GoldieBench Blog · 5 min read
Gemini 4 Carbon vs Opus 5.5: Every Shared Benchmark, Plus What GoldieBench Already Knows
Vendor-reported benchmarks for Gemini 4 Carbon and Claude Opus 5.5, cross-checked and set beside GoldieBench's own data.

Gemini 4 Carbon vs Opus 5.5 is the first real cross-lab test of Google's Gemini 4 family, and Google's own card shows Carbon ahead on five of the eight rows it published against Opus 5.5.
Book a free SEO strategy session →
GoldieBench has benched Opus 5.5 but not Carbon yet, so this page separates vendor numbers from our own.
To see how we test tools for SEO work, book a free SEO strategy session.
Vendor figures come from the Gemini 4 Carbon model card and the Anthropic Fable 5.2 post.
The vendor scoreboard
| Benchmark | Gemini 4 Carbon | Claude Opus 5.5 | Lead |
|---|---|---|---|
| AA Index v5.0 | 59 | 58 | Carbon |
| FrontierSWE 2 | 61.2% | 62.3% | Opus |
| Terminal-Bench 4.0 (AA) | 65.2% | 59.6% | Carbon |
| OSWorld-2.1 offline | 81.5% | 87.2% | Opus |
| AutomationBench | 53.9% | 42.5% | Carbon |
| Chartography | 72.5% | 66.3% | Carbon |
| GraphWalks 256K to 1M | 96.2% | 66.8% | Carbon |
| RiemannBench | 80.0% | 69.6% | Carbon |
Opus 5.5 holds the two rows most tied to coding agents in the wild: real code changes and computer control.
Carbon's biggest gap is on very long context, close to thirty points.
Cross-checked rows
Anthropic's own table, published with Fable 5.2, prints the same FrontierSWE v2, AutomationBench and OSWorld 2.1 numbers for both models.
Anthropic also adds FrontierCode 2.0, where Opus 5.5 posts 45.5% and Carbon 41.1%.
Rows that both labs agree on are the strongest evidence in this comparison.
What GoldieBench already shows for Opus 5.5
On the GoldieBench leaderboard, Claude Opus 5.5 averages 7.57 out of 10 across 50 one-shot builds, every one scored, with every game playtested.
That puts it at #13 on the overall board today.
Gemini 4 Carbon has no runs yet, so there is no bench-on-bench result to report.
Settings behind the numbers
- Both default to medium reasoning:
thinking_levelon Carbon,efforton Opus 5.5. - Opus 5.5's adaptive thinking is always on.
- Carbon spends more thinking tokens per turn than Argon on hard problems, which Google says increases latency.
- Vendor rows usually reflect each model's best setting, not its default.
Book a free SEO strategy session →
How a fair bench run would work
GoldieBench gives every model the same fixed prompt and takes the first HTML file it produces, with no retries and no best-of-N.
For this pair, that means matched medium settings, one shot each, scored 0-10 on whether it ran, how close it hit the brief and how it looked.
Build time is logged next to each score, which matters for a model Google calls deliberate.
Model choice by task
| Task | Lean towards | Basis |
|---|---|---|
| One-shot web builds | Opus 5.5 (benched) | GoldieBench 7.57 average |
| Long-context analysis | Carbon | GraphWalks lead |
| Desktop agents | Opus 5.5 | OSWorld-2.1 lead |
| App automation | Carbon | AutomationBench lead |
Where Opus 5.5 sits on the bench
Opus 5.5's 7.57 average places it behind Anthropic's own Claude Opus 5, which averages 8.27 at #2, and Claude Fable 5 at 8.10.
Its best category on our board is Sims.
That context matters: Opus 5.5 is a strong builder, but it is not the top Anthropic model on one-shot builds.
Cost and speed signals
From 1 December 2026, Carbon's list price per token matches Opus 5.5's, after Google's introductory rate ends on 30 November.
Google labels Carbon "deliberate"; Anthropic lists Opus 5.5 at moderate latency.
On a one-shot bench, build time per task is the practical speed measure, and we log it next to every score.
What we expect to check first
- Whether Carbon's Terminal-Bench lead shows up as cleaner one-shot builds.
- Whether Opus 5.5's computer-use lead matters for playtested games.
- How long Carbon takes per build at medium thinking.
- Whether Carbon's long decode helps on tasks with large briefs.
Reading vendor cards carefully
Google's Carbon card notes that its cybersecurity rows reflect access through its Fairwind Program, which is not the standard API.
Anthropic's numbers for Opus 5.5 in the Fable 5.2 post show results with safeguards on unless noted.
Book a free SEO strategy session →
Small print like this is why we prefer one fixed prompt, one shot and a public demo for every score.
Where the rest of the field sits
Anthropic's Fable 5.2 table also lists GPT-6 Astra, which scores 65.5% on FrontierSWE v2, ahead of both Opus 5.5 and Carbon on that row.
On Google's card, GPT-6 Astra lands between Carbon and Opus 5.5 on the AA Index, at 57.
So the Carbon vs Opus fight is close, and neither model is the clear frontier leader on every coding measure.
What this means for picking a builder model
If you need a model for one-shot builds today, Opus 5.5 is the one with a public GoldieBench record and playable demos.
If your builds depend on large briefs or long outputs, Carbon's published long-context and long-decode strengths are worth a trial.
We will add Carbon's runs to the matrix so the two can be compared on the same fixed prompts.
Specs that affect builds
Carbon's release guide gives it a 1,048,576-token unified window and 65,536-token default output, extendable with long decode.
Opus 5.5's model page lists a 1M window and 128K max output, with adaptive thinking always on.
For our single-HTML-file tasks, neither output limit is the bottleneck; reasoning quality and design taste are.
Those are exactly what one-shot scoring captures and vendor benchmarks do not.
Verdict for now
On vendor numbers Carbon leads more rows, but Opus 5.5 leads the rows closest to coding agents, and it is the one with a GoldieBench record.
For agent setups that use both, the AI Profit Boardroom has the trainings and the Agent OS.
Book a free SEO strategy session →
Trust the rows both labs print and Gemini 4 Carbon vs Opus 5.5 comes into focus; for data-led SEO, book a free SEO strategy session.
FAQ
Has GoldieBench tested Gemini 4 Carbon?
Not yet. Opus 5.5 is benched at a 7.57 average over 50 one-shot builds; Carbon runs are pending.
Which leads on Google's card?
Carbon leads five of the eight rows Google published against Opus 5.5; Opus leads FrontierSWE 2 and OSWorld-2.1, and the AA Index is within one point.
Do Google and Anthropic agree on any numbers?
Yes. Both publish the same FrontierSWE v2, AutomationBench and OSWorld 2.1 numbers for these two models.
Which is better for long context?
Carbon, at 96.2% on GraphWalks 256K to 1M against Opus 5.5's 66.8% on Google's card.
- 📺 Video notes + links to the tools
- 🆓 Get a FREE AI course + community + 1,000 AI agents
- 🎥 Learn how I make these videos


