Gemini 4 Carbon vs Claude Opus 5.5: benchmarks, cost per task and routing
Google opened Gemini 4 Carbon to all paid Gemini API developers on November 10, 2026 at $2 / $10 per million tokens ($4 / $20 from December 1). On KeelRouter, gemini-4-carbon costs $1 / $5.20 per million tokens next to $1.20 / $6.10 for Claude Opus 5.5. Below are the published benchmark scores, per-task token costs across effort settings, and when to route to each.
Our take: On KeelRouter, Gemini 4 Carbon is slightly cheaper per token ($1 / $5.20 vs $1.20 / $6.10) and edges out Opus 5.5 at max effort on the Artificial Analysis Intelligence Index (60 vs 59), Terminal-Bench-Science (64.8% vs 58.7%) and Chartography (72.5% vs 64.4%), while also accepting video and audio input. Opus 5.5 at medium or high effort finishes tasks in a third of the time (7.1–9.6 min vs 26.1 min), costs less per task than Carbon’s deep-thinking run, leads on FrontierSWE v2 (62.3% vs 61.2%) and OSWorld 2.1 (87.2% vs 81.5%), and plugs directly into Claude Code via /v1/messages.
Models: gemini-4-carbon, claude-opus-5-5. Prices from our live catalog; benchmark scores and token counts from Artificial Analysis, Google and Anthropic. Sources checked 2026-11-12.
Price on KeelRouter
| Option | Gemini 4 Carbon | Claude Opus 5.5 |
|---|---|---|
| input | $1 | $1.20 |
| output | $5.20 | $6.10 |
per 1M tokens. Prices follow our catalog and update when providers change theirs.
Score, time and cost per task across settings
Artificial Analysis evaluated Gemini 4 Carbon at high thinking and Claude Opus 5.5 across all five effort levels on the same suite. Note that Artificial Analysis calculated Carbon’s list cost at Google’s post-December $4 / $20 rate ($7.76 per task for 123,000 output tokens); at Google’s November $2 / $10 rate that same run is $3.88 at list prices.
| Model and setting | Intelligence Index | Cost per task (list) | Est. on KeelRouter | Time per task |
|---|---|---|---|---|
| Gemini 4 Carbon, high | 60 | $3.88 ($7.76 from Dec 1) | $2.02 | 26.1 min |
| Claude Opus 5.5, max | 59 | $10.04 | $3.06 | 24.7 min |
| Claude Opus 5.5, xhigh | 57 | $6.05 | $1.85 | 15.9 min |
| Claude Opus 5.5, high | 52 | $3.46 | $1.06 | 9.6 min |
| Claude Opus 5.5, medium | 48 | $2.45 | $0.75 | 7.1 min |
| Claude Opus 5.5, low | 35 | $0.98 | $0.30 | 2.9 min |
Coding, terminal and multimodal benchmarks
Across coding and agent evaluations reported on Anthropic’s launch table and Artificial Analysis, the two models trade leads depending on whether the task is repository software engineering or terminal/chart reasoning:
| Benchmark | Gemini 4 Carbon | Claude Opus 5.5 |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0.1 | 60 (high) | 59 (max) / 52 (high) |
| FrontierSWE v2 (Anthropic table) | 61.2% | 62.3% |
| FrontierCode 2.0 (Anthropic table) | 41.1% (medium) | 45.5% (xhigh) |
| Terminal-Bench-Science 0.1 | 64.8% | 58.7% |
| Chartography (no tools) | 72.5% | 64.4% |
| OSWorld 2.1 (offline set, partial credit) | 81.5% | 87.2% |
| Max output tokens per response | 65,536 | 128,000 |
| Supported input modalities | Text, image, audio, video, PDF | Text, image, PDF |
How to route between them on KeelRouter
When Gemini 4 Carbon wins on both quality and cost: if you were pushing Opus 5.5 to xhigh or max for hard debugging, scientific terminal tasks, or chart/document verification, Carbon at high reaches a higher Intelligence Index score (60 vs 59) for roughly $2.02 per index task on KeelRouter vs $3.06 for Opus 5.5 at max. Carbon is also the only choice of the two when the prompt includes video or audio files.
When Claude Opus 5.5 is the better workhorse: for daily coding, interactive agents, and Claude Code sessions, Opus 5.5 at medium ($0.75 est. per index task, 7.1 min) or high ($1.06, 9.6 min) responds much faster than Carbon (26.1 min per index task), leads on FrontierSWE v2 (62.3% vs 61.2%) and OSWorld 2.1 (87.2% vs 81.5%), outputs up to 128K tokens in a single call (vs 65K on Carbon), and supports explicit Anthropic prompt caching ($0.12 per 1M cache read tokens on KeelRouter).
Practical split: keep Claude Opus 5.5 at medium/high (or Gemini 4 Argon at $1 / $5.20) for fast interactive coding, and escalate stalled root-cause bugs, terminal automation, or multimodal video/PDF audits to gemini-4-carbon using the same KeelRouter API key.
Published head-to-head results
Gemini 4 Carbon (high) scores 60 with 123,000 output tokens and 26.1 min per task ($7.76 at $4/$20 list; $3.88 at $2/$10 list). Claude Opus 5.5 scores 59 at max ($10.04, 24.7 min) and 52 at high ($3.46, 9.6 min).
Artificial Analysis: Intelligence Index v5.0.1 (Carbon & Opus 5.5), 2026-11-12
Reports both Opus 5.5 and Gemini 4 Carbon across the same suite: FrontierSWE v2 62.3% (Opus 5.5) vs 61.2% (Carbon); Terminal-Bench-Science 58.7% vs 64.8%; Chartography 64.4% vs 72.5%; OSWorld 2.1 87.2% vs 81.5%.
Anthropic: Claude Fable 5.2 launch benchmark table, 2026-11-09
Opened Gemini 4 Carbon to all paid Gemini API developers at $2 / $10 per 1M tokens through November 30, 2026, with #1 results on CursorBench 4.0 and Perplexity WANDR, and 92% on its agentic honesty benchmark vs 70% for Gemini 4 Argon.
Google: Gemini 4 Carbon API availability announcement, 2026-11-10
When to use which
Gemini 4 Carbon
- Lower per-token price on KeelRouter: $1 / $5.20 per 1M tokens (52% of Google's $10 output list).
- Scores 60 on Artificial Analysis, edging out Opus 5.5 at max (59) for about one-third LESS cost per task on KeelRouter.
- Accepts text, image, audio, video and PDF input with a 1M context window.
- Strongest on Terminal-Bench-Science (64.8% vs 58.7%), Chartography (72.5% vs 64.4%) and multi-app automation.
- Thinks deeply by default (26.1 min per index task), so routine turns take longer than Opus 5.5 at medium/high.
- 65,536 max output tokens per call vs 128,000 on Opus 5.5.
- Supports implicit prefix caching only; explicit cache objects are not available.
Claude Opus 5.5
- Much faster and cheaper at medium and high effort (7.1–9.6 min per index task).
- Ahead on FrontierSWE v2 (62.3% vs 61.2%), FrontierCode 2.0 (45.5% vs 41.1%) and OSWorld 2.1 (87.2% vs 81.5%).
- 128K max output tokens and explicit Anthropic prompt caching on /v1/messages (works natively in Claude Code).
- At max effort, costs more per task than Gemini 4 Carbon while scoring 1 point lower on the Intelligence Index.
- No native audio or video input.
FAQ
- Is Gemini 4 Carbon better than Claude Opus 5.5?
- At the top end, Gemini 4 Carbon at high thinking scores 60 on the Artificial Analysis Intelligence Index against 59 for Claude Opus 5.5 at max effort, and leads on Terminal-Bench-Science (64.8% vs 58.7%) and Chartography (72.5% vs 64.4%). Claude Opus 5.5 leads on FrontierSWE v2 (62.3% vs 61.2%) and OSWorld 2.1 (87.2% vs 81.5%), and finishes routine tasks faster at medium or high effort.
- Which costs less on KeelRouter, Gemini 4 Carbon or Claude Opus 5.5?
- Per token, Gemini 4 Carbon is $1 / $5.20 per 1M input/output tokens and Claude Opus 5.5 is $1.20 / $6.10. Per task, Carbon at high costs about $2.02 on the Artificial Analysis suite against $3.06 for Opus 5.5 at max, while Opus 5.5 at medium ($0.75) is cheaper for everyday work.
- Can I call both Gemini 4 Carbon and Claude Opus 5.5 with one API key?
- Yes. Both run on POST https://api.keelrouter.com/v1/chat/completions with one KeelRouter API key and one prepaid USD balance; Claude Opus 5.5 also runs on POST https://api.keelrouter.com/v1/messages.
- When should I use Gemini 4 Argon instead of Carbon or Opus 5.5?
- Google recommends Gemini 4 Argon ($1 / $5.20 per 1M tokens on KeelRouter) for most everyday coding and knowledge tasks because it spends far fewer thinking tokens per call than Carbon.
More comparisons
- Wan 3.0 vs Seedance 2.5: one photo, one prompt, side by side
- Kling 4.0 vs Seedance 2.0 Mini: a pug takes the clapping cat’s place
- Kling 4.0 vs Higgsfield Genjutsu: does the swap respect height?
- Claude Fable 5.2 vs Opus 5.5: price per token vs cost per task
- Seedream 5.5 vs Seedream 5.0 Flash: editing curved, multi-line and RTL text in images
- Seedream 5.5 Layerize vs Seedream 5.0 Pro Layerize: occlusion order and 20+ Live2D parts
KeelRouter is an independent service and is not affiliated with the makers of these models. Model names are trademarks of their owners, used here only to name the model a request runs on. Benchmark and community figures are quoted from the linked sources as of the date above; your own tasks may cost and score differently.