Anthropic released Claude Opus 5.5 at the end of September 22. It promises Fable-level performance on most work at 40% less than Opus 5. Two months later, yesterday, Google released Gemini 4 Neon, a mid-tier model adapted from Gemini 4 Carbon
Google was very direct in the Neon release about comparing Neon to Opus 5.5. It was a central focus of the release. The benchmarks all show Neon against Opus 5.5 by name, and Neon's list price is lower on every token.
Before we dive into the comparion, I recommend subscribing to our weekly newsletter, The Median, which keeps you up to speed on the biggest model releases.
TL;DR
On Google's model card, Neon leads Opus 5.5 on most benchmarks, with one of its two largest leads on long-context reasoning.
Opus 5.5 keeps a slight edge on terminal coding and merge-ready code changes.
In our chancellor chess test, both wrote a correct rule set every time, but only Neon's engines won every game.
Neon costs 25% less per token, and each of its test runs cost about half as much.
Choose Neon for long documents, media and cost-sensitive agents; choose Opus 5.5 for terminal-heavy coding, Claude Code and the widest cloud reach.
What Is Gemini 4 Neon?
Gemini 4 Neon is Google's mid-tier Gemini 4 model, released in November 2026 and adapted from Gemini 4 Carbon.
Google pitches it as frontier-level intelligence with fast end-to-end task completion: it finishes complex software, knowledge and science tasks in fewer steps than Carbon, with four thinking levels to trade depth for speed.
Claude Opus 5.5 is Anthropic's flagship for long-running agentic coding and knowledge work, released in September 2026.
Its selling point is efficiency: Anthropic says it performs at the level of Claude Fable 5.1 on most work while using fewer tokens per task than Opus 5, and it is the model Anthropic tells developers to start with.
Gemini 4 Neon vs Claude Opus 5.5: Head-to-Head Comparison
Every head-to-head score below comes from Google's Neon model card, the only source that tests both models on the same benchmarks. Neon leads 16 of the 18 rows where both have a score, and Opus 5.5 leads the other two by about a point each.
Feature
Gemini 4 Neon
Claude Opus 5.5
Developer and release
Google, November 18, 2026
Anthropic, September 22, 2026
Context window
1,048,576 tokens, shared by input, thinking and output
1M tokens
Synchronous output limit
64K tokens (up to 1M with long decode)
128K tokens
Reasoning control
Thinking levels: low, medium, high (default), xhigh
Effort: low, medium (default), high, xhigh, max
GraphWalks (256K to 1M, BFS F1)
95.2%
66.8%
CursorBench 4.1
65.4%
58.1%
Terminal-bench 4.1
65.2%
65.8%
HLE-Diamond
71.3%
67.9%
Price per token
25% lower on input, output and cached reads
Baseline
Long-context and knowledge work
Long-context reasoning shows one of the two widest gaps in the comparison, and the only one that holds up without a harness caveat. On GraphWalks, which asks a model to traverse a graph spread across 256K to 1M tokens of input, Neon scores 95.2% against Opus 5.5's 66.8%. Both models accept about a million tokens, but on Google's numbers, Neon is far better at using the far end of that window.
Benchmark
Gemini 4 Neon
Claude Opus 5.5
Notes
GraphWalks (256K to 1M, BFS F1)
95.2%
66.8%
Long-context graph traversal
AA Index v5.0.1
61
59
Artificial Analysis composite index
Vals Index v2.2
70.9%
70.1%
Composite index
Agents' Last Exam v1.1
75.8%
71.6%
Agentic tasks
MKA (Climb WSR)
56.5%
53.0%
Reported as a win rate
The broad indexes are close: a two-point lead on the AA Index and under a point on the Vals Index are the kind of margins Anthropic itself says have become a less reliable guide to real-world differences. Anthropic's own showcase for Opus 5.5 is knowledge work, led by a GDPval-AA v2.1 score of 1846 Elo, but Google's card does not report GDPval, so that claim has no Neon counterpart.
For long documents, contracts and codebases loaded whole into context, Neon is the stronger pick on the published evidence.
Coding and agentic workflows
Coding is the closest dimension. Opus 5.5 edges Neon on terminal work and merge-ready code changes, while Neon leads by about 7 points on CursorBench 4.1, which draws ambiguous multi-file tasks from real Cursor sessions, and by about 14 points on Zapier's AutomationBench.
Benchmark
Gemini 4 Neon
Claude Opus 5.5
Notes
CursorBench 4.1
65.4%
58.1%
Multi-file tasks from Cursor sessions
Terminal-bench 4.1
65.2%
65.8%
Cheating-adjusted accuracy
FrontierCode 2.0
46.2%
47.3%
Whether the change would be merged
SWE-sweep (provider harness)
39.0%
10.1%
12.8% vs 3.6% on the Mini-SWE harness
AutomationBench
56.3%
42.5%
Business workflows across connected apps
Treat these coding numbers with some caution:
Google ran Opus 5.5 for its own model card, and Anthropic has not published Neon scores.
Anthropic reports Opus 5.5 on older versions of the same suites (Terminal-Bench 4.0, CursorBench 4.0, FrontierCode v1.1), so its figures cannot be set against Google's.
The SWE-sweep scores swing by a factor of three between harnesses, so that row says more about scaffolding than about either model.
For agentic coding, call it a draw on the benchmarks, with Opus 5.5 slightly ahead in the terminal and Neon ahead in an editor-style workflow.
Science and reasoning
Neon leads all three science and reasoning rows. The clearest gap is Terminal-Bench-Science 0.1, an agentic research benchmark run in a command line, where Neon scores 70.5% against 63.3%. Neon also leads on RiemannBench (84.0% vs 80.0%) and HLE-Diamond. The gaps are steady rather than dramatic, but they all point the same way, so Neon is the better default for quantitative and scientific workflows.
Multimodal understanding
Neon leads on video, with 89.0% on LVBench against 83.7%, and on chart reading, with 73.5% on Chartography against 66.3%. The chart result comes with a caveat: Anthropic reports 89.0% for Opus 5.5 on Chartography with tools, a different setup from Google's run, so the gap depends on whether tools are allowed.
Cybersecurity
Neon leads both shared security rows: 70.7% against 67.1% on CWE-Bench v2 and 51.3% against 37.1% on WebExploitBench v1.2. Both vendors gate their strongest offensive capabilities behind verification programs, Google's Fairwind Program and Anthropic's Cyber Verification Program, so for defenders the deciding factor is which program they qualify for rather than a few benchmark points.
Pricing: what you actually pay
Neon's rates are exactly 0.75x Opus 5.5's on input, output and cached reads, so the whole pricing comparison comes down to one number: Neon is 25% cheaper per token.
Token rates side by side
Rate
Gemini 4 Neon
Claude Opus 5.5
Input, per 1M tokens
$3.00
$4.00
Output, per 1M tokens
$15.00
$20.00
Cached input read, per 1M tokens
$0.15
$0.20
Cache write, per 1M tokens
Not applicable (implicit caching only)
$5.00 (5-minute cache)
Batch discount
50%
50%
Long-context surcharge
None across the 1M window
None across the 1M window
Consumer access
Google AI Pro and Ultra
Claude Pro, Max, Team and Enterprise
Because the multiple is the same on input and output, no workload shape flips the result. Neon's output price includes thinking tokens, so a higher thinking level raises the bill at the $15 rate. The one structural difference is caching: Opus 5.5 lets you write explicit cache entries at a premium, while Neon caches implicitly only, so its cached-read rate applies only when Google's cache hits.
What a real workload costs
Workload (monthly)
Gemini 4 Neon
Claude Opus 5.5
Difference
Balanced assistant: 1M in / 250K out
$6.75
$9.00
Neon $2.25 less (25%)
Generation-heavy: 1M in / 4M out
$63.00
$84.00
Neon $21 less (25%)
Cache-heavy loop: 100K prefix written once, 1,000 cached reads, 5K in / 1K out per request
$45.30
$60.50
Neon $15.20 less (25%)
Each total is (volume ÷ 1M) × rate, summed across input and output at standard rates. The cache-heavy row assumes every one of the 1,000 reads hits the cache; Opus 5.5 guarantees that with an explicit cache write, while Neon's implicit cache does not. The two models use different tokenizers and neither vendor has published token counts for a shared workload, so treat these totals as a rate comparison rather than a bill.
How Gemini 4 Neon and Claude Opus 5.5 Performed
To compare the models, I ran one build with an identical prompt and identical tooling.
My ask: build a single web page that plays a complete game of chess against itself. This time, I asked for a chess variant called Chancellor chess, where each side's queen is replaced by a chancellor, which a piece that moves like a rook or like a knight. (The Chancellor is also sometimes called an Empress or Elephant.)
At any rate, each model had to write the rules and a search engine from scratch, draw the board and the new piece, keep a move list in algebraic notation, and finish with the correct result.
I liked the idea of asking for a chess variant because the model would not be able to fall back on chess code it has memorized. The chancellor changes the rules in places that are easy to miss, like, for example:
from the starting position, the chancellor can leap out to c3 or e3, so White has 22 legal first moves instead of 20;
a chancellor's knight leap can give check, or attack a square the king needs to castle through;
an en passant capture is illegal when it would expose the king to a chancellor on the same rank;
pawns promote to a chancellor, rook, bishop or knight, never a queen.
The prompt I used named the variant's rules but none of these harder-to-see edge cases.
Here is how Neon's game ended, with king, chancellor, and two pawns mating a lone king. It's interesting to see because you were expect to see white's final piece on G7, but in this variant that would be a stalemate, so it chose correctly.
Opus 5.5 drew harder-to-see pieces. Black finished a knight and a pawn ahead, but White's chancellor checked the black king from square to square. Black's engine never found a way out, so the game ended in a draw:
The main findings:
Both models know the rules, chancellor included. I ran the simulation two dozen times. There was never an illegal move and I never saw a notation error, either, although I didn't check the notation for every move.
Neon's engines finish games, while Opus 5.5's often settle for a draw. Oddly, all 12 of Neon's games ended in checkmate, but only half of Opus 5.5's games ended in checkmate.
Neon drew the new piece more clearly. All three Neon pages drew the chancellor as one composite piece, a knight's head on a rook. Two of the three Opus 5.5 pages squeezed a small rook and a small knight side by side into one square, which I thought looked like a mistake.
The two models built the page differently. Opus 5.5 planned the whole program and wrote it in a single response. Neon wrote it over five steps and produced a larger file each time.
That perpetual check is good defense by White's engine, but a side that is a piece up should not allow it.
Measure
Gemini 4 Neon
Claude Opus 5.5
Games ending in checkmate (of 12)
12
7
Draws (of 12)
0
5
Steps per run
5
1
Cost per run (USD)
$0.90
$1.70
Neon wins this test. Neon's engines turned every game into a checkmate, and each Neon run cost about half as much, plus Neon's pieces were easier to see. Maybe Demis Hassabis, Google DeepMind's CEO, had a hand in it. (Demis was a child chess prodigy.)
That said, Opus 5.5's stalemate example was clever and not easy to find.
When to Choose Gemini 4 Neon vs Claude Opus 5.5
The fork is usually context length and price. Neon's long-context lead matters most when your inputs run past 256K tokens, and its 25% lower rates matter most at volume.
Choose Gemini 4 Neon if...
You load whole documents or repositories into context. Neon's long-context lead is one of the two largest gaps on Google's card.
You work with video, charts or scientific data. Neon leads every multimodal and science row Google reports for both models.
You run agents at high volume. Every token costs 25% less, and in our test each Neon run cost about half as much as an Opus 5.5 run.
You build on Google's stack. Neon is in Google AI Studio, Vertex AI and Google Antigravity, and pairs with Hydrogen in multi-model workflows.
Choose Claude Opus 5.5 if...
Your coding happens in the terminal or in Claude Code. Opus 5.5 holds a slight lead on Terminal-bench 4.1 and FrontierCode 2.0, and Claude Code is Anthropic's own agent.
You need long answers in one response. Opus 5.5 returns up to 128K tokens synchronously, twice Neon's default limit, without switching to background long decode.
You deploy on AWS or Azure. Opus 5.5 is on Amazon Bedrock and Microsoft Foundry as well as Vertex AI.
You want predictable caching. Explicit cache writes let you guarantee the cheaper cached-read rate on a repeated prefix.
How to Get Started With Gemini 4 Neon and Claude Opus 5.5
Surface
Gemini 4 Neon
Claude Opus 5.5
Consumer app
Gemini app (Google AI Pro and Ultra)
Claude apps (Pro, Max, Team, Enterprise)
First-party API
Gemini API (paid tier), Google AI Studio
Claude API
Cloud platforms
Vertex AI
Amazon Bedrock, Vertex AI, Microsoft Foundry
Coding agents
Google Antigravity, GitHub Copilot, Cursor
Claude Code, GitHub Copilot, Cursor
Third-party routers
OpenRouter
OpenRouter
API model ID
gemini-4-neon
claude-opus-5-5
The biggest access difference is cloud reach: Opus 5.5 runs on Amazon Bedrock, Vertex AI and Microsoft Foundry, while we found Neon only on Google Cloud. Neon is not listed in the Gemini CLI model docs as of November 19, 2026. The native model IDs are gemini-4-neon and claude-opus-5-5; on OpenRouter they are google/gemini-4-neon and anthropic/claude-opus-5.5.
Making your first API call
The two native SDKs differ, so the quickest way to compare them on your own prompts is OpenRouter's OpenAI-compatible endpoint, where switching models is a one-string change:
from openai import OpenAI
client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key="YOUR_OPENROUTER_KEY")
response = client.chat.completions.create(
model="google/gemini-4-neon",# or "anthropic/claude-opus-5.5"
messages=[{"role":"user","content":"Refactor this module and explain each change."}],)print(response.choices[0].message.content)Powered by
On Google's own benchmarks, Gemini 4 Neon beats Claude Opus 5.5 on most tests at a lower price, and our chess test pointed the same way: both models got every rule right, but only Neon's engines closed out every game, at about half the cost per run. The coding benchmarks are close enough that Opus 5.5 remains a reasonable pick for terminal work.
My recommendation: if you work with long inputs, media or high-volume agents, try Neon first. If your team already works in Claude Code, or deploys on AWS or Azure, Opus 5.5 is still the more convenient choice, and the quality gap is small. And if you want to create chess variant simulations on your own computer (unlikely, I know), go with Neon.
On Google's model card, Gemini 4 Neon beats Claude Opus 5.5 on 16 of 18 shared benchmarks, with large leads in long-context reasoning and one coding benchmark, and it costs 25% less per token. Opus 5.5 keeps a slight edge on terminal coding. In our chancellor chess test, both got every rule right, but only Neon's engines won every game.
How much cheaper is Gemini 4 Neon than Claude Opus 5.5?⌄
Gemini 4 Neon costs $3 per million input tokens and $15 per million output tokens, against $4 and $20 for Claude Opus 5.5. Cached reads are $0.15 against $0.20. That makes Neon 25% cheaper on every token type, so any workload costs about a quarter less at standard rates.
Which is better for coding, Gemini 4 Neon or Claude Opus 5.5?⌄
They are close. On Google's model card, Opus 5.5 leads slightly on Terminal-bench 4.1 and FrontierCode 2.0, while Neon leads on CursorBench 4.1 and AutomationBench. In our hands-on test, both wrote a correct chess variant from scratch every time; Neon's engines played stronger chess and its runs cost about half as much.
Where can I use Gemini 4 Neon and Claude Opus 5.5?⌄
Gemini 4 Neon is in the Gemini API, Google AI Studio, Vertex AI, the Gemini app for Google AI Pro and Ultra subscribers, Google Antigravity, GitHub Copilot and Cursor. Claude Opus 5.5 is in the Claude API and apps, Claude Code, Amazon Bedrock, Vertex AI, Microsoft Foundry, GitHub Copilot and Cursor. Both are on OpenRouter.
What are the API model IDs for Gemini 4 Neon and Claude Opus 5.5?⌄
On the native APIs, the model IDs are gemini-4-neon for Google's Gemini API and claude-opus-5-5 for Anthropic's Claude API. On OpenRouter, they are google/gemini-4-neon and anthropic/claude-opus-5.5, so you can switch between them by changing one string in an OpenAI-compatible client.
Author
Josef Waples
I'm a data science editor with contributions to research articles in scientific journals. I'm especially interested in linear algebra, statistics, R, and the like.
Gemini 4 Argon leads Google's benchmark table on knowledge work and long context, while Claude Opus 5.5 holds terminal coding and won our hands-on test.
Learn about Gemini 3.1 Pro, Google's latest reasoning model. Explore its features, benchmarks, hands-on tests, and how it compares to Claude Opus 4.6, Claude Sonnet 4.6, and GPT-5.2.
Compare Claude Opus 4.8 and Gemini 3.5 Flash on MCP Atlas, SWE-bench Pro, and GDPval benchmarks, plus pricing and speed, to find the right model for your work.
Both Gemini 4 Hydrogen and Claude Sonnet 5.5 aced our hands-on robot-arm test, so the choice comes down to everything else: Hydrogen leads Google's benchmark table and costs 25% less, while Sonnet 5.5 stays ahead on terminal-based coding and computer use.
Learn how to build, configure, and share Skills in Claude Code — reusable markdown instructions that Claude automatically applies to tasks at the right time.
Anthropic released Claude Opus 5.5 at the end of September 22. It promises Fable-level performance on most work at 40% less than Opus 5. Two months later, yesterday, Google released Gemini 4 Neon, a mid-tier model adapted from Gemini 4 Carbon
Google was very direct in the Neon release about comparing Neon to Opus 5.5. It was a central focus of the release. The benchmarks all show Neon against Opus 5.5 by name, and Neon's list price is lower on every token.
Before we dive into the comparion, I recommend subscribing to our weekly newsletter, The Median, which keeps you up to speed on the biggest model releases.
TL;DR
What Is Gemini 4 Neon?
Gemini 4 Neon is Google's mid-tier Gemini 4 model, released in November 2026 and adapted from Gemini 4 Carbon.
Google pitches it as frontier-level intelligence with fast end-to-end task completion: it finishes complex software, knowledge and science tasks in fewer steps than Carbon, with four thinking levels to trade depth for speed.
Our Gemini 4 Carbon guide covers the model Neon was adapted from, and our Gemini 4 Hydrogen API tutorial walks through the shared Gemini 4 API.
What Is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic's flagship for long-running agentic coding and knowledge work, released in September 2026.
Its selling point is efficiency: Anthropic says it performs at the level of Claude Fable 5.1 on most work while using fewer tokens per task than Opus 5, and it is the model Anthropic tells developers to start with.
Our Claude Opus 5.5 guide covers the launch, and our Opus 5.5 API tutorial builds an AI incident investigator with it.
Gemini 4 Neon vs Claude Opus 5.5: Head-to-Head Comparison
Every head-to-head score below comes from Google's Neon model card, the only source that tests both models on the same benchmarks. Neon leads 16 of the 18 rows where both have a score, and Opus 5.5 leads the other two by about a point each.
Long-context and knowledge work
Long-context reasoning shows one of the two widest gaps in the comparison, and the only one that holds up without a harness caveat. On GraphWalks, which asks a model to traverse a graph spread across 256K to 1M tokens of input, Neon scores 95.2% against Opus 5.5's 66.8%. Both models accept about a million tokens, but on Google's numbers, Neon is far better at using the far end of that window.
The broad indexes are close: a two-point lead on the AA Index and under a point on the Vals Index are the kind of margins Anthropic itself says have become a less reliable guide to real-world differences. Anthropic's own showcase for Opus 5.5 is knowledge work, led by a GDPval-AA v2.1 score of 1846 Elo, but Google's card does not report GDPval, so that claim has no Neon counterpart.
For long documents, contracts and codebases loaded whole into context, Neon is the stronger pick on the published evidence.
Coding and agentic workflows
Coding is the closest dimension. Opus 5.5 edges Neon on terminal work and merge-ready code changes, while Neon leads by about 7 points on CursorBench 4.1, which draws ambiguous multi-file tasks from real Cursor sessions, and by about 14 points on Zapier's AutomationBench.
Treat these coding numbers with some caution:
For agentic coding, call it a draw on the benchmarks, with Opus 5.5 slightly ahead in the terminal and Neon ahead in an editor-style workflow.
Science and reasoning
Neon leads all three science and reasoning rows. The clearest gap is Terminal-Bench-Science 0.1, an agentic research benchmark run in a command line, where Neon scores 70.5% against 63.3%. Neon also leads on RiemannBench (84.0% vs 80.0%) and HLE-Diamond. The gaps are steady rather than dramatic, but they all point the same way, so Neon is the better default for quantitative and scientific workflows.
Multimodal understanding
Neon leads on video, with 89.0% on LVBench against 83.7%, and on chart reading, with 73.5% on Chartography against 66.3%. The chart result comes with a caveat: Anthropic reports 89.0% for Opus 5.5 on Chartography with tools, a different setup from Google's run, so the gap depends on whether tools are allowed.
Cybersecurity
Neon leads both shared security rows: 70.7% against 67.1% on CWE-Bench v2 and 51.3% against 37.1% on WebExploitBench v1.2. Both vendors gate their strongest offensive capabilities behind verification programs, Google's Fairwind Program and Anthropic's Cyber Verification Program, so for defenders the deciding factor is which program they qualify for rather than a few benchmark points.
Pricing: what you actually pay
Neon's rates are exactly 0.75x Opus 5.5's on input, output and cached reads, so the whole pricing comparison comes down to one number: Neon is 25% cheaper per token.
Token rates side by side
Because the multiple is the same on input and output, no workload shape flips the result. Neon's output price includes thinking tokens, so a higher thinking level raises the bill at the $15 rate. The one structural difference is caching: Opus 5.5 lets you write explicit cache entries at a premium, while Neon caches implicitly only, so its cached-read rate applies only when Google's cache hits.
What a real workload costs
Each total is (volume ÷ 1M) × rate, summed across input and output at standard rates. The cache-heavy row assumes every one of the 1,000 reads hits the cache; Opus 5.5 guarantees that with an explicit cache write, while Neon's implicit cache does not. The two models use different tokenizers and neither vendor has published token counts for a shared workload, so treat these totals as a rate comparison rather than a bill.
How Gemini 4 Neon and Claude Opus 5.5 Performed
To compare the models, I ran one build with an identical prompt and identical tooling.
My ask: build a single web page that plays a complete game of chess against itself. This time, I asked for a chess variant called Chancellor chess, where each side's queen is replaced by a chancellor, which a piece that moves like a rook or like a knight. (The Chancellor is also sometimes called an Empress or Elephant.)
At any rate, each model had to write the rules and a search engine from scratch, draw the board and the new piece, keep a move list in algebraic notation, and finish with the correct result.
I liked the idea of asking for a chess variant because the model would not be able to fall back on chess code it has memorized. The chancellor changes the rules in places that are easy to miss, like, for example:
The prompt I used named the variant's rules but none of these harder-to-see edge cases.
Here is how Neon's game ended, with king, chancellor, and two pawns mating a lone king. It's interesting to see because you were expect to see white's final piece on G7, but in this variant that would be a stalemate, so it chose correctly.
Opus 5.5 drew harder-to-see pieces. Black finished a knight and a pawn ahead, but White's chancellor checked the black king from square to square. Black's engine never found a way out, so the game ended in a draw:
The main findings:
That perpetual check is good defense by White's engine, but a side that is a piece up should not allow it.
Neon wins this test. Neon's engines turned every game into a checkmate, and each Neon run cost about half as much, plus Neon's pieces were easier to see. Maybe Demis Hassabis, Google DeepMind's CEO, had a hand in it. (Demis was a child chess prodigy.)
That said, Opus 5.5's stalemate example was clever and not easy to find.
When to Choose Gemini 4 Neon vs Claude Opus 5.5
The fork is usually context length and price. Neon's long-context lead matters most when your inputs run past 256K tokens, and its 25% lower rates matter most at volume.
Choose Gemini 4 Neon if...
Choose Claude Opus 5.5 if...
How to Get Started With Gemini 4 Neon and Claude Opus 5.5
The biggest access difference is cloud reach: Opus 5.5 runs on Amazon Bedrock, Vertex AI and Microsoft Foundry, while we found Neon only on Google Cloud. Neon is not listed in the Gemini CLI model docs as of November 19, 2026. The native model IDs are
gemini-4-neonandclaude-opus-5-5; on OpenRouter they aregoogle/gemini-4-neonandanthropic/claude-opus-5.5.Making your first API call
The two native SDKs differ, so the quickest way to compare them on your own prompts is OpenRouter's OpenAI-compatible endpoint, where switching models is a one-string change:
On the native APIs, Neon's reasoning depth is set with
thinking_leveland Opus 5.5's witheffort. Our Opus 5.5 API tutorial and Gemini 4 Hydrogen API tutorial cover each SDK in full.Final Thoughts
On Google's own benchmarks, Gemini 4 Neon beats Claude Opus 5.5 on most tests at a lower price, and our chess test pointed the same way: both models got every rule right, but only Neon's engines closed out every game, at about half the cost per run. The coding benchmarks are close enough that Opus 5.5 remains a reasonable pick for terminal work.
My recommendation: if you work with long inputs, media or high-volume agents, try Neon first. If your team already works in Claude Code, or deploys on AWS or Azure, Opus 5.5 is still the more convenient choice, and the quality gap is small. And if you want to create chess variant simulations on your own computer (unlikely, I know), go with Neon.
To build with Anthropic's models, try our Introduction to Claude Models course.