by huge demand for our AI infrastructure and AI solutions.”
Sundar Pichai, on Alphabet’s third-quarter 2026 results
Read the full announcement here.
Related Reading
- Gemini 4 Neon on Artificial Analysis for independent scores, cost per task and speed.
- The Vals Index, which lists cost and time per test beside accuracy.
- Zapier AutomationBench, where Gemini models hold nine of the top ten places on business automation.
- Redwood Research on the Gemini 4 safety reports, the main outside critique of Google’s safety case.
- Gemini API pricing, the full price list for every Gemini model.