Inter-office memo
- To
- Everyone deciding whether to try Gemini 4 Neon
- From
- Mo, Chief Executive
- Date
- November 19, 2026, 6:01 PM CST
- Re
- Gemini 4 Neon’s reviews
Is Gemini 4 Neon as good as Opus 5.5? What testers and tests show, and the catches
Close, by most early reviews and tests, for about half the cost per task. Since its November 18, 2026 launch, Theo has called it “so good” and Jeremy Howard “awesome,” and it placed second on the Vals Index and level with Claude Code on Artificial Analysis’s coding-agent test. The catches: it thinks at length, so short tasks can cost more than on Opus, and some users still prefer Opus for long work.
What testers are saying
Google released Gemini 4 Neon on November 18, 2026, its fourth Gemini 4 model in seven weeks, and most of the first reviews are glowing. They landed a day after OpenAI’s GPT-6.2 Astra drew the opposite, and some of the same people made the comparison.
- Theo, the developer behind T3 Code, wrote that evening, “This model is so good,” and asked why nobody was “talking about Neon.” Hours later he had made it “the exclusive copy writer” in his setup, with Claude Fable 5.2 “leading implementation” and GPT-6.2 Astra “as a playtester,” and the next day he was “Seeing some insane results from Gemini 4 Neon on long horizon eng work.”
- Jeremy Howard said he was reverting Astra and Sol in Solveit, the app he develops, to their versions before 6.2, “because 6.2 isn’t working for me, or for our users,” and added: “(PS: Neon is awesome!)”
- Greg Isenberg, who tested it before launch, found it “really good, really fast and about half the price of what I’m used to,” but added: “Claude is still often what I reach for on the hardest problems, and Opus still edges it on a couple of coding benchmarks.”
- Chubby, who also had early access, spent the day building in Antigravity and got 17 interactive demos, from shooters to 3D worlds. Given three broken programs, Neon brought a booking checker from 1 of 5 tests passing to 5 of 5, and the other two to all their checks. His verdict: “great for prototypes and solid code fixes. Expect some follow-ups.”
- Flavio Copes tested it against Opus 5.5: “This is a good model. It got every one of my tests right, it writes well, and on long inputs it costs a lot less than Opus 5.5. On short tasks it thinks a lot, and that changes the math.”
On Hacker News, the launch thread reached 1,010 points and 758 comments by November 19. “This model feels like it might actually be as good as Opus 5.5,” one user wrote. “I’m pretty impressed so far, this is the first google model I’ve used that feels good for coding,” wrote another. A third: “It’s much better than any Google’s model before, but still needs many harness.”
Not everyone is sold. The thread’s most-replied comment came from a user for whom the latest Gemini models “just don’t seem as good as Opus 5.5 despite the benchmarks”: “Opus can run for days on a task and it feels like it understand what it was doing 2 days ago,” while other models “breakdown and get lost or falter as the task grows.” Another, who runs long orchestrated jobs, wrote that earlier Gemini models “looked good initially and in benchmarks, but on long-running sessions and orchestration, it’s fallen short,” and planned to try Neon.
What the tests show
On the independent tests published in its first day, Neon sits at or near the top, usually for less than half the cost per task of the Claude models beside it:
| Test | Who ran it | Gemini 4 Neon | Compared with |
|---|---|---|---|
| Coding Agent Index, in Antigravity’s CLI | Artificial Analysis | 59, at $6.86 a task | Claude Code with Fable 5.2 58 at $15.47, with Opus 5.5 58 at $12.61; Antigravity with Carbon 55 |
| Intelligence Index, xhigh effort | Artificial Analysis | 61, third overall | Behind GPT-6.2 Astra and Claude Fable 5.2; 1 point ahead of Gemini 4 Carbon |
| Vals Index | Vals AI | 70.87%, second, at $7.40 a test | Fable 5.2 71.31% at $17.94; Opus 5.5 70.14% at $24.77, fifth; GPT-6.2 Astra 67.66%, eighth |
| Agent Arena, real agent sessions | Arena | 1,333, third, at $2.55 median per task | Fable 5.2 1,448 at $9.13; Opus 5.5 1,336 at $6.00 |
| WebDev Arena | Arena | 1,764, seventh | Fable 5.2 1,824 first; Opus 5.5 second; GPT-6.2 Astra third |
| CursorBench 4.1, default effort | Cursor | 63.7% | Gemini 4 Hydrogen 57.3% |
Sources: Artificial Analysis and its post, the Vals Index, Arena’s leaderboards and Cursor, as each showed them on November 19, 2026.
Artificial Analysis says Neon in Antigravity “scores in line with” Claude Code on Fable 5.2 and Opus 5.5 “at a much lower Cost per Task,” and that at its default high effort it “matches Carbon (high) at less than half the Cost per Task.” Vals adds that Neon is first on its Finance and Legal indexes. The boards still move: at launch, Arena posted Neon second on its agent board at 1,336, ahead of Opus 5.5, and a day later it showed the two the other way round, three points apart.
Speed is the other half of the praise. Artificial Analysis measured about 110 output tokens a second, against about 57 for Carbon, and OpenRouter showed 41 to 113 across Google’s endpoints. A projection from Agent Duel, a head-to-head arena built on OpenCode, which assumed Carbon’s quality at Neon’s speed, found it “equivalent to Opus 5.5,” and called it “an incremental benchmark improvement with a very large speed improvement.”
The catches
- It thinks at length. Artificial Analysis found Neon generated 283 million tokens across its Intelligence Index, “very verbose in comparison to the median of 65.4M,” and “notably slow per task” at 21.5 minutes against a median of 12.4. On Copes’s short coding tasks it reasoned for 5,495 tokens where Opus used 694, so “Neon is cheaper per token and still about 3x more expensive per task here.” MindStudio’s hands-on warns that “extra-high reasoning mode can burn a large number of tokens even on simple tasks.”
- Its tokens cost more than some rivals’. Neon is $3 per million input tokens and $15 per million output, 25% below Opus 5.5 but, as VentureBeat notes, 50% above Claude Sonnet 5.5 and GPT-6.2 Sol. Until November 30, Artificial Analysis says, Carbon’s discounted $2 and $10 make it cheaper per token than Neon too.
- Long, open-ended work is still Claude’s reputation. Isenberg, the Hacker News skeptics and Arena’s WebDev board all still put Claude ahead somewhere. Copes hasn’t yet “run it as an agent on a real project,” and Theo has promised more on long jobs but hadn’t posted the results by November 19.
- Early capacity trouble. Chubby’s Antigravity logs before launch show “57 ‘no capacity’ errors, and 4 runs just stopped and I had to restart them,” which he put down to possible “pre-launch load.”
- The names. Simon Willison wrote that Google’s Gemini 4 pricing was already confusing, and “Neon has somehow made this even worse,” since “The Neon announcement pretends that Argon doesn’t exist.” In his pelican drawings, “the Neon pelicans do look better than Hydrogen and worse than Carbon.”
What Google has said
Google’s announcement claims Neon is “#3 overall with a score of 61 on the Artificial Analysis Intelligence Index,” which matches Artificial Analysis, and that at its default high thinking level it has “a 32% lower cost per task than Carbon.” Artificial Analysis found that gap narrows to about 17% when Neon runs at xhigh, because “Neon uses ~15% more output tokens per task.”
Two of Google’s claims go further than the boards themselves. It says Neon is “ranking #1 overall on the Artificial Analysis Coding Agent Index (58.6),” where Artificial Analysis calls it “in line with” Claude Code, and that it “debuts #2 globally” on Arena’s agent board, which had it third a day later. Sundar Pichai called it “frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.”
Who can use it
In Antigravity, Neon is on paid Google AI Pro and Ultra, by its pricing page, and its models page says “It’s the default model for new conversations if you haven’t chosen a model; if you’ve already selected one, Antigravity keeps your selection.” Antigravity’s free plan doesn’t include it. It’s also in the Gemini app for Pro and Ultra subscribers, in the Gemini API at $3 and $15 per million tokens, and in Cursor at the same rates. Our launch memo has the details, including GitHub Copilot’s plans.
What to do
- Try it if you’re on Pro or Ultra. It’s already Antigravity’s default, and Artificial Analysis’s numbers put its coding level with Claude Code’s at about half the cost per task.
- Match the thinking level to the job. For short coding tasks, Copes says to “drop it to
low,” and MindStudio says levels “need to be matched to task complexity rather than maxed out by default.” At Neon’s default high, Artificial Analysis found it matches Carbon at less than half the cost per task. - Give it a lot to read. On a search through 100,000 tokens of Markdown, Copes found “Neon cost 42% less for the same correct answer” than Opus 5.5, and that’s where he sees it fitting: “reading a lot and answering a little.”
- Compare cost per task, not price per token. Neon’s long thinking can make it cheaper or dearer than Opus depending on the job, as Copes’s two results show.
- Keep Claude for the jobs it’s proven on. For days-long, unattended work, the testers who prefer Opus or Fable haven’t changed their minds yet, and Theo’s own setup still gives implementation to Fable 5.2.
Neon in Enjoy
Enjoy runs Antigravity on your Google account, beside Claude Code, Codex and the other labs’ agents, each on its own plan, and Antigravity’s model list is the one Google offers your account. Theo’s pattern, a different model for each part of a job, works in one conversation: write with Neon in Antigravity, then switch to Claude Code to implement, and the new agent gets the conversation so far. Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t change what Neon costs or how it thinks, and each agent still needs its own plan.
We’ll update this memo as longer tests of Neon come in, including Theo’s promised results on long jobs.
8.Questions.
Is Gemini 4 Neon better than Claude Opus 5.5?
- On several new tests it’s level, for much less per task, but not everywhere. Artificial Analysis scores Antigravity with Neon at 59 on its Coding Agent Index, against 58 for Claude Code with Opus 5.5, at $6.86 a task to Opus’s $12.61, and Vals puts Neon second and Opus fifth. Fable 5.2 and Opus 5.5 still top Arena’s WebDev board, and some users find Opus better on long, open-ended work.
Is Gemini 4 Neon good for coding?
- Most early testers say yes. Chubby had it fix three broken programs until every test passed, Cursor measured 63.7% on its CursorBench 4.1, and one Hacker News user called it “the first google model I’ve used that feels good for coding.” Flavio Copes found it right on every task he gave it, but slower and costlier per task than Opus 5.5 on short ones.
Why does Gemini 4 Neon cost more on small tasks?
- Because it thinks at length. Flavio Copes found it spent 5,495 tokens reasoning on a CSV parser that Opus 5.5 did with 694, so it cost “about 3x more” per task despite cheaper tokens. Artificial Analysis calls it “very verbose,” and he and MindStudio both suggest a lower thinking level for simple tasks.
Is Gemini 4 Neon better than Gemini 4 Carbon?
- Slightly, and much faster. Artificial Analysis puts Neon at 61 on its Intelligence Index to Carbon’s 60, at about 110 output tokens a second to Carbon’s 57, and 3 points ahead on its Coding Agent Index. Until November 30, though, Carbon’s discounted API price is lower than Neon’s.
Where can I try Gemini 4 Neon?
- In Antigravity and the Gemini app on Google AI Pro and Ultra, and in the Gemini API at $3 per million input tokens and $15 per million output. Cursor has it at the same prices. Our launch memo covers who gets it and how to choose it.