Gemini 4 Neon Pros & Cons Deep Dive: #3 Globally at 61, Why a 3x Harness Gap and Pricier Than Astra?
HyphenTech · Field notes

Gemini 4 Neon Pros & Cons Deep Dive: #3 Globally at 61, Why a 3x Harness Gap and Pricier Than Astra?

7 min read
Categories:
Tags:
Summary

Google DeepMind just launched Gemini 4 Neon, adapted from flagship Carbon at 25% lower list price ($3/$15). By cross-examining the 10-page Model Card, 6-page Evals Methodology PDF, Artificial Analysis, Zapier AutomationBench, and CursorBench 4.1, we unpack why Neon wins two independent agent leaderboards at High effort yet drops 3x outside Antigravity and costs $6.44/task at Xhigh—more than GPT-6.2 Astra ($4.00).

Summary Model Deep Dive · Evidence-Backed Pros, Cons & Real Task Bills Priced at $3/$15 with #1 finishes on Zapier and CursorBench, why does Neon drop 3x in generic harnesses and burn $6.44 per task at Xhigh? Understanding the two paid knobs behind Google’s new workhorse HyphenTech · 2026-11-19

  On November 18, 2026, Google officially released Gemini 4 Neon, its new mid-tier frontier reasoning model. At first glance on the Google DeepMind launch post, the pitch looks almost unbeatable: 61 on the Artificial Analysis Intelligence Index (#3 globally, trailing only far larger models) while cutting list prices by a quarter compared to flagship Gemini 4 Carbon.

  Yet two rows further down the exact same official benchmark table sits a jarring contrast: inside a shared open-source mini-SWE harness, Gemini 4 Neon resolves just 12.8% of tasks; move that exact same model into Google’s proprietary Antigravity environment, and the score leaps to 39.0%—a 3x swing from switching the tool harness alone.

DeepMind official benchmark table showing a 3x gap on SWE-sweep between Antigravity Provider harness (39.0%) and generic Mini-SWE harness (12.8%)

  An even sharper surprise hides in the invoice: Neon lists at $3 input / $15 output per million tokens—less than 40% of GPT-6.2 Astra ($8 / $40). Yet across the full Artificial Analysis run, cranking Neon to xhigh pushed its average bill to $6.44 per task, making it more expensive per task than GPT-6.2 Astra Max ($4.00 per task). Instead of running cherry-picked toy prompts, let’s lay out the 10-page Model Card PDF, the 6-page Evals Methodology PDF, and raw leaderboards from Artificial Analysis, Zapier, and Cursor side by side.

Animated breakdown: 3x SWE-sweep score divergence across harnesses for Gemini 4 Neon ↗

1. Family Positioning & Three Verified Strengths: Undercutting Rivals in the 60+ Bracket

  First, place Neon inside the Gemini 4 lineup. Google now splits the family into three clear tiers: Carbon tackles the hardest long-horizon research, Hydrogen handles ultra-fast low-cost execution, and Neon sits squarely in the middle. As the Gemini 4 Neon Model Card page states plainly, Neon is adapted directly from Carbon and exposes 4 thinking levels: low, medium, high (default), and xhigh.

  On Google’s comparison graphic, AA Quality stands at 60 for Carbon, 61 for Neon, and 54 for Hydrogen. Per million tokens, Carbon lists at $4 / $20 (with a $2 / $10 promotional rate through November 30), Neon lists at $3 / $15 with $0.15 implicit cached input, and Hydrogen sits at $1.50 / $7.50. Speed bars rate Carbon at 2, Neon around 3.5–4, and Hydrogen at 5/5—delivering a 25% lower list price and 32% lower per-task cost at default high effort than Carbon.

Google official Gemini 4 family comparison across Carbon, Neon, and Hydrogen

  Where Neon genuinely shines comes down to three concrete pillars:

  • Broad leads over similarly priced peers on official suites: Against Claude Opus 5.5 and GPT-6.2 Sol, Neon leads on the AA Intelligence Index (61 vs. 59 vs. 57), Vals Index (70.9% #2 vs. 70.1% and 66.1%), HLE-Diamond (71.3% vs. 67.9% and 64.0%), and long-context GraphWalks (95.2% vs. Opus’s 66.8%), while taking #1 on the AA Coding Agent Index (58.6) alongside 90.5% on SEC-bench Pro, 84.0% on RiemannBench, and $16,843 on Vending-Bench 2.
  • Record #1 on Zapier AutomationBench—and 55% cheaper than Carbon at High: In the independent Zapier AutomationBench report spanning 657 end-to-end enterprise tasks, Neon xhigh solved 370 (56.3%) to set a new benchmark record—beating Carbon (53.9%) by 2.4 points and GPT-6.2 Astra Max (52.4%) by 3.9 points. Better yet, dropping to default high retains 54.3% accuracy at $0.88/task (55% cheaper than Carbon’s $1.94), while medium hits 53.3% at $0.58/task (30% cheaper than Astra).
  • 1-2 finish on CursorBench 4.1 at a fraction of rival task bills: On the CursorBench 4.1 leaderboard (see also Cursor’s launch post), Neon Extra High takes #1 at 65.4% ($10.67/task, 73 steps), and default High takes #2 at 63.7% for $6.54/task (56 steps)—dramatically undercutting Carbon High (61.5%, $8.52), Claude Opus 5.5 Max (58.1%, $11.88), and Claude Fable 5.2 Max (61.3%, a whopping $48.11/task across 295 steps).

Zapier AutomationBench independent report: 56.3% record pass rate and 55% savings vs. Carbon at High effort

CursorBench 4.1 official leaderboard: Neon takes #1 and #2 with $6.54/task at High vs. $48.11 for Fable 5.2 Max

  Behavioral discipline improved as well: DeepMind reports its lowest internal fabrication rate to date and a 2.45% reward-hacking shortcut rate on coding evals (lowest in Gemini 4), while the GitHub Copilot changelog notes fewer unsolicited edits compared to Carbon. Inside Google Antigravity, Orchestrator mode pairs Neon as planner with 8 parallel Hydrogen workers—building the 3D synthwave mini-golf game NEONPUTT '84 with automated browser verification from a single prompt, before handing Player 2 to Hydrogen in real time.

2. Four Real Trade-Offs: Why $3/$15 Ended Up Costing More Per Task Than $8/$40 Astra

  Now flip the coin over. Treating $3/$15 as an invitation to leave xhigh pinned on every call is the fastest way to blow your monthly API budget.

1. Runaway Reasoning Verbosity: 283M Output Tokens Push Xhigh to $6.44/Task

  Artificial Analysis (full standings on the AA Models Leaderboard) bluntly characterizes Neon (xhigh) as notably slow, very verbose, and somewhat expensive. Each Index task averaged 21.5 minutes (versus a 12.3-minute peer median) with a 56.51-second wait for the first answer token. Across the evaluation suite, Neon emitted 283M output tokens—over 4.3x the 65.1M peer median—driving its average cost to $6.44 per task.

Artificial Analysis Gemini 4 Neon (Xhigh) page: 61 Intelligence, 21.5 min/task, $6.44/task, and 283M output tokens

  Compare GPT-6.2 Astra (Max) side by side: despite a 2.67x higher list price ($8 / $40), Astra generated just 65.8M tokens, finished in 13.9 minutes, scored 62, and averaged $4.00 per task. Token price means nothing without token volume.

Artificial Analysis GPT-6.2 Astra (Max) page: 62 Intelligence, 13.9 min/task, $4.00/task, and 65.8M output tokens

Animated comparison: how 283M output tokens make $3/$15 Neon Xhigh costlier per task than $8/$40 Astra ↗

2. Steep Diminishing Returns Across Effort Tiers—and Failed Runs Cost the Most

  Across AA’s four effort settings, Neon scores 61 at xhigh (21.5 min), 60 at high (14.8 min, 123M tokens), 55 at medium (8.3 min, 62.9M), and 44 at low (3.7 min, 28.3M). On Zapier, cranking from low to xhigh raised solved tasks by just 13% (326 to 370) while per-task cost jumped 320% ($0.40 to $1.69) and reasoning tokens expanded 8.8x to consume 82% of all output. Worse still, failed xhigh runs averaged $2.21 versus $1.29 for successful runs—meaning stuck tasks spin in reasoning loops and pad your bill. Conversely, dropping to low on CursorBench sinks accuracy to 45.5%, trailing cheaper Hydrogen High (57.3%).

Animated Zapier effort curve: +13% pass rate vs. +320% cost surge from Low to Xhigh ↗

3. Harness Lock-In, 512K Client Caps, and Missing Modalities

  • Heavy reliance on Google’s own harness: Dropping from Antigravity (39.0%) to generic mini-SWE (12.8%) shrinks Neon’s lead over GPT-6.2 Sol (11.6%) from 30.7 points down to 1.2 points.
  • 512K context cap inside Antigravity: While the Google AI Studio model docs (and release notes) provide 1M context via API, the Google AI Forum rollout post confirms Antigravity caps Neon at 512K and Enterprise plans omit Orchestrator.
  • Category & modality gaps: Neon trails Opus 5.5 slightly on Terminal-Bench 4.1 (65.2% vs. 65.8%) and FrontierCode 2.0 (46.2% vs. 47.3%), trails Sol on Chartography (73.5% vs. 75.1%), and tops out at 37.5% on Zapier HR tasks (5.8 points behind Astra). It supports no image generation, no audio generation, no Live API, no explicit caching, and no free API tier, with a March 2026 knowledge cutoff and its full safety report still marked coming soon (inheriting Argon’s CCL assessment).

3. Underlying Mechanism & Six-Model Buying Guide: Lock In Default High

  Every paradox in Neon’s launch comes down to two paid knobs: the thinking-depth knob (buying 1–2 extra points with 82% internal reasoning tokens and wait time) and the harness knob (buying a 3x repo-repair boost by locking into Antigravity’s subagent swarm).

Six-model unified AA comparison and scenario selection scorecard

Model & Effort AA Score List Price (In/Out) Avg. AA Task Bill / Time Best Practical Fit
Claude Fable 5.2 (Max) 65 (#1) $8 / $40 $26.54 / 41.8 min Budget-unconstrained frontier science & hardest proofs
GPT-6.2 Astra (Max) 62 (#2) $8 / $40 $4.00 / 13.9 min Concise output (65.8M); faster & cheaper per task than Neon Xhigh
Gemini 4 Neon (Xhigh) 61 (#3) $3 / $15 $6.44 / 21.5 min Hard edge cases only—always pair with strict token/timeout caps
Gemini 4 Neon (High, Default) 60 $3 / $15 ~$3.74 / 14.8 min Daily sweet spot: $6.54 on CursorBench, $0.88 on Zapier
Gemini 4 Carbon (High) 60 $4 / $20 ($2/$10 promo) $7.76 / 37.1 min Worth grabbing before Nov 30 half-price promo ends; heavy projects
Claude Opus 5.5 (Max) 59 $4 / $20 $10.04 / 25.2 min Slight edge in pure CLI repair (Terminal-Bench 4.1 65.8%)
GPT-6.2 Sol (Max) 57 $2 / $10 $1.01 / 8.7 min Chart reading (75.1%) & fast low-cost execution (alongside Hydrogen)

  For day-to-day software engineering, long-context analysis, and multi-step business workflows on your own machine, lock Gemini 4 Neon to its default high effort—it captures Carbon-class 60-point reasoning while keeping bills squarely in the sweet spot. Pair it with GPT-6.2 Sol or Hydrogen for high-frequency lightweight runs, and reserve GPT-6.2 Astra (or Carbon during its pre-November 30 promo) for massive architectural refactors. If you mix multiple cloud endpoints and local models on your desktop, you can route keys cleanly through HyphenBox, wire local TTS/Whisper/video MCP tools in LocalBrain, and edit full-resolution technical breakdowns on the DaVinci-style timeline in HyphenScreen (which produced our companion video for this episode).

Limits and Applicable Boundaries

  • All benchmark scores, latencies, and per-task dollar figures reflect public Model Cards and independent leaderboard snapshots verified as of November 19, 2026; real production bills vary with repository size, implicit cache hit rates, and agent retry loops.
  • Gemini 4 Carbon’s $2 / $10 half-price promotion expires on November 30, 2026, reverting to $4 / $20 thereafter.

Primary Verified Sources

Self-made software

Versions, changes and supported installers all come from the same public release records.

Share:

Comments

Loading comments…