Google DeepMind Ships Gemini 4 Neon With 1M Context, Claims Lower Fabrication — But Most Benchmarks Are Self-Reported
Google DeepMind released Gemini 4 Neon on November 18, the fourth model in its Gemini 4 family in seven weeks. The model offers a 1-million-token context window, four configurable thinking levels, and pricing of $3 per million input tokens ($0.15 cached) and $15 per million output tokens. Google says Neon achieves its lowest fabrication rate yet in internal testing, though the 1.5% figure cited in some coverage comes from internal measurements of its sibling Carbon, not Neon. Independent verification is thin: most benchmarks come from Google's own announcement, the safety report is not yet public, and no Neon-specific fabrication number has been published.
What shipped and where
Neon is available starting November 18 in the Gemini API (Google AI Studio, Android Studio), Gemini for Business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra users. A cybersecurity-hardened variant is rolling out to vetted defenders through the Fairwind Program [1][6]. The model accepts text, images, audio, video, and PDFs within a unified 1-million-token input-output window, with up to 64K default synchronous output tokens and up to 1M via background long decode [2]. Knowledge cutoff is March 2026 [2][3].
Google positions Neon between Hydrogen (fast, cheap, everyday execution) and Carbon (hardest, longest tasks). It is adapted from Carbon [2][4]. The family now spans four models released over the past seven weeks: Argon, Carbon, Hydrogen, and Neon [3].
Pricing and the thinking-level lever
List prices are $3/$15 per million input/output tokens, with a 95% cached-input discount ($0.15/M). Google says at the default "high" thinking level, Neon costs 32% less per task than Carbon and sits on the Pareto frontier with 50–80% lower cost per task than competing frontier models [1]. At "xhigh," it claims 36%+ savings versus rivals at their highest settings [1].
Independent analysis from Artificial Analysis, cited by i-scoop.eu, complicates the picture. Carbon carries an introductory price of $2/$10 until November 30, then rises to $4/$20. Until that date, Neon is 50% more expensive per token than Carbon; after December 1 it is 25% cheaper. Neon also writes roughly 15% more output tokens per task than Carbon, so the per-task savings come entirely from the lower list price, not token efficiency [3][9]. The four thinking levels (low, medium, high, xhigh) act as a budget lever: moving from medium to high doubles cost for five index points; high to xhigh adds one point for 72% more money and 45% more latency. Total cost per task varies by nearly a factor of nine across levels [3]. The API defaults to high, whereas Carbon defaulted to medium, so swapping model IDs without setting thinking_level explicitly increases spend [3].
Benchmarks: mostly Google-reported
Google's launch post and model card cite a suite of first-place or near-top scores [1][2][4][8]:
- Artificial Analysis Intelligence Index: 61 at xhigh (third overall behind GPT-6.2 Astra and Claude Fable 5.2) [9]
- Vals Index v2.2: 70.9% — Google's blog claims #2 on the full leaderboard [1], while the model card's published comparison table shows Neon ahead of the two models included there (Opus 5.5 at 70.1%, GPT-6.2 Sol at 66.1%) [2]
- AA Coding Agent Index: 58.6 (#1) [1][4]
- SWE-sweep provider harness: 39.0% (vs. 10.1% for Claude Opus 5.5) [1][2][4]
- GraphWalks 256K–1M: 95.2% (vs. 66.8% Opus 5.5, 80.1% GPT-6.2 Sol) [2][4]
- Cyber: 90.5% SEC-bench Pro, 70.7% CWE-bench v2, 51.3% WebExploitBench v1.2 [1][2][4]
- Science/math: 84.0% RiemannBench, 70.5% Terminal-Bench-Science, 71.3% HLE-Diamond [1][2][4]
Neon does not win everywhere. Opus 5.5 leads Terminal-Bench 4.1 (65.8% vs 65.2%) and FrontierCode 2.0 (47.3% vs 46.2%); GPT-6.2 Sol leads Chartography (75.1% vs 73.5%) [2][4]. Google's comparison table omits GPT-6.2 Astra and Claude Fable 5.2, which outrank Neon on the AA Index at their maximum settings [9].
Only a subset of these benchmarks — AA Index, Vals Index, some coding agent scores — have been independently measured by Artificial Analysis. The majority are Google's own evaluations. When Argon launched, Bloomberg reported internal Google doubts about whether strong scores held up in daily work [3]. The Neon safety and security report is marked "coming soon" in the model card [2].
Fabrication and safety claims
Google's blog post states Neon records "our lowest fabrication rate yet in internal Google Antigravity testing" [1]. Google's blog post cites a 0.09% misaligned action rate on an agentic safety suite, 99.2% pass rate on an agent honesty evaluation, 0.0% cheating on Terminal-Bench 4.0, 2.45% shortcut rate on a coding cheating evaluation (lowest of any Gemini 4 model), and 0.0% attack success on Gray Swan's indirect prompt injection benchmark at highest thinking level [1].
The 1.5% fabrication figure referenced in some coverage comes from a podcast interview with Melvin Johnson, Google DeepMind's post-training lead. Johnson said that on an internal hard evaluation, fabrication fell from 8% to "about 3% for Hydrogen and about 1.5% for Carbon" — not Neon — with a target of zero [10]. Neon is adapted from Carbon, so similar rates are plausible, but no Neon-specific fabrication number is published. Johnson also noted Carbon showed a 0% cheating rate on Terminal Bench 4 versus "a much higher" rate for prior models, and that these are Google's internal evals, not independently verified [10].
What's missing and what to watch
- Independent replication: Most benchmark scores are self-reported. The safety report is not yet public.
- Pricing transition: Carbon's introductory pricing ends November 30; fair per-task comparison requires waiting for list prices [3].
- Thinking-level defaults: API defaults to high; teams should set
thinking_levelexplicitly per workload to avoid unintended spend [3]. - No free API tier: Batch and flex modes halve prices but every test costs money [3].
- EU data residency: Neon appears on the Gemini Enterprise Agent Platform with EU multi-region endpoints; teams should verify before sending personal data [3].
- Model card gaps: Architecture, training data details, training data processing, hardware, and software are deferred to the Carbon model card [2].
Google's release cadence — four models in seven weeks — is notable [3]. The company attributes it to an internal feedback loop involving 100,000+ engineers using Antigravity daily, AI writing ~75% of new code, and a monorepo that generates hard tasks automatically [10]. Whether the pace is sustainable as "easy" tasks are exhausted, and whether fabrication and cheating rates hold on harder, independent benchmarks, remains to be seen.
Sources
- Introducing Gemini 4 Neon (blog.google)
- Gemini 4 Neon - Model Card (deepmind.google)
- Gemini 4 Neon matches Carbon for half the cost per task, and the thinking ... (i-scoop.eu)
- Google DeepMind Releases Gemini 4 Neon: 1M Context, #1 on AA Coding ... (marktechpost.com)
- Gemini 4 Hydrogen: 95% Cache Discount, 1M Context (shattered.io)
- Google Introduces Gemini 4 Neon with 1-Million-Token Context Window (af.net)
- Gemini 4 Carbon, Hydrogen, Decisions API, self-hosted agent ... (philschmid.de)
- Google Gemini 4 Neon launched with 1 million-token context window, four ... (moneycontrol.com)
- Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks ago (the-decoder.com)
- Google DeepMind's Melvin Johnson says Gemini 4 teams halved agent fabrication to 1.5%, now courting zero — BigGo Finance (finance.biggo.com)
- Google Rolls Out Gemini 4 Neon With One Million-Token Context Window (ndtvprofit.com)
- Gemini 4 Neon Model Card (storage.googleapis.com)