Gemini 4 Carbon vs Claude Opus 5.5: Complete Comparison and Report on Pricing, Restricted Access, Cybersecurity Benchmarks, and Cost per Task

Gemini 4 Carbon is the frontier model Google DeepMind announced on October 28, 2026, two weeks after Gemini 4 Argon reached general availability, and Google describes it as its strongest model for long-horizon engineering and cybersecurity defense. Claude Opus 5.5 is the Anthropic flagship released on September 22, 2026 as the first model of the Claude 5.5 family, sold through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry and used inside Claude's paid plans and Claude Code. The two now sit at the top of most independent leaderboards within a point or two of each other, which makes them the natural pair to set side by side.

On the numbers that both companies and the independent evaluators publish, Carbon leads more often than it trails. Google's own table puts Carbon ahead of Opus 5.5 on 14 of the 16 benchmarks where both have a score, including Terminal-Bench 4.0 as run by Artificial Analysis (65.2% vs 59.6%), ExploitGym (54.8% vs 34.5%) and long-context GraphWalks (96.2% vs 66.8%), while Opus 5.5 keeps the lead on the OSWorld-2.1 computer-use subset (87.2% vs 81.5%) and on FrontierSWE 2 (62.3% vs 61.2%). Artificial Analysis scores Carbon at 60 and Opus 5.5 at 59 on its Intelligence Index v5.0.1, measures Carbon at $7.76 per index task against $10.04 for Opus 5.5 at maximum effort, and still ranks Claude Code with Opus 5.5 three points above Carbon on its Coding Agent Index (62 vs 59).

The deciding fact for most readers is access. Carbon is in pre-release testing with the U.S. government and a subset of vetted cyber defenders, Google has given no date for developer, business or consumer availability, and its API rates ($2/$10 per 1M tokens during an introductory period, then $4/$20) apply only to admitted partners. Opus 5.5 is available today at $4 per 1M input tokens and $20 per 1M output tokens, with batch at half price and fast mode at $8/$40. For anyone who has to ship work this month, Opus 5.5 is the only one of the two that can be bought; Carbon is the stronger choice for vetted security teams already inside Google's program, and the one to plan around for long-context and cyber workloads once it opens. Prices in this comparison were verified on the official Google and Anthropic pages on November 5, 2026.

··········

GEMINI 4 CARBON AND CLAUDE OPUS 5.5: RELEASE TIMELINE, CONTEXT WINDOW AND REASONING CONTROLS.

Carbon builds on Gemini 4 Argon, which Google released globally on October 13, 2026, and the announcement by Koray Kavukcuoglu, Google DeepMind's SVP and Google's Chief AI Architect, frames it as a model built for sustained agentic work that Google has run internally for days at a time on single problems. Opus 5.5 arrived five weeks earlier with Anthropic's claim that it matches Claude Fable 5.1 on most work while costing about 40% less to run than Opus 5 at default settings. We covered both launches in detail, in our report on the Gemini 4 Carbon launch for cyber defenders and in our article on Anthropic's Claude Opus 5.5 release.

The technical documentation is uneven between the two. Anthropic publishes a model page with the context window, the output limit, the knowledge cutoff and a retirement date, while Google has published no developer documentation for Carbon: the Gemini API models page, the release notes (last updated October 30) and the pricing page do not list it, and the Gemini Enterprise Agent Platform guide to Gemini 4 covers only Argon. The 1M-token context window for Carbon in the table below comes from Artificial Analysis and Vals AI, both of which list it, and has not been stated on a Google page we could fetch.

........

Specification	Gemini 4 Carbon	Claude Opus 5.5
Announced	October 28, 2026	September 22, 2026
Developer	Google DeepMind	Anthropic
API model ID	Not published	claude-opus-5-5 (anthropic.claude-opus-5-5 on Bedrock)
Context window	1M tokens (listed by Artificial Analysis and Vals AI)	1M tokens (official)
Maximum output	Not published	128K tokens; 300K on the Batch API (beta)
Input modalities	Text, image, speech, video (Artificial Analysis)	Text and images
Reasoning controls	Thinking levels; tested at low, medium and high by Artificial Analysis	Adaptive thinking always on; effort low, medium (default), high, xhigh, max
Knowledge cutoff	Not published	June 2026
Weights	Proprietary	Proprietary

........

The reasoning controls deserve attention because every benchmark in this comparison depends on them. On Opus 5.5 thinking cannot be switched off, a request without an effort setting runs at medium, and Anthropic warns that the model thinks more per turn than Opus 5 at the same level, most of all at xhigh and max. Requests that set thinking to disabled or pass a fixed token budget return a 400 error, and forced tool use is no longer accepted. Google has not documented Carbon's thinking settings; its evaluation notes say only that Gemini runs used the highest thinking setting available in the Gemini API.

··········

GEMINI 4 CARBON VS CLAUDE OPUS 5.5 API PRICING, CACHE AND BATCH RATES.

At standard rates the two models cost exactly the same per token. Google's footnote to the Carbon announcement gives an introductory price equal to Argon's, $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% below the input price, and a post-introductory price of $4 and $20. Anthropic charges $4 and $20 for Opus 5.5, with cache reads at $0.20, which is the same 95% discount. Artificial Analysis computes an identical blended price of $2.94 per 1M tokens for both on its 7:2:1 cache-hit, input and output mix.

........

Price per 1M tokens (USD)	Gemini 4 Carbon (introductory)	Gemini 4 Carbon (after introductory period)	Claude Opus 5.5
Input	$2.00	$4.00	$4.00
Output	$10.00	$20.00	$20.00
Cached input / cache read	95% off input ($0.10)	95% off input ($0.20)	$0.20
Cache write	Not published	Not published	$5.00 (5-minute), $8.00 (1-hour)
Batch input / output	Not published	Not published	$2.00 / $10.00
Fast mode input / output	Not offered	Not offered	$8.00 / $40.00
Long-context surcharge	Not published	Not published	None across the 1M window
US-only inference	Not published	Not published	1.1x on all token types

........

Google has not said when Carbon's introductory period ends. Argon's lower rates run through November 30, 2026 on the Gemini API pricing page and double on December 1, and Google says Carbon's introductory price is the same as Argon's, but Carbon does not appear on that page and Artificial Analysis records the launch discount as having no confirmed end date. Our guide to Gemini API pricing in 2026 tracks the Argon change and will be the place where Carbon's own price list appears once Google publishes one.

Anthropic's list has more moving parts. Prompt caching charges a premium on the write (1.25x input for a five-minute cache, 2x for a one-hour cache) and a 0.05x rate on the read, the Batch API halves input and output, and fast mode, which Anthropic says runs at up to 2.5 times normal speed, doubles both rates and cannot be combined with batch. Requests pinned to US-only inference cost 1.1 times the standard rate, and Amazon Bedrock regional endpoints carry a 10% premium over global ones. No surcharge applies above 200K input tokens, so a 900K-token prompt is billed at the same per-token rate as a short one.

··········

WHO CAN USE GEMINI 4 CARBON AND CLAUDE OPUS 5.5 TODAY.

Access separates the two models more sharply than any benchmark does. Google's October roundup says Carbon is "not yet broadly available" and is in pre-release testing with the U.S. government and a subset of trusted cyber defenders, and that Google is refining safety guardrails and expanding partner testing before a broader launch to developers, businesses and consumers. Admission runs through the Fairwind Program, which prioritizes governments and national cyber authorities, critical infrastructure operators in healthcare, telecom, energy and finance, and core technology platforms. Applicants go through vetting that includes background checks, access requires user-level authentication with phishing-resistant MFA, use is limited to internal security, incident response and penetration testing teams, and partners may not resell or share access. The Fairwind page itself still describes the program in terms of Gemini 4 Argon, and Google has not published Carbon-specific terms or costs for it.

........

Surface	Gemini 4 Carbon	Claude Opus 5.5
Developer API	Fairwind partners only; no public API	Claude API, generally available
Cloud platforms	Not announced	Amazon Bedrock, Google Cloud (Google managed and Anthropic managed), Microsoft Foundry, Claude Platform on AWS
Consumer app	Not available in the Gemini app	Claude apps on Pro, Max, Team and seat-based Enterprise plans
Coding tools	Not announced publicly	Claude Code, GitHub Copilot CLI and VS Code, per Anthropic
Free tier	No	Not stated for Opus 5.5
Broad launch date	None given	Released September 22, 2026

........

Opus 5.5 is a normal commercial product. Anthropic lists the same model ID across its own API, Google Cloud and Microsoft Foundry, uses a separate ID on Amazon Bedrock, and says it raised five-hour usage limits for Pro, Max, Team and seat-based Enterprise plans at launch. The Claude Pro plan costs $20 a month in the US, with a five-hour session limit and a weekly limit across all models. Anthropic gives Opus 5.5 a retirement date no sooner than September 22, 2027, which is the kind of commitment a company deploying it in production needs and one that Google has not yet been in a position to make for Carbon.

··········

HOW THE GEMINI 4 CARBON AND CLAUDE OPUS 5.5 BENCHMARK FIGURES WERE PRODUCED.

Most of the head-to-head numbers come from a single table on the Google DeepMind Gemini page, and its methodology document explains where each figure originates. Carbon scores are pass@1 at the highest Gemini API thinking setting. For non-Gemini models Google says it reports maximum thinking or reasoning settings, falls back to the best available reasoning results where those are missing, and uses provider self-reported figures unless a note says otherwise. Several rows mix sources inside the row: on SWE-sweep, Agents' Last Exam and OSWorld-2.1, Google ran Carbon itself with its own harness while the Opus 5.5 score comes from a public leaderboard or from Anthropic's system card.

A second difference concerns safeguards. Anthropic's benchmark footnotes say Opus 5.5 was tested with production safeguards on, so when a classifier intervened, cybersecurity tasks went to Opus 4.8 and biology tasks went to Opus 5, which the company says likely lowers the reported scores. Google says Carbon's external benchmarks were evaluated without guardrails. On cyber benchmarks this gap in setup is large enough to change how the scores should be read, and the section on cybersecurity below returns to it.

Some figures also disagree between sources for the same model. Google's table lists Opus 5.5 at 87.2% on the OSWorld-2.1 offline subset, while Anthropic's own launch table shows 81.8% on OSWorld 2.1 (partial). Google cites 42.5% on AutomationBench from Zapier's public leaderboard; Anthropic shows 40.0% from Zapier's early-access run. Chartography appears at 66.3% in Google's table, run without tools per the Surge leaderboard, and at 89.0% with tools in Anthropic's. None of these is an error on its face, since each source names a different subset or setting, but the tables in this report keep the source attached to every number for that reason.

··········

GEMINI 4 CARBON VS CLAUDE OPUS 5.5 ON GOOGLE'S SHARED BENCHMARK TABLE.

The table below reproduces every non-cyber row in which Google published a score for both models, with GPT-6 Astra's figure added as a reference point because it is the third model in the same table.

........

Benchmark (source of figures)	Gemini 4 Carbon	Claude Opus 5.5	GPT-6 Astra
AA Intelligence Index v5.0 (Artificial Analysis)	59	58	57
Vals Index v2.2 (Vals AI)	70.4%	70.2%	65.0%
SWE-sweep (Carbon self-run; others from leaderboard)	9.6%	3.6%	9.5%
FrontierSWE 2 (Proximal leaderboard)	61.2%	62.3%	65.5%
Terminal-Bench 4.0 (Artificial Analysis)	65.2%	59.6%	59.1%
Agents' Last Exam v1.1 (Carbon self-run; others from leaderboard)	77.4%	71.6%	73.1%
OSWorld-2.1, offline subset (Carbon self-run; Opus from system card)	81.5%	87.2%	80.9%
AutomationBench (Zapier leaderboard)	53.9%	42.5%	41.4%
Vending-Bench 2, final balance (Andon Labs)	$16,484	$9,235	$15,515
Chartography, no tools (Surge)	72.5%	66.3%	71.0%
Blueprint-Bench 2 (Andon Labs)	58.6%	51.2%	49.7%
GraphWalks 256K to 1M, BFS F1 (Google self-run for all)	96.2%	66.8%	71.8%
RiemannBench (Surge)	80.0%	69.6%	72.0%
Terminal-Bench-Science 0.1 (Artificial Analysis)	64.8%	59.0%	63.3%

........

The widest gaps sit in areas Google chose to emphasize. GraphWalks, which tests graph traversal over contexts between 256K and 1M tokens, shows the largest margin of the whole table at almost 30 points, and Google ran every model on it itself. AutomationBench, Zapier's business-workflow benchmark, puts Carbon 11.4 points ahead, and Andon Labs' Vending-Bench 2 simulation ends with Carbon holding a bank balance 78% higher than Opus 5.5's. The rows where a third party ran both models under one harness (Artificial Analysis for Terminal-Bench 4.0 and Terminal-Bench-Science, Proximal for FrontierSWE 2, Surge for RiemannBench) are the most comparable, and on those Carbon wins three of four by between 5.6 and 10.4 points.

Opus 5.5 keeps two rows. Its OSWorld-2.1 figure is 5.7 points higher than Carbon's, although it comes from Anthropic's system card and Carbon's from Google's own CUA harness, so the gap measures two different setups as much as two models. On FrontierSWE 2 the 1.1-point lead is small enough to treat as level.

··········

CYBERSECURITY RESULTS FOR GEMINI 4 CARBON AND CLAUDE OPUS 5.5: EXPLOITGYM, SEC-BENCH PRO AND CWE-BENCH.

Google built Carbon's launch around defensive security, and its cyber rows show the largest relative gaps over Opus 5.5. The setups explain a good part of that distance.

........

Benchmark	Gemini 4 Carbon	Claude Opus 5.5	How the figures were produced
ExploitGym (exploitation)	54.8%	34.5%	Carbon self-run, Antigravity harness, single attempt, no internet, mitigations off; Opus from published reports
SEC-bench Pro (vulnerability finding)	88.2%	Not published	Carbon self-run, 183-task JavaScript engine split, average of 3 attempts
CWE-bench (patching)	71.0%	67.0%	Collinear AI leaderboard, programmatic pass@1, all models run by Collinear
Gray Swan indirect prompt injection, attack success	0.0%	No figure published	Google's figure; Anthropic says Opus 5.5 ties Fable 5.1 for the lowest rate on a Gray Swan benchmark

........

CWE-bench is the only cyber row where an outside organization ran both models, and there the margin is four points. ExploitGym shows a 20.3-point gap, but Carbon's run had no guardrails and up to six hours per task (with a 10-hour extended setting also reported), while Anthropic's system card describes production classifiers that block vulnerability discovery in compiled binaries and route flagged cyber requests to Opus 4.8. Any exploit task that tripped those classifiers would have been scored on a weaker fallback model. Anthropic's system card also states that Opus 5.5 meets or exceeds Claude Mythos 5.1 and Opus 5 on every cyber evaluation it ran, and places the model in the lower of its two cyber risk tiers.

For a security team the practical difference lies in what each company will let the model do. Carbon is offered to vetted defenders precisely so it can run exploitation and reverse-engineering tasks, with dual-use work such as threat simulation and malware analysis allowed under the Fairwind terms. Opus 5.5 permits source-code vulnerability discovery by default, and Anthropic says a less restricted version will come through its Cyber Verification Program, which is expanding to three access tiers.

··········

BENCHMARKS ONLY GOOGLE OR ANTHROPIC PUBLISHED.

Each company also reported results the other did not cover. These figures cannot be compared across the pair, but they show where each vendor expects its model to be judged.

........

Benchmark	Published by	Figure	Settings stated
Terminal-Bench 4.0 (Anthropic's run)	Anthropic	Opus 5.5: 66.4%	xhigh effort, Claude Code harness, ±2.6 points
FrontierCode v1.1 Main	Anthropic	Opus 5.5: 54.4% (54.6% at medium effort)	Max effort unless noted
CursorBench 4.0	Anthropic	Opus 5.5: 57.8% (52.5% at medium)	Max effort unless noted
GDPval-AA v2.1	Anthropic	Opus 5.5: 1846 Elo	Max effort, 44 occupations
Humanity's Last Exam, with tools	Anthropic	Opus 5.5: 67.7%	Max effort
WebDev Arena	Google	Carbon: 1768, ranked #4, up 89 Elo	Not stated
Secure code generation (internal)	Google	Carbon: more than 2x Argon's success rate	Internal benchmark
Terminal-Bench 4.0 cheating	Google	Carbon: 0.0% successful cheating, 0.4% attempts	Internal safety evaluation

........

Anthropic's own Terminal-Bench 4.0 run at xhigh effort gives Opus 5.5 66.4%, which is above Carbon's 65.2% on the Artificial Analysis run, while Artificial Analysis measured Opus 5.5 at 59.6%. The 6.8-point spread for the same model on the same benchmark comes from harness and effort choices, and it is the clearest reminder in this comparison that vendor-run and third-party-run scores should not be mixed in a single ranking.

··········

ARTIFICIAL ANALYSIS AND VALS AI RESULTS FOR GEMINI 4 CARBON AND CLAUDE OPUS 5.5.

Two independent organizations have tested both models with their own harnesses. Artificial Analysis evaluated Carbon at high reasoning and Opus 5.5 at max effort with Anthropic's default fallback enabled; its charts mark Carbon with a "no publicly available API" flag, and the company has not said how it obtained access. Vals AI posted Carbon results on October 28 at high reasoning effort and Opus 5.5 results on September 22, mostly at max effort.

........

Evaluation (organization, date checked)	Gemini 4 Carbon	Claude Opus 5.5
Intelligence Index v5.0.1 (Artificial Analysis, Nov 5)	60 (high)	59 (max)
Coding Agent Index (Artificial Analysis, Oct 28)	59 via Antigravity CLI	62 via Claude Code
AA-Briefcase v1.1, Elo (Artificial Analysis)	1610	1800
AutomationBench-AA (Artificial Analysis)	77.8%	69.5%
FrontierCode 2.0 (Artificial Analysis)	40.6%	46.4%
AA-Omniscience v1.1 (Artificial Analysis)	54	48
Hallucination rate (Artificial Analysis, Oct 28)	11%	50%
HLE-Diamond (Artificial Analysis)	69.1%	67.9%
Vals Index (Vals AI)	70.45%, rank 2 of 32	70.18%, rank 3 of 32
Vals Finance Index (Vals AI)	67.05%, rank 1 of 35	65.59%, rank 2 of 35
Vals Legal Index (Vals AI)	75.93%, rank 3 of 33	72.33%, rank 8 of 33
Terminal-Bench 4.0 (Vals AI)	63.13%	65.15%
Cost per Vals Index test (Vals AI)	$10.51	$28.12

........

The independent results split along a clear line. Carbon leads on knowledge-retrieval and reliability measures, with Artificial Analysis recording a hallucination rate of 11% against 50% for Opus 5.5 at max effort on AA-Omniscience, and it leads on the Vals finance and legal indexes. Opus 5.5 leads on the agentic coding and document-work evaluations that rely on a full harness: the Coding Agent Index, AA-Briefcase by 190 Elo points, FrontierCode 2.0 by 5.8 points, and Vals AI's own Terminal-Bench 4.0 run by two points. On the headline indexes from both organizations the two are within one point, which is about the size of the ±1-point margin Vals AI reports for its index.

Both Vals AI pages carry figures that differ between their header cards and their launch-day posts (Carbon's Vals Index appears as 70.45% and 70.36%, Opus 5.5's as 70.18% and 69.69%), so the table uses the current header cards.

··········

TOKEN CONSUMPTION AND COST PER TASK BY EFFORT LEVEL.

Because standard per-token prices are identical, the cost difference between the two comes entirely from how many tokens each model spends. Artificial Analysis measured 123K output tokens per Intelligence Index task for Carbon at high reasoning, of which 72K were reasoning tokens, against 173K for Opus 5.5 at max effort, of which 120K were reasoning. That is 29% fewer output tokens for Carbon and a cost per task 22.7% lower at standard pricing.

........

Configuration (Artificial Analysis)	Intelligence Index	Cost per task	Time per task
Gemini 4 Carbon, high	60	$7.76	29.6 min
Gemini 4 Carbon, medium	58	$6.59	25.9 min
Gemini 4 Carbon, low	54	$7.19	145.6 min
Claude Opus 5.5, max	59	$10.04	25.8 min
Claude Opus 5.5, xhigh	57	$6.05	16.1 min
Claude Opus 5.5, high	52	$3.46	9.2 min
Claude Opus 5.5, medium (default)	48	$2.45	6.7 min
Claude Opus 5.5, low	35	$0.98	2.9 min

........

The effort curves run in opposite shapes. Opus 5.5 scales cost down steeply: at xhigh it scores one point below Carbon at medium for $0.54 less per task and finishes in 16.1 minutes against 25.9, and at its default medium setting it costs $2.45 per task, about 2.7 times less than Carbon at medium, at the price of a 10-point lower index score. Carbon's curve is flat by comparison, and its low setting was both slower and more expensive per task than medium in Artificial Analysis's runs, which the organization's data shows but does not explain. These figures use Carbon's $4/$20 standard rate; during the introductory period every Carbon cost in the table would be half.

··········

WHAT THREE TYPICAL WORKLOADS WOULD COST ON GEMINI 4 CARBON AND CLAUDE OPUS 5.5.

To show what the price lists mean outside a benchmark, we calculated three workloads in Python from the official rates. The assumptions are deliberately simple. A long agentic coding session sends 4 million input tokens over its life, 90% of them served from cache, and produces 200K output tokens. A document review job runs 100 contracts of 800K tokens each with no caching and 10K tokens of output per contract. A support assistant handles 300,000 requests a month at 3,000 input tokens each, 2,000 of them a cached system prompt, and 400 output tokens. The comparison assumes both models spend the same tokens on the same job, which the Artificial Analysis data above suggests will not hold exactly, and it bills Opus 5.5's uncached input at the base rate; charging that input as five-minute cache writes adds $0.40 to the coding session and $300 to the monthly support bill.

........

Workload	Carbon, introductory rates	Carbon, standard rates	Claude Opus 5.5, standard	Claude Opus 5.5, Batch API
Agentic coding session (4M input, 90% cached, 200K output)	$3.16	$6.32	$6.32	Not suited to batch
Contract review, 100 x 800K tokens, 10K output each	$170.00	$340.00	$340.00	$170.00
Support assistant, 300,000 requests per month	$1,860.00	$3,720.00	$3,720.00	Not suited to batch

........

At standard rates the bills are identical by construction, so the real difference on each workload is the token count each model produces for it. Carbon's introductory rates halve every line, and for offline work that can wait up to a day, Opus 5.5's Batch API reaches the same $170 for the contract job at its regular price. Google has not published a batch rate for Carbon. Applied to Artificial Analysis's measured consumption, Carbon at high reasoning spends 29% fewer output tokens than Opus 5.5 at max, which would move the coding and contract lines in Carbon's favor; Opus 5.5 at xhigh or high spends fewer tokens than either and would move them the other way.

··········

GEMINI 4 CARBON AND CLAUDE OPUS 5.5 SPEED AND LATENCY.

Artificial Analysis measures speed on five-day rolling averages refreshed every six hours, so its figures shift from day to day, and its pages showed two snapshots on November 5. On the per-effort comparison, Opus 5.5 generated 107 to 112 output tokens per second across its effort levels against 66 to 69 for Carbon at medium and high, while Carbon's time to first answer token at high reasoning was 64.6 seconds against 351.2 seconds for Opus 5.5 at max. The single-pair comparison page showed 88 tokens per second for Carbon and 95 for Opus 5.5, with time to first token of 154 seconds and 641 seconds.

Both snapshots agree on the shape. Opus 5.5 at max effort thinks for a long time before it writes, then writes faster; Carbon at high starts answering sooner and writes more slowly. At medium effort Opus 5.5's time to first answer token falls to 13.7 seconds and its full task time to 6.7 minutes, which is the setting Anthropic ships as the default. Carbon's speed figures were measured through an endpoint that the public cannot call, so they may not reflect the serving capacity Google will offer at a broad launch. For Opus 5.5, fast mode is the documented way to buy more speed, at double the standard rates.

··········

SAFETY CLASSIFICATIONS, CYBER SAFEGUARDS AND DATA HANDLING FOR CARBON AND OPUS 5.5.

Google says Carbon uses the same layered safety stack as Argon, combining model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring, and that it is taking part in the U.S. government's voluntary pre-release testing process. Its internal figures report that high-severity misalignment alerts in live coding environments fell by almost 80% compared with Argon, unsupported claims were roughly halved, and overreach and under-service fell by 26% and 29%. Google has published a safety and security report for the Gemini 4 generation, written around Argon; we did not find a Carbon-specific model card.

Anthropic's system card classifies Opus 5.5 as having CB-1 capabilities without crossing the CB-2 threshold, rates its AI research and development risk below threshold, and assesses catastrophic misalignment risk as low. Its automated behavioral audit across nearly 2,000 scenarios gave Opus 5.5 the best score of any model the company had tested, and the model attempted to cross containment boundaries about 85% less often than Opus 5 or Mythos 5.1. Anthropic also says Opus 5.5 often suspects when it is being evaluated, which limits how well its evaluations predict behavior in deployment.

For data handling, Opus 5.5 offers zero data retention, a US-only inference option at 1.1 times the standard price, watermarking for EU AI Act compliance, and protection of its reasoning against distillation for API accounts created on or after August 31, 2026. Carbon's data terms are those of the Fairwind agreements, which Google has not published beyond the access controls listed above.

··········

WHAT GOOGLE AND ANTHROPIC HAVE NOT PUBLISHED ABOUT CARBON AND OPUS 5.5.

The gaps fall mostly on Carbon's side because it has not launched to the public. Google has not published a model ID, a maximum output length, a batch or priority rate, a cache-write rate, the end date of the introductory price, a broad availability date, or a model card. The 1M-token context window comes from independent evaluators. Google also has not said whether Carbon will reach the Gemini app plans or which plan tiers will include it.

On Anthropic's side, the company has not published Opus 5.5 scores for SEC-bench Pro or Gray Swan's indirect prompt-injection test as numbers, and its system card does not give the ExploitGym figure that Google cites from published reports. Anthropic does not state which Claude plans include Opus 5.5 by name on the Pro plan help page, although its launch page raises limits for Pro, Max, Team and seat-based Enterprise users. Neither company has published an independent run of the other's showcase benchmarks, and the only cyber benchmark in which a single outside organization ran both models is Collinear AI's CWE-bench.

··········

CHOOSING BETWEEN GEMINI 4 CARBON AND CLAUDE OPUS 5.5 BY WORKLOAD.

The verified facts point to a clear division for now. Carbon leads on most of Google's shared table, on independent long-context, knowledge and finance measures, and on token efficiency at the top setting, but only vetted security partners can use it. Opus 5.5 is level with Carbon on the headline independent indexes, ahead on harness-based agentic coding and document work, cheaper per task at every effort level below max, and available everywhere today. For a broader view of where Carbon sits inside Google's own lineup, our Gemini 4 Carbon vs Gemini 4 Argon comparison covers the step up from the generally available model.

........

Workload or buyer	Recommended model	Reason
Production coding agents shipping now	Claude Opus 5.5	Generally available; leads the Artificial Analysis Coding Agent Index (62 vs 59) and AA-Briefcase
Vetted security team in the Fairwind Program	Gemini 4 Carbon	Built and licensed for exploitation and vulnerability work; leads ExploitGym, CWE-bench and SEC-bench Pro
Analysis over 256K to 1M tokens of context	Gemini 4 Carbon, once available	96.2% vs 66.8% on GraphWalks in Google's own runs; same 1M window on both
High-volume work where cost per task decides	Claude Opus 5.5 at medium or high effort	$2.45 to $3.46 per Artificial Analysis task against $6.59 or more for Carbon
Overnight document batches	Claude Opus 5.5 on the Batch API	$2/$10 per 1M tokens; no published Carbon batch rate
Factual research where wrong answers are costly	Gemini 4 Carbon, once available	11% vs 50% hallucination rate on AA-Omniscience
Teams on Bedrock, Azure or Google Cloud contracts	Claude Opus 5.5	Listed on all three; Carbon has no announced cloud availability
Consumer chat subscription	Claude Opus 5.5	Available on paid Claude plans; Carbon is not in the Gemini app

........

A team choosing a model for work it must deliver this quarter has one real option, and the data supports Opus 5.5 run at xhigh, where it scores within three points of Carbon at high for $6.05 per task against $7.76, or at high effort, where it costs less than half as much per task as Carbon at any setting. A security organization already admitted to Fairwind should run Carbon on its exploitation and patching backlog, where Google's numbers and Collinear's independent CWE-bench run both put it ahead. Everyone else should keep Carbon on the evaluation list for long-context retrieval and low-hallucination research, and wait for Google to publish the model ID, the end of the $2/$10 introductory rate and a general availability date before budgeting for it.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org