Gemini 4 Carbon vs Gemini 4 Argon: Complete Comparison and Report on Pricing, Access, Cybersecurity Benchmarks, and Long-Horizon Agents
Gemini 4 Carbon and Gemini 4 Argon are the two frontier models of Google's Gemini 4 generation, released two weeks apart. Argon was announced on September 30, 2026 for a limited group of cyber defenders and became globally available to developers, businesses and Google AI subscribers on October 13. Carbon was announced on October 28, 2026 as the company's strongest model, built for cybersecurity work and for agents that run for days, and it is currently limited to a subset of trusted cyber defenders in the Fairwind Program and to pre-release testing with the U.S. government.
On the measurements that both models share, Carbon is clearly ahead. Artificial Analysis gives Carbon 59 and Argon 50 on its Intelligence Index v5.0, with Carbon leading by 8 points on Terminal-Bench 4.0 (65% vs 57%) and by 12 points on HLE-Diamond (69% vs 57%), while Google reports 71.0% vs 68.0% on CWE-bench v1 and a prompt-injection attack success rate of 0.0% vs 0.7% on the Gray Swan benchmark. The list prices are identical: $2 input and $10 output per million tokens at the introductory rate, $4 and $20 at the standard rate, with cached input 95% cheaper.
Argon keeps three concrete advantages. It is the only one of the two that a developer or subscriber can actually use today, from the $4.99 Google AI Plus plan up to the API, Antigravity and Gemini for business. It answers much sooner, with a measured time to first token of 40.47 seconds against 147.83 seconds for Carbon. And it hallucinates less on the Artificial Analysis Omniscience test (4% against 11%). For almost every reader, Argon is the model to deploy now; Carbon is the better model for vulnerability research and multi-day engineering agents, available only to organizations admitted to Google's program. Prices in this comparison were verified on Google's official pages on November 3, 2026.
··········
GEMINI 4 ARGON AND GEMINI 4 CARBON RELEASE TIMELINE AND MODEL IDENTITY.
Google staged both launches the same way, starting with cyber defenders and widening access later. Argon followed that path in two weeks: the first announcement on September 30 described a rollout "to a set of trusted cyber defenders" through the Fairwind Program, and the global release on October 13 opened it to the Gemini API paid tier, Google AI Studio, Gemini for business, Google Antigravity, the Gemini app and AI Mode in Search. Carbon is at the first stage of the same path. Its announcement says developers, businesses and consumers will get it "as soon as possible", with no date, while Google gathers feedback and refines guardrails.
The Argon model card describes a large-scale sparse mixture-of-experts transformer with native multimodal input (text, images, audio and video) and text output, a knowledge cutoff of March 2026 and a combined input-output context of up to 1 million tokens. Google has not published a model card for Carbon. The DeepMind model card index lists Argon (updated October 13) and no Carbon entry, and the Gemini API models page, last updated October 28, lists `gemini-4-argon` as a stable model with no Carbon model code.
........
Item | Gemini 4 Carbon | Gemini 4 Argon |
First announcement | October 28, 2026 | September 30, 2026 |
General availability | Not yet, no date given | October 13, 2026 |
Positioning by Google | Strongest model, cybersecurity and long-horizon agents | Frontier reasoning model for coding, finance, legal and science |
Model card | Not published | Published October 13, 2026 |
Architecture disclosed | No | Sparse mixture-of-experts transformer |
Knowledge cutoff | Not published | March 2026 |
API model code | Not listed on the Gemini API models page | gemini-4-argon (stable) |
Current access | Subset of cyber defenders (Fairwind), U.S. government pre-release testing | API, AI Studio, Antigravity, Gemini app, AI Mode, Gemini for business |
........
The coverage on Data Studios followed the same sequence, from the global Argon launch with $2/$10 introductory pricing to the Gemini 4 Carbon announcement for cyber defenders and long-horizon agents.
··········
GEMINI 4 CARBON AND ARGON API PRICING, CACHE AND BATCH RATES.
Google priced Carbon exactly like Argon. Both carry an introductory rate of $2 per million input tokens and $10 per million output tokens, and both move to $4 and $20 after the introductory period, with cached input billed at 95% off the input price. The difference lies in the dates and in what is documented. For Argon, the Gemini API pricing page sets the end of the introductory period on November 30, 2026 and lists the new prices from December 1. For Carbon, the announcement gives no end date for the introductory rate, and the Gemini API pricing page does not list Carbon at all, so batch, Flex and Priority prices for Carbon are not published.
........
Price item (USD per 1M tokens) | Gemini 4 Carbon | Gemini 4 Argon |
Input, introductory | $2.00 (no end date published) | $2.00 through November 30, 2026 |
Output, introductory | $10.00 | $10.00 through November 30, 2026 |
Input, standard | $4.00 | $4.00 from December 1, 2026 |
Output, standard | $20.00 | $20.00 from December 1, 2026 |
Cached input | 95% off input ($0.10 intro, $0.20 standard) | $0.10, then $0.20 (implicit caching only) |
Batch input / output | Not published | $1.00 / $5.00, then $2.00 / $10.00 |
Flex input / output | Not published | $1.00 / $5.00, then $2.00 / $10.00 |
Priority input / output | Not published | $3.60 / $18.00, then $7.20 / $36.00 |
Free tier | No public access | Not available |
........
Argon's pricing page adds two billing details that apply to every agentic workload. Output is billed including thinking tokens, and with grounding the internal tool system instructions and function declarations are billed as processed tokens. Grounding with Google Search costs $14 per 1,000 requests after 5,000 free requests per month shared across Gemini 3 and newer models. Our Gemini API pricing guide tracks the December 1 increase across the rest of the Gemini line.
··········
CONTEXT WINDOW, OUTPUT LIMITS AND THINKING CONTROLS IN GEMINI 4 CARBON AND ARGON.
Argon's limits are documented in detail. The Gemini API model page gives a combined context window of 1,048,576 tokens, a default synchronous output cap of 65,536 tokens and up to 1,048,576 output tokens with long decode, in which the model generates in hops of up to 64K tokens and the server joins them when the request runs in background mode. The developer guide sets three thinking levels, `low`, `medium` and `high`, with `medium` as the default, tells developers to remove temperature, top_p and top_k because they are ignored, and recommends HTTP timeouts of at least 15 minutes for hard prompts. Explicit context caching is not supported; implicit caching is on by default for prompts of 4,096 tokens or more.
For Carbon, Google's own pages state none of these values. The figures available come from independent evaluators that tested the model through Google's API: Artificial Analysis lists a 1M-token context window, and Vals AI lists a 1M context window and 1,048,576 maximum output tokens for its high-reasoning runs. Artificial Analysis also publishes a separate "Gemini 4 Carbon (medium)" entry, which indicates more than one reasoning setting, but Google has not documented Carbon's thinking levels.
........
Specification | Gemini 4 Carbon | Gemini 4 Argon |
Context window | 1M tokens (Artificial Analysis, Vals AI) | 1,048,576 tokens (Google) |
Default output cap | Not published | 65,536 tokens |
Maximum output | 1,048,576 tokens (Vals AI settings) | 1,048,576 tokens with long decode (Google) |
Thinking levels | Not documented by Google | low, medium (default), high |
Input modalities | Text, image, speech, video (Artificial Analysis) | Text, image, audio, video, PDF (Google) |
Output | Text | Text |
Context caching | 95% discount stated; mechanism not documented | Implicit only, from 4,096 tokens |
Computer use, function calling, search grounding | Not documented | Supported |
........
··········
WHERE GEMINI 4 CARBON AND ARGON ARE AVAILABLE: PLANS, API AND PLATFORMS.
Access is the deciding difference between the two models today. Argon is spread across Google's consumer, developer and enterprise products. In the Gemini app, Google's subscriptions page says Google AI Plus ($4.99 a month) "includes Gemini 4 Argon", Google AI Pro ($19.99) gets "access to Gemini 4 Argon", and Google AI Ultra (from $99.99, with a $199.99 tier) gets the "highest access". The Gemini Apps help center does not give prompt counts per plan; it states usage multiples (Plus 2x and Pro 4x the standard limits, Ultra 5x or 20x Pro), context windows of 128K tokens on Plus and 1 million on Pro and Ultra, and limits that refresh every 5 hours up to a weekly cap.
Carbon appears on none of these pages. Its availability is the Fairwind Program for trusted cyber defenders, expanding gradually to more partners, plus Google's participation in the U.S. government's voluntary pre-release testing. Google's October recap, published November 2, still describes Carbon as "in pre-release testing" with no date for the broader launch.
........
Channel | Gemini 4 Carbon | Gemini 4 Argon |
Gemini app, Google AI Plus ($4.99/month) | No | Included |
Gemini app, Google AI Pro ($19.99/month) | No | Access |
Gemini app, Google AI Ultra (from $99.99/month) | No | Highest access |
AI Mode in Search | No | Pro and Ultra subscribers |
Gemini API paid tier and AI Studio | Not listed | Yes, global |
Google Antigravity | Not announced | Enterprises and Ultra subscribers |
Gemini for business | No | Yes |
Fairwind Program (cyber defenders) | Yes, a subset | Yes, version without cyber guardrails |
U.S. government pre-release testing | Yes, in progress | Part of the staged rollout before October 13 |
........
Argon's Fairwind version differs from the public one. The Argon announcement says trusted defenders and Google's internal teams get a version without cyber guardrails, and that the $4/$20 standard pricing applies to the version intended for public access. Google has not said whether the Carbon used by Fairwind participants carries the same guardrails as the eventual public release.
··········
BENCHMARK METHODOLOGY AND SOURCE RELIABILITY FOR THE CARBON VS ARGON COMPARISON.
The two models are documented unevenly, so the source of each number decides how far it can be compared. Argon's figures come from a published model card with a full table against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Carbon's figures come from the announcement post and a model evaluation document whose score tables are published as images. That document states its method in text: results are pass@1 unless noted, with no majority voting or parallel test-time compute in single-attempt settings, at the highest thinking settings through the Gemini API, and with competitors' numbers mostly taken from their providers' own reports. It also says responses flagged by safety filters are returned as empty strings.
Two independent evaluators tested both models with the same harness, which makes their results the cleanest comparison available. Artificial Analysis ran both at high reasoning and published a direct comparison page. Vals AI ran both at high reasoning effort with default sampling. Both organizations used Google's API at the standard $4/$20 prices for their cost figures, and their pages show small internal inconsistencies between summary tables and update notes; the figures in this report are those shown in each page's main table on November 3, 2026.
··········
DIRECTLY COMPARABLE SCORES: GEMINI 4 CARBON VS ARGON.
The table below includes only measurements where both models were tested by the same organization under the same settings. Carbon leads on 13 of the 18 rows and ties on one, while Argon leads on four, with the largest Argon margin on Vals AI's HLAB task set.
........
Benchmark (source, setting) | Gemini 4 Carbon | Gemini 4 Argon | Gap |
Intelligence Index v5.0 (Artificial Analysis, high) | 59 | 50 | Carbon +9 |
Terminal-Bench 4.0 (Artificial Analysis) | 65% | 57% | Carbon +8 |
FrontierCode 2.0 (Artificial Analysis) | 41% | 26% | Carbon +15 |
HLE-Diamond (Artificial Analysis) | 69% | 57% | Carbon +12 |
CritPt v2, hard subset (Artificial Analysis) | 63% | 53% | Carbon +10 |
Terminal-Bench-Science 0.1 (Artificial Analysis) | 65% | 53% | Carbon +12 |
AA-Omniscience v1.1 (Artificial Analysis) | 54 | 46 | Carbon +8 |
GDP.pdf (Artificial Analysis) | 28% | 22% | Carbon +6 |
AA-Briefcase v1.1, Elo (Artificial Analysis) | 1610 | 1484 | Carbon +126 |
AutomationBench-AA (Artificial Analysis) | 78% | 78% | Tie |
Vals Index (Vals AI, high) | 70.45% | 69.24% | Carbon +1.21 |
Terminal-Bench 4.0 (Vals AI) | 63.13% | 57.58% | Carbon +5.55 |
Tax Agent Bench (Vals AI) | 51.78% | 44.56% | Carbon +7.22 |
Vals Finance Index (Vals AI) | 67.05% | 64.89% | Carbon +2.16 |
Vibe Code Bench v1.1 (Vals AI) | 89.06% | 91.91% | Argon +2.85 |
Finance Agent v2.1 (Vals AI) | 67.16% | 67.66% | Argon +0.50 |
Vals Legal Index (Vals AI) | 75.93% | 76.72% | Argon +0.79 |
HLAB-Vals-1.1 (Vals AI) | 17.50% | 27.08% | Argon +9.58 |
........
Several Vals AI gaps sit inside the published confidence intervals. Finance Agent v2.1 carries about ±2 points for each model, the Legal Index about ±1.8 and the Vals Index about ±1, so those three rows read as statistical ties. The Terminal-Bench 4.0 result is consistent across both evaluators, with Carbon ahead by 8 points at Artificial Analysis and 5.55 points at Vals AI. In the Artificial Analysis launch analysis dated October 28, Carbon took the top score in 4 of the 9 index evaluations, and the evaluator wrote that Carbon is "not leading in coding and knowledge work" against the whole field even though it improves on Argon in both.
··········
[[ADPLUS_300x250_1]]
GEMINI 4 CARBON AND ARGON BENCHMARKS PUBLISHED BY GOOGLE ONLY.
Google's own figures cover more ground, and a few of them are directly comparable because Google reported both models on the same test. CWE-bench v1, the patching benchmark from Collinear AI, is the clearest case: Argon's model card gives 68.0% (tied for first with Grok 4.7 and GPT-6 Astra) and Carbon's announcement gives 71.0%. Carbon's results on external cybersecurity benchmarks were run without guardrails, according to the announcement, and the evaluation document specifies the harnesses: SEC-bench Pro on a 183-task JavaScript engine split averaged over three attempts, and ExploitGym as a single attempt with 6-hour and 10-hour timeouts, both in the Antigravity harness.
........
Benchmark (Google-reported) | Gemini 4 Carbon | Gemini 4 Argon |
CWE-bench v1 (patching) | 71.0% | 68.0% |
Gray Swan indirect prompt injection, attack success (lower is better) | 0.0% | 0.7% |
SEC-bench Pro (vulnerability finding) | 88.2% | Not published |
ExploitGym (exploitation) | 54.8% | Not published |
WebDev Arena | 1768 Elo, ranked #4 (+89 Elo) | Not published |
DeepSWE v1.1 | Not published | 77.9% |
AutomationBench (Zapier) | Not published | 51.3% |
GraphWalks 256K to 1M (BFS F1) | Not published | 84.2% |
LVBench (long video) | Not published | 91.7% |
OSWorld-2.0, offline subset | Not published | 69.2% |
Agent's Last Exam | Not published | 39.5% |
........
Google also reports internal comparisons that only make sense against Argon. Carbon's announcement claims more than a 2x uplift in success rate on internal secure-code generation, unsupported claims cut by about half, overreach down 26% and under-service down 29%, all measured against Argon in Google's internal testing. On Terminal-Bench 4.0, Google says successful cheating fell to 0.0% and cheating attempts to 0.4%, against the 2% shortcut-attempt rate it reported for Argon on its coding benchmarks. These are vendor measurements without an external replication.
··········
INDEPENDENT EVALUATIONS OF GEMINI 4 CARBON AND ARGON BY ORGANIZATION.
The independent results add rankings against the rest of the market, which helps place the gap between the two Google models. At Artificial Analysis, Carbon is ranked first of 60 models on the Intelligence Index, against a median of 38 for its comparison group. On the evaluator's Coding Agent Index, Carbon with the Antigravity CLI scored 59, three points behind Claude Code with Claude Opus 5.5 at maximum effort (62). On AA-Briefcase, its agentic knowledge-work test, Carbon gained 126 Elo over Argon but still trailed Grok 4.7, Muse Wonder and Kimi K3.1 Pro, according to the October 28 analysis.
........
Organization and date | Measure | Gemini 4 Carbon | Gemini 4 Argon |
Artificial Analysis, Oct 28 to Nov 3, 2026 | Intelligence Index v5.0 rank | #1 of 60 (59) | 50 |
Artificial Analysis, Oct 28, 2026 | Coding Agent Index (Antigravity CLI) | 59 | Not reported |
Artificial Analysis, Oct 28, 2026 | AA-Omniscience hallucination rate | 11% | 4% |
Artificial Analysis, Oct 28, 2026 | AA-Omniscience accuracy | 58% | 48% |
Vals AI, Nov 3, 2026 | Vals Index rank | 2 of 32 | 4 of 32 |
Vals AI, Nov 3, 2026 | Legal Research Bench | 55.77%, rank 1 of 89 | 54.81%, rank 5 of 89 |
Vals AI, Nov 3, 2026 | EMB v1.1 (Excel modeling) | 76.47%, rank 2 of 37 | 74.83%, rank 6 of 37 |
Vals AI, Nov 3, 2026 | Cost per Vals Index test | $10.51 | $13.37 |
Wiz (reported by Google) | Web penetration-testing benchmarks | Leads | 70.9% on Wiz pentest benchmark |
........
The hallucination result cuts the other way. Artificial Analysis measured Carbon's hallucination rate on AA-Omniscience at 11%, up from Argon's 4%, while Carbon's accuracy rose 10 points to 58%; Argon kept the lowest hallucination rate of any model the evaluator tracks and Carbon the second lowest, with Claude Opus 5.5 and GPT-6 Astra at 50%. The Vals AI cost per test also favors Carbon, $10.51 against $13.37 at the same token prices, which implies fewer billed tokens on that suite.
··········
OUTPUT TOKENS, LATENCY AND COST PER TASK: GEMINI 4 CARBON VS ARGON.
Equal list prices do not mean equal bills, because Carbon generally writes more. Artificial Analysis measured 123K output tokens per Intelligence Index task for Carbon against 116K for Argon (about 6% more) and 72K against 68K reasoning tokens, for a cost per task of $7.76 against $7.68. Across the full index, Carbon generated 205 million output tokens against 113 million for Argon, and the run cost $8,240 against $6,103, about 35% more. Vals AI found the opposite on its own suite, with Carbon cheaper per test.
Speed is mixed. Carbon generates faster once it starts, at 94 output tokens per second against 83 for Argon (about 13% faster), but Artificial Analysis measured a time to first token of 147.83 seconds against 40.47 seconds, about 3.7 times longer, and an end-to-end response time of 153.14 against 46.49 seconds. On complete agentic tasks the gap closes: time per task was 1,271 seconds for Carbon and 1,363 for Argon, and Vals AI recorded latencies of 47 minutes 45 seconds and 44 minutes 58 seconds for its index runs.
........
Measure | Gemini 4 Carbon | Gemini 4 Argon |
Output tokens per index task (AA) | 123K | 116K |
Reasoning tokens per index task (AA) | 72K | 68K |
Cost per index task at $4/$20 (AA) | $7.76 | $7.68 |
Cost to run the full Intelligence Index (AA) | $8,240 | $6,103 |
Output speed (AA) | 94 tokens/s | 83 tokens/s |
Time to first token (AA) | 147.83 s | 40.47 s |
End-to-end response time (AA) | 153.14 s | 46.49 s |
Time per agentic task (AA) | 1,271 s | 1,363 s |
Vals Index latency (Vals AI) | 47 min 45 s | 44 min 58 s |
........
The workload costs below were computed in Python from the official prices, with the token volumes stated in each row. Because the per-token prices are identical, the same token volume costs the same on either model; the difference between them comes from how many tokens each model spends, and only Argon can use the batch discount, since Carbon has no published batch price. Applying Artificial Analysis's 6% higher output per task to the security agent run raises its standard-price cost from $1.66 to $1.73.
........
Workload (assumptions) | Introductory price, per run | Standard price, per run | Monthly volume and standard-price total |
Security code-review agent: 400K input, 75% cached, 60K output | $0.83 | $1.66 | 500 runs: $830.00 |
Long-document analysis: 300K input, no cache, 8K output | $0.68 | $1.36 | 2,000 requests: $2,720.00 |
Same document analysis on Argon Batch API | $0.34 | $0.68 | 2,000 requests: $1,360.00 |
Multi-day engineering agent: 6M input, 85% cached, 400K output | $6.31 | $12.62 | 20 tasks: $252.40 |
........
For Argon users, the December 1 change doubles every figure in the introductory column. Carbon's introductory column has no published end date, so a budget built on it carries that uncertainty.
··········
GEMINI 4 CARBON AND ARGON SAFETY CLASSIFICATIONS AND CYBER GUARDRAILS.
Argon's Frontier Safety Framework results are public. The model card says the Cybersecurity Uplift Level 1 critical capability level and the biological Uplift Level 1 level "cannot be ruled out", so safeguards were applied, while Harmful Manipulation and the two machine-learning R&D levels were not reached; on stealth and situational awareness the model passed 14 of 15 challenges and a residual risk assessment was performed. Google says Argon refuses harmful cyber and CBRN requests while allowing legitimate dual-use research, monitors internal activations for misuse, and runs a misalignment monitor over chain-of-thought and actions that can halt execution.
Carbon uses the same stack, according to its announcement (training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring), with added alignment training for complex or high-pressure tasks and a behavior that retrieves sources and surfaces discrepancies before acting on ambiguous instructions. Google reports that high-severity misalignment alerts in internal live coding dropped by almost 80%. Carbon has no published Frontier Safety Framework table, which is the main safety document missing for a direct comparison, and its limited release is itself the guardrail Google chose for a model with stronger exploitation results.
··········
GEMINI 4 CARBON VS ARGON: DATA GAPS IN THE PUBLISHED EVIDENCE.
The comparison rests on an uneven record. Google has published for Argon a model card, API documentation, a pricing page with batch, Flex and Priority rates, and plan-by-plan access; for Carbon it has published an announcement, an evaluation methodology document whose tables are images, and prices without an end date for the introductory period. No source states Carbon's knowledge cutoff, architecture, thinking levels, default output cap or rate limits, and no Gemini app plan lists it.
The independent coverage also has limits. Artificial Analysis notes that Carbon has no publicly available API for most users and tested it through Google's first-party API; its time-to-first-token figure for Carbon differs slightly between the model page (152.52 seconds) and the comparison page (147.83 seconds). Vals AI's pages show summary and update figures that differ by fractions of a point. No arena ranking for Carbon in text chat was available, and WebDev Arena's 1768 Elo comes from Google's announcement.
··········
WHICH TO CHOOSE: GEMINI 4 CARBON OR GEMINI 4 ARGON BY WORKLOAD.
For a team that needs a model in production this month, the choice is already made by access: Argon is the only one of the two on the public API and in Google's subscriptions. Carbon becomes the stronger option once it can be bought, for workloads where its measured leads are large and consistent across evaluators: terminal and engineering agents, scientific reasoning, vulnerability discovery and patching.
........
Workload or user | Recommended model | Reason from the published figures |
Developer building on the Gemini API today | Gemini 4 Argon | Only model of the two on the public API, with batch and Flex rates |
Gemini app subscriber (Plus, Pro, Ultra) | Gemini 4 Argon | Carbon is not offered in any plan |
Fairwind cyber defender, vulnerability research | Gemini 4 Carbon | 88.2% SEC-bench Pro, 54.8% ExploitGym, 71.0% vs 68.0% CWE-bench v1 |
Multi-day engineering or terminal agents (when available) | Gemini 4 Carbon | Terminal-Bench 4.0 65% vs 57% (AA), 63.13% vs 57.58% (Vals AI) |
Interactive chat and short answers | Gemini 4 Argon | 40.47 s vs 147.83 s time to first token |
Fact-sensitive answers without tools | Gemini 4 Argon | 4% vs 11% AA-Omniscience hallucination rate |
High-volume document processing | Gemini 4 Argon | Batch API at $1/$5 until November 30, then $2/$10 |
Front-end and vibe coding | Gemini 4 Argon | 91.91% vs 89.06% on Vals AI Vibe Code Bench v1.1 |
........
Teams already on Argon should plan for the December 1 doubling of its prices to $4 and $20 per million tokens, and test whether Carbon's 9-point lead on the Intelligence Index justifies a switch once Google opens access, since at those standard rates the two models cost the same per token and Artificial Analysis measured only a 1% difference in cost per task.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
