Google says Gemini 4 Neon is #3 on Artificial Analysis and #2 on Vals; the boards agree, Fable 5.2 is first
Google's launch post of 18 November 2026 says Gemini 4 Neon is #3 on Artificial Analysis and #2 on Vals. The saved pages agree: third counting each model once, second at 70.87%, and second on CWE-bench at 70.7%, each behind Claude Fable 5.2, which the launch table leaves out.
By Parminder Kumar Sharma · LinkedIn · · 23 min read
Artificial Analysis prints Neon as 6th of 96; counted once per model it is third, as Google says (derived)
Google's launch post, dated 18 November 2026, says Gemini 4 Neon "ranks ahead of comparable models and #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level". Artificial Analysis's page for Neon prints the 61, labels it Xhigh, and prints a rank beside it: "# 6 / 96". That looks like a contradiction. It is not one. The page's hover list under that rank, which this site read in the saved HTML, names ten entries, each a model at an effort level: Claude Fable 5.2 at Max, Xhigh and High, GPT-6.2 Astra at Max and Xhigh, and then Gemini 4 Neon (Xhigh) sixth. So the 96 counts entries, a model once for each effort level it was run at. Count each model once and the order is Claude Fable 5.2, GPT-6.2 Astra, Gemini 4 Neon: third, which is Google's "#3 overall" (derived from the hover list). The data behind the page's charts agrees: Fable 5.2 at 64.56, Astra at 62.14, Neon at 61.40.
The other two boards agree as well. Vals lists Neon second on its Vals Index at 70.87%, behind Claude Fable 5.2 at 71.31%, with Claude Sonnet 5.5 third at 70.85%, and Google's own table prints that as 70.9%. CWE-bench v2 lists Neon second at 70.7% on all 250 of its tasks, behind Fable 5.2 at 75.0%. An independent second place on two boards, and third on a third, is a real result. On all three pages the model ahead of Neon is the same one, and it is not in Google's launch table.
What this does not establish. The launch table compares Neon with Claude Opus 5.5 and GPT-6.2 Sol only. The independent pages put those two lower than Neon: Opus 5.5 is fifth and Sol tenth on Vals, fifth and eighth on CWE-bench, and eighth and eleventh on Artificial Analysis's entry ranks as read on 17 November. Neither Google's page nor its methodology says why those two were chosen. Their list prices bracket Neon's ($4 and $20, and $2 and $10, against $3 and $15 per million tokens on Artificial Analysis's pages), which is one plausible reason, and it is this site's observation, not Google's statement. Most of the table's other figures are not on any saved independent page: by the methodology page's own wording, 8 of the 18 benchmarks are self-computed or self-evaluated by Google (derived count). And the cyber numbers were measured with safety filters off, per the same page.
Whose account this is. Google made Neon, sells it, and ran or chose most of the numbers in its table; it also cites independent boards that favour it. Artificial Analysis, Vals and CWE-bench are the independent side, described only as their own pages describe them: Vals says its suite is "benchmarks that no other party has full access to"; CWE-bench says it is "built and verified by Collinear AI's research and engineering team", and Collinear also offers a separate training corpus. This brief does not rank Google above or below any lab, and it is not investment advice.
The launch table: 18 benchmarks, two rivals, three rows where a rival is ahead
Google's post carries one table, and the model card prints the same table under "Results", "as of November, 2026". It compares Gemini 4 Neon with Claude Opus 5.5 and GPT-6.2 Sol on 18 benchmarks in eight groups; SWE-sweep has two lines, so there are 19. This site viewed the image and typed the figures below from it, and the model card's page 5 shows the same values.
The same table as text. In the image a blue cell marks the highest printed value in its row and a grey cell a rival's higher value; the image has no legend, so that is this site's reading of the pattern. (ahead) marks a rival's higher value. Not reported is a dash in the image.
| Group: benchmark, qualifier | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|
| Capability indices: AA Index v5.0.1 | 61 | 59 | 57 |
| Capability indices: Vals Index v2.2 | 70.9% | 70.1% | 66.1% |
| Agentic coding: CursorBench 4.1, Score | 65.4% | 58.1% | not reported |
| Agentic coding: Terminal-bench 4.1, Cheating adjusted accuracy | 65.2% | 65.8% (ahead) | 61.5% |
| Agentic coding: FrontierCode 2.0, Main score | 46.2% | 47.3% (ahead) | 43.8% |
| Agentic coding: SWE-sweep, Provider harness | 39.0% | 10.1% | 8.3% |
| Agentic coding: SWE-sweep, Mini-SWE harness | 12.8% | 3.6% | 11.6% |
| Knowledge work: AutomationBench | 56.3% | 42.5% | 41.3% |
| Knowledge work: Agents' Last Exam v1.1 | 75.8% | 71.6% | 73.0% |
| Science and math: RiemannBench | 84.0% | 80.0% | 80.0% |
| Science and math: Terminal-Bench-Science 0.1 | 70.5% | 63.3% | 68.0% |
| Reasoning: HLE-Diamond | 71.3% | 67.9% | 64.0% |
| Multimodal understanding: MKA, Climb WSR | 56.5% | 53.0% | 52.7% |
| Multimodal understanding: Chartography | 73.5% | 66.3% | 75.1% (ahead) |
| Multimodal understanding: LVBench | 89.0% | 83.7% | 87.6% |
| Long context: GraphWalks, 256k to 1M, BFS (F1) | 95.2% | 66.8% | 80.1% |
| Cybersecurity: CWE-Bench v2 | 70.7% | 67.1% | 66.0% |
| Cybersecurity: SEC-bench Pro | 90.5% | not reported | 88.3% |
| Cybersecurity: WebExploitBench v1.2 | 51.3% | 37.1% | not reported |
Counts, derived from the image: Neon holds the highest printed value on 16 of the 19 lines, a rival on 3, and three of the 38 rival cells are dashes (GPT-6.2 Sol on CursorBench 4.1 and WebExploitBench v1.2, Claude Opus 5.5 on SEC-bench Pro). The three rows with a rival ahead are Terminal-bench 4.1 (Opus 5.5 by 0.6 points), FrontierCode 2.0 (Opus 5.5 by 1.1) and Chartography (GPT-6.2 Sol by 1.6). The post quotes Neon's own number on exactly those three, "65.2% on Terminal-Bench 4.1 and 46.2% (at its default high setting) on FrontierCode 2.0" and "73.5% on Chartography", without the rival's figure, which the table prints beside them.
Two cautions about reading the table. First, on SWE-sweep the table's "Provider harness" line runs each model in its maker's own tool (the methodology page names the Antigravity-CLI harness for Neon, Claude Code for Opus 5.5 and Codex for Sol 6.2), and the post says Neon "more than triples to a state-of-the-art 39.0% in Google Antigravity" (3.05 times, derived: 39.0 / 12.8) against its 12.8% on the shared mini-SWE harness. On that shared harness GPT-6.2 Sol scores 11.6%, 1.2 points behind Neon, and on its own provider harness 8.3%. Second, the methodology page says LVBench frames were sampled differently for each model, at "1FPS (low resolution) for Gemini, 500 frames for GPT-6.2 Sol, and 600 frames for Opus 5.5 due to API limitations", so that row compares three different inputs.
Each rank claim against the saved pages
One diagram sets Google's rank claims beside what the saved independent pages print. Three of the seven have a matching result on a saved page, and the four without one are not shown to be wrong: the pages were not saved.
Vals. The post says Neon "lands on the Pareto frontier", and the saved Vals page's chart text names six models, Neon among them, under a Frontier toggle; the text does not say what the labels mark. The gap between Neon and Sonnet 5.5 is 0.02 points (derived: 70.87 - 70.85) and the gap up to Fable 5.2 is 0.44 (derived: 71.31 - 70.87). The page says its error bars show "±1 standard error"; the bar values are in a chart this site could not read as text, so the saved page does not say whether a 0.02 gap is outside them. The page prints no effort setting for Neon, while the post says the Pareto claim is "at high".
CWE-bench. Google's "3.4% uplift over Carbon" is 3.4 percentage points (derived: 70.7 - 67.3). The 70.7% is on the page's combined 250 tasks, 100 public and 150 held out, at "high reasoning", not on the 100 public tasks alone. Fable 5.2 is 4.3 points higher at 75.0% (derived). The methodology page says ranking is by pass@1 with ties broken by pass@4, and on pass@4 Neon (76.4%) is below Claude Opus 5.5 (79.2%), Muse Wonder (78.8%), Hy4 (78.4%), Gemini 4 Hydrogen (78.0%) and Gemini 4 Argon (77.2%) as well as Fable 5.2 (82.8%). Google's post claims no rank on this board.
Not on a saved page. The Agent Arena Index (1336 across "4,200+ real-world agent sessions"), the Artificial Analysis Coding Agent Index (58.6), Zapier's AutomationBench (56.3%), Andon Labs' Vending-Bench 2 ($16,843) and the two Vals sub-benchmarks the post names are all Google's claims here, with a link in the post and no saved page behind them.
Cost per task: both of Google's claims hold on Artificial Analysis's data, at different effort levels
Google's post makes two cost claims in one paragraph. "At its default high thinking level" Neon pairs a high Artificial Analysis score "with a 32% lower cost per task than Carbon"; and at xhigh it is "still costing 36%+ less per task than competing models at their highest setting". Artificial Analysis's headline for Neon, $6.44 a task at Xhigh, is above Carbon's $5.48 on its page, which looks like a contradiction of the first claim and is not one: the data behind the Neon page's own cost chart prints Neon at every effort level. At high it is $3.74, which is 31.8% lower than Carbon's $5.48 (derived: 1 - 3.74 / 5.48 = 0.3175), and both print as an index of 60. At xhigh Neon is 17.5% above Carbon (derived: 6.44 / 5.48 = 1.175). The second claim holds against Anthropic's models on that page: 35.9% below Claude Opus 5.5 at max ($10.04, derived) and 75.7% below Fable 5.2 ($26.54). It does not hold against OpenAI's: GPT-6.2 Astra at max is $4.00, so Neon xhigh costs 61.0% more per task, and GPT-6.2 Sol at $1.01 costs 6.38 times less. The post does not say which competing models it means.
Artificial Analysis, as read on 18 November 2026 about 22:47 UTC from the Neon page's own chart data, except Carbon, read from its own page on 17 November about 19:30 UTC. Each model at the effort setting its page lists first; Neon at all four. Index v5.0.1, whole numbers as displayed. Cost per task is dollars per Intelligence Index task.
| Model and setting | Index | Cost per task | Minutes per task |
|---|---|---|---|
| Gemini 4 Neon, low | 44 | $0.75 | 3.7 |
| Gemini 4 Neon, medium | 55 | $1.90 | 8.3 |
| Gemini 4 Neon, high | 60 | $3.74 | 14.8 |
| Gemini 4 Neon, xhigh | 61 | $6.44 | 21.5 |
| Gemini 4 Carbon, high (17 Nov page) | 60 | $5.48 | 33.4 |
| GPT-6.2 Astra, max | 62 | $4.00 | 13.7 |
| GPT-6.2 Sol, max | 57 | $1.01 | 8.5 |
| Claude Opus 5.5, max | 59 | $10.04 | 24.6 |
| Claude Fable 5.2, max | 65 | $26.54 | 41.2 |
The effort levels are steep. From high to xhigh the page's full precision index rises 1.8 points (59.6 to 61.4) while cost per task rises 72.2% ($3.74 to $6.44) and output tokens on the index rise from 123M to 283M, 2.3 times (derived). Google calls xhigh the setting for "the hardest engineering, analytical, and scientific workloads"; the table above says what the last index point costs.
Vals prints cost on its own scale, "Cost / Test", and Neon's is $7.40 against $17.94 for Fable 5.2, $16.18 for Sonnet 5.5 and $24.77 for Opus 5.5: 58.8%, 54.3% and 70.1% lower (derived). That fits the post's "roughly 50% to 80% lower cost per task than top competing frontier models" for those three. Neon's cost is 30.7% below GPT-6.2 Astra's $10.68, and GPT-6.2 Sol's $3.18 is lower than Neon's by a factor of 2.33 (derived). Vals does not print Neon's effort setting on the saved page.
Speed. The post says "fast end-to-end task completion" and a quoted Glean customer reports "3x the speed and 60% of the cost" of Carbon in early evaluations. Artificial Analysis prints, for Neon at Xhigh, 21.5 minutes per index task, 56.51 seconds to the first answer token and 110 tokens per second, and calls it "notably slow per task and very verbose" against a median of 12.4 minutes. At high Neon takes 14.8 minutes against Carbon's 33.4, 2.26 times as fast (derived, pages read a day apart); at xhigh it is 1.55 times. GPT-6.2 Astra at max takes 13.7 minutes and Sol 8.5, so at xhigh Neon is slower than both. Artificial Analysis also lists Neon's 283M output tokens at Xhigh against a median of 65.1M. The Glean figure is a customer's own test, with no effort setting printed.
The cyber figures: one on an independent board, the rest Google's, measured with safety filters off
The post's section "Built for fast, efficient, frontier-level cyber defense" gives five figures: AutoBaxBench's hard tier at 76% ("state-of-the-art"), WebExploitBench v1.2 at 51.26% ("state-of-the-art ... at half the cost of competing large models"), SEC-bench Pro at 90.5% ("roughly half as much per challenge"), ExploitGym at "frontier-level performance using half the tokens of Carbon", and CWE-bench v2 at 70.7%. The methodology page says who ran each, and the sentence that matters comes first: "All results are with safety filters on with the exception of the cybersecurity evaluations."
CWE-bench v2 is the one independent number. The board is Collinear AI's, "built and verified" by its team, with 100 public and 150 held out tasks across 185 weakness types, graded by deterministic verifiers, four rollouts per task at high reasoning. The tasks are audit and patch: "agents are never asked to create exploits". Neon is second at 70.7% and $6.52 per rollout. Fable 5.2 is first at 75.0% and $7.85, so Neon is 16.9% cheaper per rollout for 4.3 points less (derived). Claude Opus 5.5 scores 67.1% at $1.85 and GPT-6.2 Sol 66.0% at $0.81, so Neon costs 3.52 and 8.05 times as much per rollout (derived). The page's own sentence about its Pareto frontier: "196× the cost buys 1.30× the pass rate". The methodology page adds that the leaderboard was "independently evaluated with Artificial Analysis"; the saved CWE-bench page does not say who ran Neon's rollouts, and for version 1 it lists "Collinear and Artificial Analysis" as the ones who could run it.
The others are Google's. SEC-bench Pro is "self computed" on one slice, the "183-task May 2026 JavaScript engine split", averaged over three attempts in Google's Antigravity harness; a task counts when a proof of concept crashes the vulnerable version of the software but not the patched version, or when an LLM judge confirms the crash is the intended flaw. The table shows no Opus 5.5 figure, and the page takes competitor numbers "from their published system cards and technical reports". WebExploitBench v1.2 is "self-evaluated and validated by AgentCyberRange", with 15 multi-container web application targets and 277 vulnerabilities, Opus 5.5 at 37.1% and no GPT-6.2 Sol figure. AutoBaxBench is "as provided by the authors of the benchmark", and ExploitGym is self-computed with mitigations off and a 10 hour limit; the saved pages print no ExploitGym number. The "half the cost", "half as much" and "half the tokens" claims have no cost or token figure on any saved page. The post's customer quotes are Google's selection: Thoughtful Lab says that in its cyber evals Neon "was able to find several subtle vulnerabilities other models missed while using a fraction of the tokens", with no counts or settings.
Microsoft's page for MAI Cyber 2 Flash, covered in brief 732, also reports a WebExploitBench figure, 40.4% against 38.4% for GPT-5.6 Cyber. That page prints no version and no task count for it, while Google's methodology prints v1.2, 15 targets and 277 vulnerabilities, so this brief does not put 40.4% beside 51.3% or rank the two.
Which model was measured. The post says "Gemini 4 Neon for cyber is rolling out to trusted defenders in the coming days as part of the Fairwind Program", and that eligible members "will begin receiving access over the next few days" through a Cloud account team. The model card never uses the phrase "Neon for cyber", never mentions the Fairwind Program, and lists cyber defensive tasks among Neon's intended uses. Whether "Neon for cyber" is a separate model, a mode, or the same model with different safeguards is not stated on either page, and neither is whether the cyber figures were measured on it. The post separately says Neon ships with "strong real-time safeguards against cyber and biosecurity misuse comparable to our Carbon protections". The methodology page says the cyber figures were measured without safety filters, and the post says the shipped model has real-time safeguards. Brief 731 records the same arrangement in OpenAI's GPT-6.2 card.
Where each number comes from
The table below goes line by line through the methodology page. "Self-computed" is the page's word. Of Neon's 18 benchmarks, three are on a saved independent page (the two indexes and CWE-bench v2), ten are sourced from a named leaderboard or party, and eight are self-computed or self-evaluated by Google (derived count: SWE-sweep, Agents' Last Exam, Terminal-Bench-Science, Chartography, LVBench, GraphWalks, SEC-bench Pro and WebExploitBench; the ten include the three). Non-Gemini figures are, in the page's words, "providers' self-reported numbers unless otherwise mentioned". The page also says the results were run "with the highest thinking settings unless indicated otherwise".
Where each figure in Google's post and table comes from, per Google's methodology page (a six page PDF at the address the table image prints), and whether a saved independent page confirms it. Read on 18 November 2026.
| Figure, as Google prints it | Who ran it, and the setting (methodology page) | Saved independent page? |
|---|---|---|
| AA Index v5.0.1: 61 (post: "#3") | Artificial Analysis. Neon at xhigh | Yes. Neon page: 61, Xhigh, rank 6 of 96 (entries) |
| Vals Index v2.2: 70.9% (post: "#2") | Vals AI. Setting not printed on the saved page; the post says Pareto claim "at high" | Yes. Second at 70.87% |
| CWE-Bench v2: 70.7% | Collinear AI's board, "independently evaluated with Artificial Analysis" per Google. Page: four rollouts at high reasoning, 250 tasks | Yes. Second at 70.7%, $6.52 a rollout |
| Terminal-bench 4.1: 65.2% | Neon's figure sourced from Artificial Analysis using the public Antigravity-CLI harness 1.5.9; others from the official public leaderboard; cheating adjusted | No |
| FrontierCode 2.0: 46.2% | Official public leaderboard; Neon at "high thinking" | No |
| SWE-sweep: 39.0% and 12.8% | Self-computed by Google. Provider harnesses (Antigravity-CLI, Claude Code, Codex) and a mini-swe-agent vision harness | No |
| CursorBench 4.1: 65.4% | Sourced from Cursor | No |
| AutomationBench: 56.3% | Official public leaderboard | No (Artificial Analysis's index has its own variant, AutomationBench-AA) |
| Agents' Last Exam v1.1: 75.8% | Self-computed, 122 tasks, Antigravity-CLI 1.4.8, 5 hour window, safety filters on; Sol from its system card, Opus from the leaderboard | No |
| RiemannBench: 84.0% | Official public leaderboard | No |
| Terminal-Bench-Science 0.1: 70.5% | Self-computed, mini-swe-agent vision harness, average of 5 attempts; others from the leaderboard | No |
| HLE-Diamond: 71.3% | Reported from the Artificial Analysis leaderboard | No score printed on the saved page, though HLE-Diamond is one of its nine index evaluations |
| MKA: 56.5% | Artificial Analysis leaderboard | No |
| Chartography: 73.5% | Self-computed without tools for Neon and Sol; Opus from its system card | No |
| LVBench: 89.0% | Self-computed without tools; frame sampling differs per model | No |
| GraphWalks: 95.2% | Self-computed on 200 problems of 256k to 1M tokens | No |
| SEC-bench Pro: 90.5% | Self-computed; 183-task JavaScript engine split; pass@1 over 3 attempts; safety filters off; competitors from system cards | No |
| WebExploitBench v1.2: 51.3% (post: 51.26%) | Self-evaluated and validated by AgentCyberRange; 15 targets, 277 vulnerabilities; safety filters off | No |
| AutoBaxBench 76%, ExploitGym, Gray Swan IPI | Provided by the benchmark's authors; self-computed, mitigations off; evaluated and provided by Gray Swan | No. ExploitGym and Gray Swan print no figure beyond the post's words |
What the model card states, and what it leaves out
The model card is a ten page PDF (a cover and nine numbered pages) of about 1,600 words, "Published: November 2026". Its Model Information, Architecture, Training Dataset, Hardware, Software and Acceptable Usage entries all say the same thing: Neon "is adapted from Gemini 4 Carbon", and the reader should see the Carbon card, which this site did not read. Its Safety and Security section says "The full Safety & Security Report for Gemini 4 Neon will be available in the coming days", so the report the card points to was not yet available when the card was read.
What the Gemini 4 Neon model card and launch post state, and what neither states. Read 18 November 2026.
| Topic | Stated | Not stated |
|---|---|---|
| What the model is | Adapted from Gemini 4 Carbon. Knowledge cutoff March 2026. Context window 1M tokens combined; 64K default synchronous output, up to 1M by background long decode | Parameter count (Artificial Analysis: Google has not disclosed it), training data, compute and cost to train: only by reference to the Carbon card |
| Frontier safety | A Material Capability Change assessment against Gemini 4 Argon found "an increase in general capability" below the MCC threshold, so "a full suite of critical capability evaluations was not required". Results are carried forward from Argon | Neon's own critical capability results. The card also says it evaluated Argon and Carbon, and names Argon as the latest model with a critical capability assessment; it does not say why Carbon is not the reference |
| Cyber, in the frontier safety table | Cybersecurity Uplift CCL: "CCL cannot be ruled out (carried forward from Argon)". The biological CBRN CCL also "cannot be ruled out", and so does the Loss of Control stealth and situational awareness TCL | What threshold Neon for cyber was assessed at, or whether its cyber access is a different model |
| Automated safety evaluations | Against Gemini 4 Hydrogen, not Carbon: text 0.00pp, multilingual -0.40pp, image to text +1.90pp, tone +1.25pp, unjustified refusals -3.00pp. The card says they are "automated evaluations and not human evaluation or red teaming" | Any comparison with Carbon, the model Neon is adapted from |
| Google's own post, in addition | Refusals on 1.98% of Terminal-Bench 4.0 tasks and 0% on Terminal-Bench-Science; 2.45% shortcuts on its coding cheating evaluation; 0.09% misaligned actions; 99.2% on its agent honesty evaluation; 0.0% attack success on Gray Swan's benchmark with safeguards at the highest thinking level | Who designed these evaluations beyond Google and Gray Swan, their sizes, and results with the safeguards off. All are Google's reports |
| Access and data | Channels: Gemini app, Gemini API, Google AI Studio, Gemini for business, Google Antigravity. Terms: Gemini API Additional Terms of Service and Google Cloud Platform Terms of Service | What the terms say about retention or training on customer data: this site did not read them |
The card's sentence on the cyber threshold is the one for a security reader: for Neon, as for the Argon model its results come from, Google says the cybersecurity uplift level "cannot be ruled out". That is a statement about the threshold under Google's own framework, not a finding that Neon was used to attack anything, and the card does not say what the mitigations are beyond "we continue to deploy mitigations".
Where the records differ, and what each difference is
Each row is a place where two records print something that looks like a difference. Several turn out to be rounding or counting rather than disagreement; the table says which.
Differences between the launch post, the card, the methodology page and the independent pages, as read on 17 and 18 November 2026.
| Item | One record | The other, and what it is |
|---|---|---|
| Rank on Artificial Analysis | Post: "#3 overall" | Page: "# 6 / 96". The 96 counts entries, a model at each effort level; one per model, Neon is third (derived). Consistent, not a difference |
| Vals score | Post and table: 70.9% | Vals: 70.87%. Rounding |
| WebExploitBench | Post: 51.26% | Table: 51.3%. Rounding |
| Cost against Carbon | Post: 32% lower "at its default high thinking level" | Artificial Analysis headline: $6.44 at Xhigh against Carbon's $5.48. The same page's chart data has Neon high at $3.74, 31.8% lower. A difference of effort level, resolved by that data |
| Thinking level of the numbers | Methodology: "highest thinking settings unless indicated otherwise" | FrontierCode 2.0 is at "high thinking"; the CWE-bench page says its rollouts are at "high reasoning"; the Artificial Analysis index is at xhigh; Vals prints no setting. One paragraph of the post mixes high and xhigh |
| CWE-bench task count | Board: a public set of 100 tasks | The 70.7% is on all 250 tasks, public and held out |
| SWE-sweep | Table: 39.0% on the "Provider harness" | 12.8% on the shared mini-SWE harness; the post says "state-of-the-art" for the first. Different harnesses, as printed |
| Methodology address | Model card text: an address ending deepmind.com | The table image and the page this site read: deepmind.google. The card's text address was not requested |
| Coding Agent Index | Post: "#1 overall" (58.6) | No saved page. Not a difference, a gap |
What to do, in order
First, what the post says about availability and thinking levels. Neon is "rolling out today in the Gemini API in Google AI Studio, Gemini for business (via API and model dropdown), for paid Google AI Pro and Ultra subscribers in Google Antigravity, and for Pro and Ultra subscribers in the Gemini app", priced at $3 and $15 per million input and output tokens with a 1 million token window. It has four thinking levels, "scale steadily", with high the default and xhigh the setting its Artificial Analysis score is measured at. Neon for cyber is separate in the post's account: "eligible Fairwind Program members", by request through a Cloud account team. The steps below are this site's suggestions.
Take this with you
This site's suggestions
- Choose the thinking level on cost per completed task, on your own work. Artificial Analysis prints $0.75, $1.90, $3.74 and $6.44 per index task at low, medium, high and xhigh for index scores of 44, 55, 60 and 61. Record tokens and dollars per task at the level you will run.
- For security teams, test Neon on your own code with finding, validating and patching measured separately, as the post splits them, and keep your own true and false positive record. Do not use a vendor benchmark as your rate. Ask Google whether the model you can reach is the one the cyber figures were measured on, and with what safeguards.
- For agent use, test blocker reporting and permissions in your own harness. The post says Neon is steerable and handles long, tool heavy sessions; brief 731 shows how a lab's own card can print blocker omission rates, and the Neon pages print none.
- In procurement, ask for the table's settings and the methodology page's footnotes, for results with safety filters on, and for results against the models you actually compare, Fable 5.2 and GPT-6.2 Astra, not only the two in the table.
- For data and compliance, read the terms for the route you will use (Gemini API Additional Terms of Service, or Google Cloud Platform Terms of Service for Gemini for business). This site did not read them.
- Re-read the independent pages before you decide. Artificial Analysis's ranks and costs are rolling figures, and Google's Safety and Security Report was still to come.
The question the table leaves open
Google's rank claims hold up on the three independent pages this site saved, and on each of them the model ahead of Neon is one Google did not put in its table. If a launch table compares a model with two rivals that Vals places fifth and tenth, which comparison did you need: the one on the slide, or the one against the model that is first?
Key facts
Sources
- PrimaryIntroducing Gemini 4 Neon, launch post dated 18 November 2026, rendered text copied about 22:55 UTC: the rank, cost and cyber claims, availability, safety statements, the launch tableGoogleaccessed 2026-11-18
- PrimaryGemini 4 Neon model card, 8 pages, Published November 2026: adapted from Carbon, intended use, frontier safety assessment, automated safety evaluations, the same results tableGoogle DeepMindaccessed 2026-11-18
- PrimaryGemini 4 Neon Model Evaluation, approach, methodology and results (6 page PDF): who ran each benchmark, harnesses, settings and footnotesGoogle DeepMindaccessed 2026-11-18
- PrimaryGemini 4 Neon model page as read 18 November 2026 about 22:47 UTC: Intelligence Index 61 at Xhigh, rank 6 of 96, rank hover list, chart data for four effort levels, cost per task, speed, tokensArtificial Analysisaccessed 2026-11-18
- PrimaryGemini 4 Carbon model page as read 17 November 2026 about 19:30 UTC: index 60 at High, $5.48 per task, 33.4 minutes, 205M output tokensArtificial Analysisaccessed 2026-11-17
- PrimaryClaude Opus 5.5 model page as read 17 November 2026 about 19:30 UTC: index 59, rank 8 of 91, $10.04 per task, list pricesArtificial Analysisaccessed 2026-11-17
- PrimaryGPT-6.2 Sol model page as read 17 November 2026 about 19:30 UTC: index 57, rank 11 of 91, $1.01 per task, list pricesArtificial Analysisaccessed 2026-11-17
- PrimaryGPT-6.2 Astra model page as read 17 November 2026 about 19:30 UTC: index 62, rank 3 of 91, $4.00 per task, list pricesArtificial Analysisaccessed 2026-11-17
- PrimaryClaude Fable 5.2 model page as read 17 November 2026 about 19:30 UTC: index 65, rank 1 of 91, $26.54 per task, list pricesArtificial Analysisaccessed 2026-11-17
- PrimaryVals Index page, updated 11/18/2026, version 2.2, as read about 23:00 UTC: leaderboard with accuracy, cost per test and latency, methodologyVals AIaccessed 2026-11-18
- PrimaryCWE-bench v2 page as read 18 November 2026 about 23:00 UTC: leaderboard on 250 tasks at high reasoning, pass@1, pass@4, cost per rollout, method and scopeCollinear AIaccessed 2026-11-18


