Last month Carbon got the best score on my bench and emptied a $99.99 Ultra plan in two hours. Argon before it scored nearly as well and ate tokens. Both took 16 to 30 minutes per fix.

I resubbed to Ultra for this one. Gemini 4 Neon is the first Gemini 4 model that is fast and cheap enough to use every day.

While Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution, Neon balances depth and efficiency.

— Google, Introducing Gemini 4 Neon

Price per million tokens: Neon $3 in, $15 out. Opus 5.5 $4 and $20. Carbon and Argon are $2 and $10 until November 30, then $4 and $20.

The Same Six Changes

Same bench as the Carbon and Fable 5.2 posts: six merged changes from a production monorepo, replayed on clean snapshots, graded by the tests the original fixes shipped with. Medium effort, 20 runs.

20 runs eachNeonOpus 5.5CarbonArgon
Runs that passed every test17131716
Runs that broke a passing test0000
Total cost at list price$26.93$38.45$60.38$57.68
Average time per fix7 min6 min30 min16 min
  • The four-run lead is one change. Neon built it the way the human fix did. Opus built the same behaviour another way and failed two tests that check exact list contents. I reran Opus today to check: same result. That is a tie, not a win.
  • Everything else matches. Same passes, and the same miss on Change C, where every model fails a hidden rule.
  • 70% of the cost. $26.93 against $38.45.
  • Close to Opus speed. 7 minutes a fix against 6. Carbon took 30.

A Second Tier

I also ran the diagnosis tier from the Fable post: seven real bugs, each described by its symptom only, 4 runs each.

28 runs eachNeonOpus 5.5
Runs that passed every test2827
Total cost at list price$24.83$26.56
Average time per fix5.5 min3.7 min

Another tie on quality. A smaller cost gap, and Opus is faster here.

The Quota Problem Is Gone

48 Neon runs and 6 Orchestrator runs, four at a time, on the same $99.99 Ultra plan: half of one five-hour window and 12% of the weekly limit. Carbon hit the wall in two hours.

Orchestrator Mode

Google pitches Neon “acting as a fast, high-judgment Orchestrator directing parallel Hydrogen subagents in Google Antigravity.” In Antigravity, both models run at xhigh effort. I ran it once per change, against Neon alone at medium:

  • Same results. 5 of 6 clean, like Neon alone.
  • Twice the time. 16 minutes a fix against 7.
  • About three times the quota.

My fixes take minutes, and Neon alone already hits the ceiling. Orchestrator mode may pay off on long, parallel jobs. On normal fixes it bought nothing.

Not Just Me

largbae on Hacker News ran both on a large monorepo port:

Today I started Neon on one worktree and Opus on another, and Neon did so well that I have now switched for both.

— largbae, Hacker News

Their bill: $30 of Neon in a day, against $70 of Opus the day before. Not everyone agrees. aenis found “Neon is far below Carbon in terms of quality, but indeed much faster”. On my bench they tied. asciilore says one window on the $20 Pro plan “eats roughly 25% of the weekly limit”. My numbers are from Ultra.

Different harness

Neon ran in Google’s Antigravity CLI. Opus ran in Claude Code. Last month Argon scored the same in both harnesses, so I don’t think the harness explains much. Costs are token counts at list price.

What Neon Doesn’t Solve

  • Better than Opus? No. Equal on both tiers.
  • Seeing the next bug? No. On two bugs the team shipped a second fix days later. Fable 5.2 anticipated it in 6 of 8 runs. Neon and Opus in 0.
  • Sample size. One codebase. Another could rank these differently.

Lessons

  • The middle model is the one to use. Carbon’s score at a quarter of the time and less than half the cost.
  • That makes Carbon and Argon hard to justify. Slower, dearer per fix, no better on my work. Maybe they win on long projects. I haven’t seen it.
  • Opus finally has a real rival. Same quality on my code at 70 to 95% of the cost.
  • Skip the orchestrator for normal work. Same fixes, more time, more quota.