Google DeepMind announced Gemini 4 Carbon on October 28, 2026, and within a day it sat at the top of the Artificial Analysis Intelligence Index v5 with a score of 59. That is one point above Claude Opus 5.5 at 58. The index was rebuilt on October 27 around mostly private evaluations, so Carbon took the lead one day after Opus 5.5 held it.
This post answers the question people type next: is Gemini 4 Carbon actually better than Claude Opus 5.5, GPT-6 Astra and Grok 4.7, and can I use it? We compare the four on independent scores first, Google's own claims second, then price, hallucination, coding agents and access. Every figure names its source, and where a lab's table and an independent run disagree we show both.
TL;DR: Carbon vs the field #
| Question | Direct answer |
|---|---|
| **Who leads the Intelligence Index v5?** | Carbon 59, Opus 5.5 (max) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55 |
| Who leads coding agents? | Opus 5.5 in Claude Code, 62 vs Carbon 59 in Antigravity CLI | | Cheapest per Index task (standard price)? | Astra $6.25, Carbon $7.76, Opus 5.5 $10.04 | | List price per 1M tokens (in / out) | Carbon $2 / $10 intro, then $4 / $20. Opus 5.5 $4 / $20. Astra $10 / $50. Grok 4.7 $2 / $6 | | Hallucination rate (AA-Omniscience) | Carbon 11%, versus 50% reported for Opus 5.5 and Astra | | Can I call Carbon today? | No public ID. Selected cyber defenders only, expanding through Fairwind | | Can I call the other three today? | Yes | | Biggest caveat | Carbon's cyber headline numbers are Google's own, run without safety guardrails |
What does the independent scoreboard say? #
Start with the evaluators who do not sell a model. Artificial Analysis, Vals AI and Arena all published Carbon results within two days of the announcement.
| Evaluation | Carbon | Claude Opus 5.5 | GPT-6 Astra | Source |
|---|---|---|---|---|
| Intelligence Index v5 | **59** (1st of 52) | 58 (max) | 57 (max) | Artificial Analysis |
| Coding Agent Index | 59 (Antigravity CLI) | 62 (Claude Code) | not stated | Artificial Analysis | | Terminal-Bench 4.0 (AA run) | 64% (2nd) | not stated | not stated | Artificial Analysis | | FrontierCode v1.1 | 51% | 56% (max) | not stated | Artificial Analysis | | Vals Index v2.2 | 70.36% (2nd) | 70.18% | not stated | Vals AI | | Text Arena | 1,528 | not stated | not stated | Arena | | Agent Arena net improvement | +16.2% (2nd) | +21.1% (1st) | +13.4% | Arena |
Two patterns show up. Carbon wins on broad composite scores: the Intelligence Index, Text Arena, and several Vals sub-boards (it leads the Finance Index and Legal Research benchmark). Opus 5.5 wins where an agent has to finish long, tool-heavy work: the Coding Agent Index, FrontierCode and Agent Arena.
Artificial Analysis also says Carbon holds the top score in four of the index's nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It does not lead coding or knowledge work. That split matters more than the headline 59.
Vals AI puts Claude Sonnet 5.5 first on its Index at 70.85%, so a Sonnet-class model belongs in your bake-off too. Our note on Sonnet 5.5's cache-read price cut covers why that matters for agent loops.
Carbon vs Claude Opus 5.5: a one-point lead, three points behind on coding #
The Intelligence Index gap is one point. Artificial Analysis itself describes the lead as small, and the index mixes nine evaluations with roughly 60% of its weight on private or held-out data, so one point is within what a rerun could move.
Where the two diverge:
- Coding agents. Claude Code with Opus 5.5 scores 62 on the Coding Agent Index. Carbon in Antigravity CLI scores 59. Carbon was measured in a single harness, so other harnesses are untested.
- Terminal-Bench 4.0. Artificial Analysis ranks Carbon second at 64%. Google's own self-reported figure is 65.2%, which is below Opus 5.5's 66.4% on the same sheet, as one developer pointed out. Different harnesses give different orderings, which is why we ran a separateTerminal-Bench 4.0 audit .
- Token use. Carbon uses about 123,000 output tokens per Index task, 29% fewer than Opus 5.5's 173,000. That is why its cost per task is lower even though list prices are close.
- Agent steerability. Arena reports the highest steerability score for Carbon (+18.6%) and a median task cost of $1.80, 72% below Opus 5.5 (max). Arena's own ordering still puts Opus 5.5 first overall.
If your workload is a coding agent that already runs on Opus 5.5, nothing here says to switch. See the Opus 5.5 launch coverage for the baseline you would be comparing against.
Carbon vs GPT-6 Astra: cheaper per task, but only if Astra's price list is what you pay #
GPT-6 Astra scored 57 on the same index and sits $6.25 per task at standard pricing, the lowest of the three leaders. Carbon is $7.76 and Opus 5.5 is $10.04, so Carbon is about 1.2 times Astra's cost per task and about 23% below Opus.
List prices tell a different story. Astra lists at $10 per million input and $50 per million output, while Carbon's standard price is $4 and $20. Astra reaches its low per-task cost by spending far fewer tokens. If your workload is not shaped like the Index tasks, the per-token price is the number that bites.
Astra still holds specific science and terminal wins on Artificial Analysis's earlier v5 table (Terminal-Bench-Science and the CritPt v2 hard subset), though Carbon now leads Terminal-Bench-Science. Our Astra launch post and the Sol vs Astra cost note cover where Astra and its cheaper sibling fit.
Carbon vs Grok 4.7 and the rest of the frontier #
Grok 4.7 is the cheap live option: $2 per million input, $6 per million output, generally available. It also beats Carbon on one independent board. On AA-Briefcase, a knowledge-work Elo, Grok 4.7 (xhigh) scores 1,638 against Carbon's 1,609. Muse Wonder (max) scores 1,659 and Kimi K3.1 Pro 1,678.
Carbon did gain 126 Elo over Argon on that board, from 1,483, so the direction is right. It is still not the best at office-style knowledge work. Compare the other labs side by side in our six-model Muse Wonder table and the Grok 4.7 launch post.
What does Carbon cost compared with each rival? #
| Model | Input / output per 1M | Cost per Index task | Output tokens per task |
|---|---|---|---|
| Gemini 4 Carbon (intro) | $2 / $10 | not shown | about 123k |
| Gemini 4 Carbon (standard) | $4 / $20 | $7.76 | about 123k |
| Claude Opus 5.5 | $4 / $20 | $10.04 (max) | about 173k |
| GPT-6 Astra | $10 / $50 | $6.25 (max) | not stated |
| Grok 4.7 | $2 / $6 | not stated | not stated |
Cost per task meter used to compare Gemini 4 Carbon vs Claude Opus 5.5 token spend
Two caveats keep the table honest. First, Google's introductory rate ($2 and $10) has no announced end date, and Artificial Analysis scores use the regular $4 and $20 price. Budget at the standard price. Second, Vals AI reports Carbon using 2.3 times Argon's reasoning tokens and 64% more output tokens, and some Legal Research tasks took over 50 minutes. Lower tokens per Index task does not guarantee a short or cheap job on your own work.
Does Carbon hallucinate less than Opus 5.5 and Astra? #
This is where Carbon looks best and needs the most care. On AA-Omniscience v1.1, Carbon scores 54 against Opus 5.5's 48, and its hallucination rate is 11%, second lowest of any model Artificial Analysis lists. Reports have Opus 5.5 (max) and Astra (max) at 50% on the corresponding line. Note that one outlet read the 50% as accuracy rather than hallucination rate, so check Artificial Analysis's own page before quoting it.
Accuracy versus hallucination tradeoff balance used to compare Gemini 4 Carbon hallucination rate with Argon
Against Argon, the trend is the other way. Accuracy rose from 48% to 58%, but hallucination rose from 4% to 11%. Artificial Analysis found Carbon guesses more often when unsure. A model that answers more questions will also be wrong more often when it does not know. For high-stakes retrieval, keep a verification step regardless of which model you pick.
Which of these can you actually call today? #
| Model | Callable? | Notes |
|---|---|---|
| Gemini 4 Carbon | No | Selected cyber defenders first; more partners through Fairwind; Google Cloud account team application; background checks; use limited to internal security, incident response or pen testing |
| Claude Opus 5.5 | **Yes** | `claude-opus-5-5` ; apps, API, major clouds |
| GPT-6 Astra | **Yes** | `gpt-6-astra` |
| Grok 4.7 | Yes | SpaceXAI API, Cursor, routers |
Google says developers, businesses and consumers get Carbon after more feedback and guardrail work, with no date. The model is also going through the U.S. government's voluntary pre-release testing. Until then, a comparison of Carbon against live models is bake-off planning, not a procurement decision.
How much of this is Google's own claim? #
A lot of the loudest numbers are. Google's cyber results run without safety guardrails and have not been independently replicated: SEC-bench Pro at 88.2%, CWE-Bench at 71.0% and ExploitGym at 54.8%. The ExploitGym headline uses a ten-hour cap, versus six hours for the Opus 5.5 comparison point. A Chrome-team comparison names no competitor model, and there is no separate Carbon system card yet. Our Argon safety report note is the closest published risk evidence.
Source citation trail linking benchmark claims back to their original evaluators
That is not a reason to dismiss Carbon. The independent Artificial Analysis, Vals and Arena results do not depend on Google's cyber table. It is a reason to separate the two layers when you read a chart. Our guide to reading AI benchmarks walks through how to do that.
Which model should you pick this week? #
- Production coding agents: stay on Claude Opus 5.5. It leads the coding-agent boards and is live.
- Science, long-context and general reasoning pilots: add Carbon as a test column once your account is approved. It leads four of nine Index evaluations.
- Cost-sensitive science or terminal agents: keep GPT-6 Astra or Sol in the mix. Astra has the lowest cost per Index task of the leaders.
- High-volume, cheap generation: Grok 4.7 at $2 and $6 is the live bargain.
- Security teams: Carbon is the model built for you. If you have a Google Cloud account team, apply to Fairwind, and run your own red-team prompts before trusting vendor numbers. If you deploy agents yourself, a guardrail layer such asAgentBeam from the explainx.ai team can stop an agent before it takes a dangerous action.
What to watch next #
Four things will settle the picture. First, whether Google publishes a Carbon model card and context window; Artificial Analysis and Vals list 1M tokens, but Google's post does not. Second, a public release date and a final end date for the intro price. Third, repeat runs on Terminal-Bench 4.0 and FrontierCode in more harnesses. Fourth, the next Anthropic and OpenAI releases, which The Information has reported are close; see our note on what was reported about a new Anthropic model, Carbon and OpenAI Terra.
For a deeper look at Carbon against its own predecessor, read Carbon vs Argon: what changed in two weeks. For research workloads specifically, see Carbon for deep research and the Perplexity WANDR claim.
Related reading #
Scores, prices and access are accurate as of October 31, 2026 and come from Google's announcement, Artificial Analysis, Vals AI and Arena as reported that week. They may change as Carbon access widens.