
Gemini 4 Neon Tops the Coding Agent Index at 58.6 — the Gap to Claude Fable 5.2 Is 0.17 Points, and Most of the Gap to GPT-6.2 Astra Is One Benchmark
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 185 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 214 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 76 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 112 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 210 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Gemini 4 Neon now holds first place on the Artificial Analysis Coding Agent Index at 58.64, with Claude Fable 5.2 at 58.47 and GPT-6.2 Astra — released on 17 November, two days before this reading — at 53.47. That is the claim going around, and every number in it checks out against the published leaderboard. It is also, read carefully, an argument about harnesses as much as about models: Neon's 58.64 was measured inside Google's own Antigravity CLI, Fable 5.2's 58.47 inside Claude Code, Astra's 53.47 inside Codex. Three labs, three terminal agents, one composite built from 396 FrontierCode 2.0 attempts, 372 SWE-Atlas-QnA attempts and 198 Terminal-Bench 4.1 attempts per row. The ranking is real. What it means for your repository is narrower — and considerably more useful — than a leaderboard screenshot suggests.
Everything below is read off the public coding-agents leaderboard at Artificial Analysis on 19 November 2026, on index version v2.1. None of it is Google's own benchmark table: the composite, the three component scores, the cost per task and the token counts are measured in the lab, each run through the vendor's own terminal agent. Google's launch-day figures for Gemini 4 Neon — Chartography 73.5%, Terminal-Bench 4.1 at 65.2%, FrontierCode 2.0 at 46.2% — are a separate set, vendor-reported and not the source of any measured figure below, and they are labelled as such wherever they appear next to these.
What the index actually counts
The Coding Agent Index does not test a model in isolation. Each row is a harness-plus-model pair, and the headline number is the equal-weighted mean of three component scores: FrontierCode 2.0, 132 software-engineering tasks written by Cognition and implemented by Artificial Analysis; SWE-Atlas-QnA, 124 technical Q&A tasks from Scale AI with grading adjusted by AA; and Terminal-Bench 4.1, 66 agentic terminal tasks from the Laude Institute. Each component is the mean score over three attempts per task. The page publishes no confidence interval on any of it, which is the first thing to know before treating a tenth of a point as a fact about capability.

That caveat matters more than usual here, because the top of the board is extraordinarily tight. Ranked by composite, to two decimals:
• Gemini 4 Neon (xhigh), Antigravity CLI — 58.64
• Claude Fable 5.2 (xhigh), Claude Code — 58.47
• Devin Fusion CLI, Claude Opus 5.5 max + GPT-6 Luna high — 58.40
• Claude Fable 5.2 (high), Claude Code — 58.24
• Claude Opus 5.5 (max), Claude Code — 58.16
Five configurations inside half a point, spanning three vendors and three different agents. Whatever separates them is not visible at this resolution, and an article that describes the top of this board as a decisive result is describing rounding as signal.
The five-point gap is mostly one benchmark
The GPT-6.2 Astra comparison is where the headline gets interesting, because a five-point composite gap sounds structural and turns out to be concentrated in a single component. It also needs one correction up front: the "six points ahead" figure circulating since launch is what you get by comparing Neon's 58.64 against Astra at high, which scores 52.28 — 6.36 points back. Against Astra's best configuration, max at 53.47, the gap is 5.17. Both numbers are on the board; they describe different settings of the same model. Break the max pairing apart:
• FrontierCode 2.0 — Neon (xhigh) 43.0 vs Astra (max) 41.7, a gap of 1.3 points across 132 software-engineering tasks.
• Terminal-Bench 4.1 — Neon 65.2 vs Astra 63.6, a gap of 1.6 points across 66 terminal tasks.
• SWE-Atlas-QnA — Neon 67.7 vs Astra 55.1, a gap of 12.6 points across 124 technical Q&A tasks.
Roughly four-fifths of the composite difference comes from the Q&A component, and that is the component least like what an agent does when it edits a repository. SWE-Atlas-QnA asks the harness to answer questions about a codebase. FrontierCode and Terminal-Bench ask it to change one, or to drive a terminal until a job completes. On the two task sets that resemble real agent work, Gemini 4 Neon and GPT-6.2 Astra sit inside two points of each other — the same order of magnitude as the spread among the five configurations at the top. On the task set that resembles retrieval-and-reasoning, Neon is better by twelve.
Both readings are true, and which one predicts your experience is a property of your workload rather than of the leaderboard. If your agent spends its day answering questions over a large private codebase, SWE-Atlas-QnA is the closest proxy on the board and a twelve-point gap is worth acting on. If it spends its day opening pull requests, FrontierCode and Terminal-Bench are the closer proxies, and the composite overstates the difference by a factor of three or four.

Cost per task runs the other way
Artificial Analysis reports cost alongside score, and on cost the ordering inverts. Mean cost per task across the full index:
• Gemini 4 Neon (xhigh) — $6.86, 13.72M input and 125.6k output tokens, 177 steps, 30.5 minutes wall time.
• Claude Fable 5.2 (xhigh) — $15.47, 9.55M input and 207.4k output tokens, 95 steps, 53.3 minutes.
• Claude Fable 5.2 (high) — $7.51, 5.02M input and 94.1k output tokens, 55 steps, 31.3 minutes.
• GPT-6.2 Astra (max) — $5.12, 4.62M input and 43.0k output tokens, 51 steps, 25.1 minutes.
• Gemini 4 Carbon (high) — $9.75, 15.31M input and 120.5k output tokens, 186 steps, 33.6 minutes.
Three things fall out of that. First, Neon is 2.3 times cheaper per task than Fable 5.2 at the setting the Fable row is scored at, and finishes in 30.5 minutes against 53.3. Second, Neon is not cheap in absolute terms: at $6.86 it costs 34% more per task than the GPT-6.2 Astra configuration it beats, and 30% less than its own sibling Gemini 4 Carbon, which it beats by 3.2 composite points — so against Carbon the case is unambiguous on both axes, and against Astra it is a straight trade of dollars for score. Third, the token profile explains the price. Neon burns 13.72M input tokens and 177 steps per task against Fable 5.2's 9.55M and 95: about 44% more input and 86% more tool calls for a result 0.17 points ahead. Cache hit rates sit between 94% and 96% on every row above, so that is not a caching artefact — it is the model taking more, smaller steps.
Google's launch material claims Gemini 4 Neon at its default high setting costs 32% less per task than Gemini 4 Carbon. That is vendor-reported and this board does not reproduce it. The only Neon row published here is xhigh, at $6.86 against Carbon's $9.75 — a 30% saving, close to the vendor figure, but measured at the deepest of Neon's four thinking settings rather than at the default the claim describes. Read the direction as supported and the precise number as unconfirmed until a high row exists.
The setting moves the score more than the model does
The most useful thing on this board is not the ranking. It is that Claude Fable 5.2 appears five times, in the same harness, at five thinking levels, and its scores run 47.96 at low, 53.04 at medium, 58.24 at high, 58.47 at xhigh and 55.74 at max — a 10.5-point range from a parameter. GPT-6.2 Astra appears five times in Codex and runs 35.58 at low up to 53.47 at max, a 17.9-point range across the same kind of knob.
Set those ranges against the gaps the headlines are about. The spread between Astra at low and Astra at max is 3.5 times the gap between Neon and Astra at their best settings, and 5.6 times the gap between Neon and Carbon. The 10.5-point spread inside Fable 5.2 alone is more than twice the entire distance from Neon's first place to Astra's tenth. Choosing a model from one number per model is choosing on the smallest variable on the page; the effort level you pass and the agent you wrap around it move the result further than the logo on the box.
It also makes the "surpassed Fable 5.2" framing need a qualifier. Neon's 58.64 is measured at xhigh inside Antigravity CLI, and it is the only Neon configuration on the board — there is no high row, so nobody outside Google can yet say what Neon's default costs in score or in dollars. Fable 5.2's 0.17-point deficit holds at its own xhigh and, more narrowly, at high where it scores 58.24 on $7.51 per task. Both of those numbers are inside the range a different harness could plausibly reverse, which is not a criticism of Neon — it is what a half-point spread at the top of a 41-row board means.
What is actually new here
Gemini 4 Neon reached general availability on 18 November 2026, a day before this board was read, at $3.00 per million input tokens and $15.00 per million output for a 1M-token context window. It is proprietary and flat-priced: no introductory window, no free tier, unlike Gemini 4 Hydrogen. The rate card and the independent intelligence scores have been covered elsewhere on this blog; what is new is the first measured agentic result, and it says several things the launch post did not.
The launch post did say it took first place on the Artificial Analysis Coding Agent Index at 58.6, and that figure is exactly what the board publishes, at the configuration described. What the post did not say is that the margin over the next configuration is 0.17 points, that most of the lead over GPT-6.2 Astra comes from one of the three components, or that the run consumed 13.72M input tokens per task. Those are the parts a reader deciding what to deploy needs, and they exist only on the board.
One more figure worth holding beside the score: zero refusals. Across all 966 attempts in the Neon row — 396 FrontierCode, 372 SWE-Atlas-QnA, 198 Terminal-Bench — Artificial Analysis recorded no safety refusals and no provider fallbacks. Fable 5.2 at xhigh recorded 1 refusal on FrontierCode, 38 on SWE-Atlas-QnA and 16 on Terminal-Bench; GPT-6.2 Astra at max recorded 8 refusals on SWE-Atlas-QnA, all of them hard stops. A refusal is not a capability failure, but it is a task the agent did not complete, and inside an agentic loop it presents as a stall rather than a wrong answer. For an agent pointed at production repositories, a model that completes every attempt is an operational difference no other configuration on this board matches.
What switching costs, and where we fit
If Neon's numbers argue for a trial, the practical question is what the trial costs in engineering. Neon is not in our catalogue: the OrcaRouter model list carries google/gemini-4-carbon and google/gemini-4-argon and no Neon entry as of 19 November, so we cannot route it for you today and will not imply otherwise. GPT-6.2 Astra is not in the catalogue either. What we do carry is anthropic/claude-fable-5.2 at Anthropic's list price of $8.00 input and $40.00 output per million, plus the earlier openai/gpt-6-astra — GPT-6 Astra, a September model and a different row on this board — at $10.00 and $50.00. Both sit on the same key at provider list price with 0% markup, which means the arithmetic above can be re-run against your own traffic without a second contract.

What matters most about a board this tight is turnover. Three of the five leading configurations use a model that shipped in the last ten days, and the ordering has moved every time one of them landed. A production path welded to whichever row is first today gets rebuilt when the next model ships, which is why the routing decision belongs outside the code. With one endpoint and a routing rule, a candidate model is a config change rather than a migration, and automatic failover lets an unproven model sit behind a proven one and absorb a slice of traffic without putting the main path at risk.
For the two models we do serve, pass-through is literal: when Anthropic or OpenAI moves a rate, the number on our invoice moves the same day, with nothing layered on top. That is the honest basis for the cost comparison above — re-price both sides at what you would actually be billed, not at a launch-post figure.
Who should act on this, and how
The board supports three different decisions depending on what you run, and they are not the same decision.
If you run an agent that answers questions over a large private codebase — onboarding assistants, "where is this implemented" tools, migration planners — the SWE-Atlas-QnA gap is the number that matters, and 67.7 against 55.1 is wide enough to justify a measured trial. Budget for the xhigh setting, since that is the only configuration scored, and do not assume the default performs the same.
If you run an agent that opens pull requests and drives a terminal, the honest summary is that Neon, Fable 5.2, Opus 5.5 and the Devin Fusion pairing sit inside half a composite point of each other and inside three points on Terminal-Bench, though their FrontierCode scores spread wider, from 43.0 to 49.2. There is no capability case here for switching anything. There is a cost case — Neon at $6.86 against Fable 5.2 at $15.47 for the same score — and an operational case in the zero-refusal row. Neither is a reason to migrate a working pipeline this week.
If you are still choosing a harness, that is where the evidence is strongest and least ambiguous. The spread across thinking levels inside a single model dwarfs the spread between the leading models at their best settings. Two afternoons spent running your own agent at two or three effort levels will tell you more about your bill and your pass rate than any amount of time spent reading this leaderboard, this article included.
What to watch: a Neon row at its default high setting. Until one is published, the model that tops the board has been measured at exactly one configuration — and the 32%-cheaper-per-task claim Google makes is about the setting nobody has scored yet.
Compared in this article2
Detected from this article · Benchmarks: Artificial Analysis · updated daily
