AI · News

Gemini 4 Carbon vs Claude and GPT: price, speed and benchmark comparison

Google’s new model tops Artificial Analysis Index v5, but it is still in a restricted rollout. A matched comparison shows where Carbon leads, where Claude or GPT does better, and how introductory pricing differs from benchmark costs.

By CherTra News DeskCherTra News29 October 20265 min read

Rows of network and server equipment in a data centre; contextual infrastructure photography, not Google hardware or a Gemini product image.Photo by Kevin Ache / Unsplash

A new model, but not an open public launch

Google announced Gemini 4 Carbon on October 28, describing it as its strongest model and emphasizing cybersecurity and long-horizon tasks. The initial rollout is deliberately limited: Google says access starts with a subset of cyber defenders, then expands to trusted partners through its Fairwind program while safeguards are tested. As of October 29, the company had not announced general developer, business or consumer availability.

Artificial Analysis reported Carbon at the top of its v5 Intelligence Index, scoring 59 in the listed high-reasoning configuration. That is a composite across knowledge work, coding, science and general evaluations, not a universal measure of quality. Google’s launch benchmarks are vendor-reported; the independent index provides a wider comparison, but it too reflects a specific set of tasks and configurations.

What the model comparison does — and does not — say

The table puts the requested models on the same Artificial Analysis Index v5 snapshot where possible, then lists vendor API prices and AA speed, first-answer latency and task-cost figures. The Index scores are rounded to whole points in the comparison pages. Carbon leads Opus 5.5 by one point and GPT-6 Astra by two; Opus remains ahead on FrontierCode 2.0, Sonnet leads AA-Briefcase, and Astra leads the hard CritPt subset. There is no single winner across the suite.

GPT-6.1 Terra launched on October 26, after GPT-6.1 Sol, so it is included rather than silently omitted. Its v5 score is lower than the models above, but its $1/$5 standard API price and fast, image-capable profile make it relevant for cost-sensitive visual or computer-use tasks. It is not a direct replacement for Astra or Sol: teams should compare the exact reasoning setting and product surface that fits their work.

Gemini 4 Carbon and comparable frontier models

Metric	Gemini 4 Carbon	Claude Opus 5.5	Claude Sonnet 5.5	GPT-6 Astra	GPT-6.1 Sol	GPT-6.1 Terra
AA Intelligence Index v5 (higher)	59	58	54	57	55	51
API input / 1M tokens, standard	$4	$4	$2	$10	$2	$1
API output / 1M tokens, standard	$20	$20	$10	$50	$10	$5
Introductory input/output per 1M	$2 / $10 through 30 Nov	—	—	—	—	—
Cached input / 1M tokens	$0.20 ($0.10 intro)	$0.20	$0.10	$1.00	$0.10	$0.05
AA output generation speed	98 t/s	98 t/s	137 t/s	78 t/s	79 t/s	91 t/s
AA first answer token, reasoning included	103 s	648 s	410 s	232 s	249 s	—
AA cost per Index task, standard price	$7.76	$10.04	$10.02	$6.24	$1.27	$0.96
Context window	1M	1M	1M	1M	1M	1M
Maximum output documented	Not published	128k	128k	128k	128k	128k
Input modalities	Text, image, speech, video	Text, image	Text, image	Text, image	Text, image	Text, image
AA reasoning configuration	High	Max + fallback	Max + fallback	Max	Max	Max
API / rollout status on 29 Oct	Restricted initial cohort; not broadly available	Public API	Public API	Public API; ChatGPT rollout	Public API; Work/Codex	Public API; Work/Codex

Artificial Analysis (AA) figures are the 29 October 2026 snapshot; Index points and model configurations are shown as AA reports them. Prices are standard API rates unless a Carbon introductory rate is explicitly shown. AA cost per task is the weighted cost of its Index evaluation suite, not a quote for a typical prompt. Time to first answer token includes reasoning time; it is not interchangeable with output tokens per second. A dash means no comparable value was verified in the cited source.

Where the benchmark results diverge

Metric	Gemini 4 Carbon (High)	Claude Opus 5.5 (Max, default fallback)	GPT-6 Astra (Max)
Artificial Analysis Intelligence Index v5.0	59	58	57
FrontierCode 2.0 · coding agents	41%	46%	42%
Terminal-Bench 4.0 · terminal coding	65%	60%	59%
AutomationBench-AA · agentic workflows	78%	70%	68%
AA-Briefcase v1.1 · knowledge work (Elo)	1609	1800	1568
CritPt v2 · hard physics subset	63%	61%	71%
HLE-Diamond · reasoning and knowledge	69%	68%	64%
Terminal-Bench-Science 0.1	65%	59%	63%
SEC-bench Pro · Google-reported security	88.2%	—	—
ExploitGym · Google-reported security	54.8%	—	—
CWE-bench · Google-reported patching	71.0%	—	—

Cross-model values are Artificial Analysis Intelligence Index v5.0 results for the named configurations. Google’s three cyber scores are vendor-reported and do not share a verified common test configuration with the other models; blanks mean not reported here, not zero.

List price, promotion and cost per benchmark task

Google’s announcement lists introductory Carbon pricing of $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95%. The Google developer pricing material sets the introductory window through November 30, 2026; standard rates then become $4/$20, with cached input at $0.20 per million. The table shows both rates to avoid confusing a launch discount with the later price. Google says tokens used for thinking are billed as output.

Artificial Analysis uses the standard $4/$20 list rate in its Carbon charts, not Google’s launch half-price. Its estimated $7.76 per Intelligence Index task is therefore a benchmark-specific normalized cost, not an average customer invoice. Reasoning-token use, task mix, cache hits and prompt design all affect real spend. At listed rates, Sol and Terra have much lower AA task costs; Carbon’s higher composite result comes with more token use and a different workload profile.

Speed is not the same as time to the first answer

Artificial Analysis reports output generation after the first streamed chunk separately from the time to the first answer token. On the evaluated reasoning runs, Carbon generates at roughly 98 tokens per second, near Opus, faster than Astra and Sol, and slower than Sonnet. Its first-answer delay is around 103 seconds in the matched comparison snapshot; Opus and Sonnet take longer in that demanding evaluation, while Astra and Sol also have substantial reasoning delays.

These numbers should not be read as the wait a person will experience on a short everyday prompt. The AA first-answer measure includes the model’s reasoning phase, and its Index tasks are long, difficult workloads. Provider, effort setting, prompt size, output length, region and service load all matter. A model can stream quickly once it starts but still take longer to begin; the two measures answer different questions.

Where Carbon stands out — and where other models lead

Carbon leads the selected Index composite and several of its sub-evaluations, including AutomationBench-AA, HLE-Diamond and Terminal-Bench-Science. Google separately reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, framing them as cybersecurity defense results. Those are Google-reported figures with no matched cross-vendor rows in the cited comparison; they should not be treated as proof that Carbon is safer in every deployment or that a benchmark equals real-world defense.

The second table also shows meaningful counterexamples: Opus 5.5 leads FrontierCode 2.0, Sonnet 5.5 scores higher on AA-Briefcase, and GPT-6 Astra leads CritPt v2’s hard physics subset. Tool ecosystems, coding harnesses, context handling and price may matter more for a particular task than a two-point Index gap. The benchmark is a shortlist for testing, not an adoption decision.

Availability and a practical selection test

Carbon was not broadly available at launch. The public API status in the table therefore refers to a restricted initial cohort and trusted-partner expansion, not a general signup path. Google’s announcement says developer, business and consumer access is planned after further evaluation. The other models in the comparison have public APIs, although product access, plans and regional availability differ; API prices should not be compared directly with subscription quotas.

For a team evaluating models, run a small representative set of real tasks with the same inputs, tool definitions and acceptance criteria. Record completion quality, review time, first-answer latency, total wall-clock time, output tokens and cost. Include failure cases and permissions. Carbon is compelling when its multimodal input, long context or reasoning performance solves a defined need; another model may be better for a faster first response, a specific coding workflow, current access or a tighter cost envelope.

Sources & further reading

Have a factual correction or a source to suggest? Contact the editorial desk.