Google's Gemini 4 Neon Ties Carbon at Equal Effort and Leads by One Point at Xhigh
Artificial Analysis scores Google's Gemini 4 Neon 59.6 against Carbon's 60.0 at high effort, and Kilian Lieret's SWE-sweep board peaks at 8.9 percent where Koray Kavukcuoglu claims 39.0 for Neon.
By Ryan Elliott Dennis · 14 sources · 9 min read

Google released Gemini 4 Neon on Nov. 18, 2026, at $3 per million input tokens and $15 per million output tokens 1. Artificial Analysis, an outside measurer, scored it 61 on its Intelligence Index that day, one point above Gemini 4 Carbon and third behind Anthropic's Claude Fable 5.2 and OpenAI's GPT-6.2 Astra 2. That 61 came at xhigh, a thinking level Google offers on Neon and has yet to offer on Carbon. At high effort, the top setting the two models share, Neon scored 59.6 and Carbon 60.0 34.
Who needs to know? A team choosing a model for coding agents pays for the difference, and this ledger counts a capability gain when an outsider measures one. Four independent measurers tested Neon within a day of launch, and on each of the three that publish an error figure, the gap separating Neon from Carbon sits inside it.
Gemini 4 Neon's tie with Carbon leaves capability flat
The moveThe data5 rows · sources
| Measure | Value |
|---|---|
| Reading before this day | 8.6 |
| This piece's move | 0.0 (Capability, confirmed) |
| Band for confirmed evidence | 0.8 to 1.5 |
| Reading after the day (with 2 other pieces that day) | 8.6 |
| Distance to 100 | 91.4 |
What did Artificial Analysis measure?
Artificial Analysis builds its Intelligence Index from nine evaluations across knowledge work, coding, science and general knowledge, and it runs every model itself 3. Neon has four effort levels, the settings that control how long the model reasons before it answers: xhigh, high, medium and low. Their scores run 61, 60, 55 and 44, while Carbon's three levels peak at 60 on high 4.
Rounding hides the order. The chart data behind those pages puts Neon's high run at 59.6, Carbon's at 60.0 and Neon's xhigh run at 61.4 34. So the one-point lead appears at the single setting where Neon runs alone. Neon's xhigh run generated 283 million output tokens across the index, 2.3 times the 123 million of its high run, and those additional tokens bought 1.8 points.
Koray Kavukcuoglu, senior vice president at Google DeepMind and Google's chief AI architect, placed the result in the second paragraph of his launch post 1. "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, landing just behind larger, slower models," he wrote. Read the last clause. Fable 5.2 scored 64.6 and Astra 62.1, and Kavukcuoglu leaves both unnamed 3. Maximilian Schreiner of The Decoder caught the same gap in Google's chart on Nov. 19, writing that Neon trails Astra and Fable, "both of which Google omits from its own comparison table" 11.
Inside the index, Neon and Carbon trade places test by test. Neon gained about four points on FrontierCode-AA 2.0, a set of software engineering tasks, and lost 8.5 points on the hard subset of CritPt, a physics test, 60.5 against 69.0 3. Terminal-Bench 4.0, which Artificial Analysis runs through the neutral mini-swe-agent harness over 66 tasks with three repeats each, went to Carbon by 65.2 percent to 63.6 14. A harness is the program wrapped around a model that hands it tools, files and a loop to work in.
Artificial Analysis Intelligence Index: Neon passes Carbon only at xhigh
ComparedDo other boards separate Neon from Carbon?
Three more independent measurers published within a day, and each one prints its own error.
Vals AI ranks Neon second on its Vals Index at 70.87 percent, behind Fable 5.2 at 71.31 and ahead of Carbon at 70.45 6. Each score carries a standard error near 0.95 points, according to the page's chart data, and Neon's margin over Carbon is 0.42.
Andon Labs runs Vending-Bench 2, a simulated year of managing a vending machine business, scored by the closing bank balance 7. Neon leads at $16,843, plus or minus $972. Carbon follows at $16,484, plus or minus $1,096, so $359 separates two models whose errors each span about $1,000.
Plotly, the charting software company, ran Neon through its private 43-question data analysis exam on launch day 12. Neon passed all 43 and Carbon 42. "One run per model is a single sample. In our repeat runs of other models, scores moved by up to three points between runs, so treat a one- or two-point gap as close to a tie," Plotly wrote.
Terminal-Bench's own board, hosted by Stanford, Harbor and the Laude Institute, had yet to list Neon on Nov. 19 8. It shows Carbon at 64.5 percent, plus or minus 3.2, inside Google's Antigravity CLI, while Google's post cites 65.2 percent for Neon 1. That figure matches the Antigravity run by Artificial Analysis, where Neon's 65.2 beat Carbon's 64.6 by a single attempt out of 198, by desk arithmetic 5.
What does Neon cost against Carbon today?
Kavukcuoglu built the launch on cost 1. "Neon delivers high performance at a lower cost per task across third-party benchmarks," he wrote, adding that Neon at high effort costs 32 percent less per task than Carbon.
Artificial Analysis's measurements support a larger saving, on one condition. At high effort Neon cost $3.74 per index task and Carbon $7.76, a 52 percent cut, with Carbon billed at its standard $4 and $20 per million tokens 34. Carbon's standard price starts Dec. 1, and through Nov. 30 Google bills the model at half that rate 2. Halve Carbon's bill and it lands near $3.88, an estimate by desk arithmetic. On that estimate Neon runs about 4 percent cheaper at equal effort and 66 percent dearer at xhigh, where a task costs $6.44.
Simon Willison, the developer behind Simon Willison's Weblog, read the price list the same way on launch day 10. "Except Carbon is currently half price, so right now Neon ($3/million input, $15/million output) is the most expensive of the three - then Carbon ($2/$10), then Hydrogen ($1.50/$7.50)," he wrote. Plotly's bill agrees with him: its exam cost $17.03 on Neon and $11.55 on Carbon at launch-day prices, though Neon used 20.1 million tokens to Carbon's 29.5 million 12.
Artificial Analysis explained the xhigh saving in its launch note 2. "Neon uses ~15% more output tokens per task, so the lower cost comes purely from its 25% lower token prices," the company wrote. Speed is the gain that survives every price list: about 110 output tokens per second against Carbon's 57, and 14.8 minutes per task at high effort against 36.6 34.

Cost per index task: Neon at $3.74 and $6.44, Carbon at $7.76 standard
ComparedWhere does Neon lead, and on whose harness?
Coding makes the strongest case for a step up. Artificial Analysis's Coding Agent Index runs each model inside the agent its own vendor builds, and Neon in Antigravity CLI scored 59, first on the board 2. Carbon scored 55 in that same harness, while Claude Fable 5.2 and Claude Opus 5.5, each inside Anthropic's Claude Code, scored 58. Neon's average task cost $6.86 against $15.47 for Fable. On FrontierCode-AA 2.0, Neon beat Carbon by five points, 43.0 to 38.0, over 396 attempts each 5.
How much of a coding score belongs to the harness? Kavukcuoglu gave the widest answer in his own post 1. "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," he wrote. SWE-sweep hands an agent a repository holding many bugs, withholds every hint about them, and counts how many the agent repairs.
SWE-sweep's own board shows a different picture 13. Updated Nov. 2, sixteen days before Neon shipped, it lists 18 results, every one run through mini-SWE-agent, and its top score is 8.9 percent for OpenAI's Astra 6. Google's two figures for Neon have yet to appear on it, and Antigravity has yet to appear as a harness. Kilian Lieret of Meta Superintelligence Labs, first of the benchmark's 12 authors, addressed vendor harnesses on that page with his co-authors. "Our paper has an ablation with Claude Code and Codex. Neither seems to significantly outperform mini-swe-agent (to the contrary, mini-swe-agent is even quite a bit better than Codex)," they wrote. Public submissions, the page says, are coming soon.
Lieret's team tested two vendor harnesses and found them level with the neutral one. Google reports a threefold lift from a third, measured by Google. Artificial Analysis's Terminal-Bench runs sit closer to Lieret's finding: Neon scored 63.6 percent through mini-swe-agent and 65.2 through Antigravity CLI, about 1.5 points apart across two versions of the test 35.

What outsiders measured for Neon, and what rests on Google's harness
The recordThe data10 rows · sources
| Column | Item | Source |
|---|---|---|
| Measured by outsiders | Coding Agent Index: Neon 59, Carbon 55, both inside Antigravity CLI | [2] |
| Measured by outsiders | FrontierCode-AA 2.0: Neon 43.0, Carbon 38.0, over 396 attempts each | [5] |
| Measured by outsiders | Terminal-Bench: 63.6 percent on mini-swe-agent, 65.2 on Antigravity CLI | [3] [5] |
| Measured by outsiders | Vending-Bench 2: Neon $16,843, $359 above Carbon, errors near $1,000 | [7] |
| Measured by outsiders | SWE-sweep board: top score 8.9 percent, all 18 rows on mini-SWE-agent | [13] |
| Google's own figures | SWE-sweep: 12.8 percent on the mini-SWE harness | [1] |
| Google's own figures | SWE-sweep: 39.0 percent inside Google Antigravity | [1] |
| Google's own figures | FrontierCode 2.0: 46.2 percent at the default high setting | [1] |
| Google's own figures | Cost per task 32 percent lower than Carbon at high effort | [1] |
| Google's own figures | Capability increase rated below Google's material change threshold | [9] |
Why does the ledger hold flat?
Capability holds flat on the ledger at confirmed confidence. Independent boards measured Neon, and their measurements show a tie with Carbon, the model this ledger counted on Nov. 16. Neon's two first places are a coding index, by one point over Fable inside Google's harness, and a vending simulation, by a third of one standard error. Both composite boards, from Artificial Analysis and Vals AI, keep Fable 5.2 first.
Google's own paperwork points the same way 9. The model card states that "Gemini 4 Neon is adapted from Gemini 4 Carbon." Its safety section reports Google's Material Capability Change assessment, the internal test that decides whether a model needs fresh frontier safety evaluations. "Gemini 4 Neon showed an increase in general capability under our MCC assessment compared with Gemini 4 Argon, however this falls below our MCC threshold and therefore a full suite of critical capability evaluations was not required in line with our FSF," Google wrote. On Nov. 19 the card still listed the full safety report as coming soon.
So what would each side need? Kavukcuoglu's claim holds if Neon keeps its coding lead once an outside board runs every model through one neutral harness, and if the cost gap opens on Dec. 1 as the price list implies. Willison's reading holds through Nov. 30 by the calendar.
For a team choosing between the two, the measurements give a rule. Buy Neon for speed today and for price starting Dec. 1. Expect Carbon's capability at high effort, and expect to spend 2.3 times the output tokens for 1.8 more points at xhigh. Count the frontier where Fable 5.2 and Astra left it.
Four Gemini 4 models in seven weeks, then a price change on Dec. 1
TimelineBy the numbers
- 61 against 60: Neon at xhigh and Carbon at high on the Artificial Analysis Intelligence Index, with Neon third overall 2.
- At high effort, Artificial Analysis's chart data scores Neon 59.6 and Carbon 60.0 34.
- $3.74 against $7.76: cost per index task at high effort, with Carbon at its standard price 34.
- Vals AI puts Neon at 70.87 percent and Carbon at 70.45, each with a standard error near 0.95 points 6.
- $16,843, plus or minus $972: Neon's Vending-Bench 2 balance, $359 above Carbon's 7.
- 8.9 percent: the top score on SWE-sweep's own board, where Google claims 12.8 and 39.0 for Neon 113.
- Four Gemini 4 models shipped in seven weeks, by Artificial Analysis's count 2.
- Terminal-Bench lists Carbon at 64.5 percent, plus or minus 3.2, and had yet to list Neon on Nov. 19 8.
What to watch
Terminal-Bench and SWE-sweep have yet to post a Neon row, and SWE-sweep says public submissions are coming soon 813. A Neon result near 39 percent on that board, in a harness its authors accept, would turn this hold into a step up. Carbon's price doubles on Dec. 1, and the next runs from Plotly and Artificial Analysis will show whether Neon's cost per task falls below Carbon's 212. Google's full safety report for Neon is due, and a revised capability assessment there would also move the reading 9.
Sources
- 1Introducing Gemini 4 Neon, Google, Koray Kavukcuoglu, Nov. 18, 2026
- 2Google continues to accelerate its release cadence with Gemini 4 Neon, Artificial Analysis, Nov. 18, 2026
- 3Gemini 4 Neon Models: Intelligence, Performance and Price Comparison, Artificial Analysis, Nov. 18, 2026
- 4Gemini 4 Carbon Models: Intelligence, Performance and Price Comparison, Artificial Analysis, Nov. 18, 2026
- 5AI Coding Agent Benchmarks and Leaderboard, Artificial Analysis, Nov. 19, 2026
- 6Vals Index Leaderboard and Methodology, Vals AI, Nov. 18, 2026
- 7Vending-Bench 2, Andon Labs, Nov. 19, 2026
- 8Terminal-Bench 4.1 leaderboard, Terminal-Bench, Stanford, Harbor and the Laude Institute, Nov. 19, 2026
- 9Gemini 4 Neon Model Card, Google DeepMind, Nov. 18, 2026
- 10llm-gemini 0.38, Simon Willison's Weblog, Simon Willison, Nov. 18, 2026
- 11Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks ago, The Decoder, Maximilian Schreiner, Nov. 19, 2026
- 12Gemini 4 Neon Benchmark Results for Data Analysis, Plotly, Nov. 18, 2026
- 13SWE-sweep: How many bugs can LMs find and fix in large codebases?, SWE-sweep, Kilian Lieret, Jeffrey Jian Ma, Rahul Kindi, Yuxiang Wei, Jeremy Ma, Sten Sootla, Parth Thakkar, Chao Beyond Zhou, Pengcheng Yin, Rui Hou, Ofir Press and John Yang, Nov. 2, 2026
- 14Intelligence Benchmarking Methodology, Artificial Analysis, Nov. 19, 2026