Gemini 4 Carbon: same price per token as Argon, not the same bill

In short

• Since 10/11/2026, Carbon has been open through the API to paying developers and businesses, and is rolling out to Google AI Ultra subscribers. $2 and $10 per million tokens until 30/11/2026, then $4 and $20, like Argon.
• "More than 2 times" and "more than 6 times" cheaper per task: Google's figures. Only the MKA test is in our sources; Artificial Analysis ranks Carbon 2nd there, without writing out its cost per task.
• Measured by Artificial Analysis (high setting, a level of reasoning effort; at $4 and $20): $7.76 per weighted average task for Carbon, $9.95 for Argon, so $7,756 versus $9,947 for 1,000 tasks split as in its index. It's the other way round over the whole exam: $8,240 versus $7,620, with 205 million tokens written by Carbon and 113 million by Argon.

What has changed since 28/10/2026

On 28/10/2026, Carbon was reserved for cyber defense (our 28/10/2026 article on Gemini 4 Carbon). On 10/11/2026, the door opens. The $2 and $10 price had been announced, but not its end date: it will be 30/11/2026.

Paying developers and businesses get it through the API, in Google AI Studio and Gemini for business. For Google AI Ultra subscribers, it is rolling out in Google Antigravity, the Gemini app and AI Mode in Google Search.

Also new: /swarm, in preview in Antigravity: one prompt, a team of agents. According to Google, Carbon takes the heavy-thinking roles, alongside lighter models. No end date for the preview and no price of its own in the announcement.

The same day, Google DeepMind announced security proofs about Signal's code, written by Carbon agents (what was really proven about Signal's code).

The price per token: on sale until 30/11/2026

A token is a piece of a word. Until 30/11/2026: $2 per million input, $10 per million output, $0.10 per million re-read from the cache. From 01/12/2026: $4, $20 and $0.20. Argon: same prices, same dates.

Same price per token. But Google warns: Carbon uses more tokens than Argon, to explore, reason and verify for longer. The price of a task can't be deduced from the price of a token. It has to be measured.

Google's figures: "more than 2 times", "more than 6 times"

Google ranks Carbon first on three benchmarks it calls independent: CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at a cost per task more than 2 times lower than the next best outside model. Not in our files: I can't settle it.

The 6 times is about MKA, a test designed by Google DeepMind and run by Artificial Analysis: 200 tasks in which an agent reads plans, schematics or medical scans and delivers a report or a spreadsheet. According to Google, Carbon is 2nd there, more than 6 times cheaper on average per task than the leader.

At Artificial Analysis, the ranking holds: Claude Fable 5.2 (max setting) leads with 57%, Carbon (high) 54%, Claude Opus 5.5 (max) 53%; Argon (high) is at 51%. It's a score: the weighted share of the hardest criteria each deliverable meets, on average. The price isn't written down: the page spells out only the three lowest costs; Carbon's and Fable 5.2's are only drawn. The 6 times stays with Google.

The Artificial Analysis receipt: $7.76 versus $9.95

Artificial Analysis, AA for short, puts every model through the same nine-test exam. At the high setting, Carbon scores 60, Argon 50. AA counts at the 01/12/2026 price, $4 and $20, not the sale price.

A weighted average task costs $7.76 with Carbon, $9.95 with Argon. Carbon's figure was already in our 28/10/2026 article; the face-off with Argon wasn't. For 1,000 tasks split according to the index weights: $7,756 versus $9,947 (our calculation).

I have the receipt. Each amount is rounded separately, hence a $1 gap with the totals:

Output tokens cost a little more with Carbon, as Google writes: $128 more, for 122.8 million tokens versus 116.4 million over 1,000 weighted tasks, about 5.5% more (our calculation). The $2,191 gap is made on another line: cache writes, $2,428 more for Argon. Nothing to do with output tokens: it means storing part of the prompt to re-read it for less. One reservation: AA counts these writes at $4 per million, the input price (shown for Carbon, derived from its totals for Argon: our calculation). Google's price list, for its part, gives no cache-write price at all and offers automatic caching only.

On FrontierCode 2.0, programming in real open-source projects, 1,000 tasks cost $12,018 with Carbon and $18,737 with Argon. On HLE-Diamond, reasoning and knowledge questions, $1,628 versus $484. Carbon is cheaper on 4 tests out of 9: exactly the ones where it writes fewer tokens than Argon per task of the test. More expensive on the other 5. Over the whole exam, every task counted: $8,240 for Carbon, $7,620 for Argon.

My verdict: same price per token, not the same bill. Where Carbon writes less, it costs less. Where it writes more, it costs more.

What we don't know yet

• The cost per task at the sale price. None at AA, which calculates at $4 and $20.
• Carbon's cost on MKA, and the three tests behind the 2 times. Not written out, or missing from our sources.
• Cache writes. Google's price list gives no price; AA counts them at $4 per million.
• The cyber version. Footnote 1 limits the price to the model intended for public access; Carbon for cybersecurity goes through the Fairwind Program, with no price in our sources.

Key terms

• Token: a piece of a word, the unit a model reads and writes. APIs are billed per million tokens.
• Million tokens: the pricing unit. For a sense of scale, Carbon wrote 205 million tokens for the whole AA exam.
• API: the gateway through which software sends its requests to the model, and pays per use.
• Cache: text already sent, kept so it can be re-read for less. Writing it there and storing it are billed separately, according to AA.
• Cost per task: what one task of a test costs, calculated by AA from the tokens used and the prices; for the index, as a weighted average: each test counts by its weight in the score.
• Benchmark: a standard test, the same for every model, to compare them.
• Artificial Analysis (AA): a site that runs its own evaluations on models and publishes their scores, speeds and costs.
• AA index: a score that combines nine evaluations (version 5.0.1); higher is better.
• Reasoning setting: how much the model thinks before answering. AA tested Carbon at low, medium and high.
• Agent: a model that chains actions on its own: calling tools, writing code, testing it.
• /swarm: an Antigravity feature, in preview, that turns a single prompt into a coordinated team of agents you can follow in real time, according to Google.

Source sheet

• Google DeepMind: announcement "Gemini 4 Carbon is now available", signed by Koray Kavukcuoglu, published on 10/11/2026 at 16:00 UTC.
• Google: Gemini API price list, read on 10/11/2026 at 16:13 UTC (same content on ai.google.dev and aistudio.google.com).
• Artificial Analysis: Gemini 4 Carbon and Gemini 4 Argon pages, and the MKA evaluation page, read on 10/11/2026 at 16:13 UTC to.
• Our calculations: costs per task multiplied by 1,000, gaps by line item and by test, from the raw values on the AA pages.
• Links: Google DeepMind announcement on Gemini 4 Carbon (10/11/2026) · Gemini API pricing · Gemini 4 Carbon page at Artificial Analysis

Source passages

Pages S1 to S5 read on 10/11/2026 between 16:12:56 and 16:13:50 UTC (dated copies in the article's folder). "Raw value" = number read in the page data (HTML), tied to the model by its identifier.

Show the 6 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+12	"Nov 10, 2026" ; datePublished "2026-11-10T16:00:00+00:00" ; "Today, we’re making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers." ; "Two weeks ago , we released Carbon to a set of cybersecurity defenders via our Fairwind Program ." ; "Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model’s intelligence like complex, multi-step coding and debugging and open-ended scientific research and discovery — using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops." ; "It ranks #1 on CursorBench 4.0 , Perplexity’s WANDR , and Zapier’s AutomationBench at more than 2X lower cost per task compared to the next best external model." ; "we’re previewing /swarm in Google Antigravity, turning a single prompt into a coordinated team of agents you can track in real time. Developers can assign Carbon to roles that demand the deepest reasoning, like architectural design, coordination, and auditing, while pairing it with faster, lighter models across parallel isolated workspaces and adversarial verifiers that test every change." ; "Gemini 4 Carbon is available via the API in Google AI Studio and Gemini for business for all paying developers and enterprises, at the same introductory price 1 as Gemini 4 Argon ($2 per million input tokens and $10 per million output tokens). It is also rolling out to all Google AI Ultra subscribers in Google Antigravity , the Gemini app , and AI Mode in Search ." ; "we partnered with Artificial Analysis to launch the new Multimodal Knowledge Agent (MKA) benchmark that tests a model’s multimodal performance in agentic workflows. Carbon ranks #2 while being more than 6X lower in average cost per task compared to the top performing model." ; "AlphaProof Code agents then used Carbon to generate millions of lines of code in Lean" ; "For cybersecurity defenders: Apply to join the Fairwind Program to get access to Carbon for cyber on a rolling basis." ; footnote 1: "After the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply. Applies to the Carbon model intended for public access only."	https://deepmind.google/blog/gemini-4-carbon-is-now-available/
S2↑1↑2↑3↑4↑5↑6+2	Gemini 4 Carbon: "Gemini 4 Carbon is not available on the Free tier. Every token the model processes is counted and billed, including thinking tokens and internal tool system instructions and function declarations when using grounding." ; Standard: "Input price | Not available | $2.00 through November 30, 2026. $4.00 starting December 1, 2026." ; "Output price (including thinking tokens) | Not available | $10.00 through November 30, 2026. $20.00 starting December 1, 2026." ; "Context caching price | Not available | $0.10 through November 30, 2026. $0.20 starting December 1, 2026. Implicit caching only (explicit caching is not supported)." ; Gemini 4 Argon: "Our efficient Gemini 4 frontier reasoning model. It delivers Gemini 4 intelligence with faster responses and lower token spend per turn compared to Gemini 4 Carbon" ; same price lines ($2.00 / $4.00; $10.00 / $20.00; $0.10 / $0.20, same dates)	https://ai.google.dev/gemini-api/docs/pricing
S3↑1↑2↑3↑4↑5↑6+13	"Gemini 4 Carbon high • Proprietary model • Released October 2026" ; "60 Artificial Analysis Intelligence Index" ; "In $4.00 Out $20.00 Cache discount 95% $7.76 Cost per Intelligence Index task" ; "205M Output tokens from Intelligence Index" ; "It's also notably slow per task and very verbose." ; "Gemini 4 Carbon (High) costs $4.00 per 1M input tokens and $20.00 per 1M output tokens (based on Google's API)." ; "Artificial Analysis Intelligence Index v5.0.1 incorporates 9 evaluations" ; "Evaluations measured independently by Artificial Analysis" ; "Weighted average cost per Intelligence Index task. Each evaluation’s cost is calculated from input, cache hit, cache write, reasoning, and answer token prices, divided by task count, and weighted by its Intelligence Index weight." ; "The cost to run the evaluations in the Artificial Analysis Intelligence Index, calculated using the model's input, cache hit, cache write, reasoning, and answer token prices and the number of tokens used across evaluations (excluding repeats)." ; "cache write and cache storage are billed separately and vary by provider" ; "FrontierCode 2.0 New Agentic software engineering in open-source repositories" ; "HLE-Diamond New Reasoning & knowledge" ; raw values (identifier gemini-4-carbon, effort high): intelligenceIndex 60.025738829526 ; intelligenceIndexCostPerTask total 7.755545745286556, output 2.4553559174861244, cacheWrite 3.214016151173799, cacheRead 2.0391456408266344, nonCacheInput 0.0470280358 ; intelligenceIndexCost total 8239.688057352134 ; canonicalIntelligenceIndexTokenCount output 204759670 ; price1mInputTokens 4 ; price1mOutputTokens 20 ; cacheHitPrice 0.2 ; cacheWritePrice 4 ; costPerTask frontiercode-2-0 12.01822900910347, hle-diamond 1.6276047800000002	https://artificialanalysis.ai/models/gemini-4-carbon
S4↑1↑2↑3↑4↑5↑6+5	"Gemini 4 Argon (High)" ; "50 Artificial Analysis Intelligence Index" ; "In $4.00 Out $20.00 Cache discount 95% $9.95 Cost per Intelligence Index task" ; "113M Output tokens from Intelligence Index" ; "Gemini 4 Argon (High) costs $4.00 per 1M input tokens and $20.00 per 1M output tokens (based on Google's API)." ; raw values (identifier gemini-4-argon, effort high): intelligenceIndex 50.4121935905389 ; intelligenceIndexCostPerTask total 9.946608711550533, output 2.3273605813566736, cacheWrite 5.642398181480672, cacheRead 1.9523062287131876, nonCacheInput 0.02454372 ; intelligenceIndexCost total 7620.221037891417 ; canonicalIntelligenceIndexTokenCount output 113329224 ; cacheHitPrice 0.2 ; cacheWritePrice null ; costPerTask frontiercode-2-0 18.73697109089479, hle-diamond 0.483555492 ; Carbon on the same page: gemini-4-carbon-low (effort low), gemini-4-carbon-medium (medium), gemini-4-carbon (high)	https://artificialanalysis.ai/models/gemini-4-argon
S5↑1↑2↑3↑4↑5↑6+2	"Artificial Analysis' implementation of Google DeepMind's Multimodal Knowledge Agent (MKA) benchmark. It tests whether agents can read visually dense project files, such as architectural drawings, engineering schematics and clinical scans, and produce the deliverables a professional would." ; "We run its 200 tasks across 10 industry verticals. Each task gives the agent a workspace of input files and asks for deliverables such as a report, a spreadsheet, a slide deck or a marked-up drawing." ; "Claude Fable 5.2 (Max, Default Fallback) scores the highest on MKA Climb WSR with a score of 57%, followed by Gemini 4 Carbon (High) with a score of 54% and Claude Opus 5.5 (Max, Default Fallback) with a score of 53%" ; "Mean weighted share of Climb rubric criteria satisfied per task" ; "Climb criteria are the harder ones" ; "GPT-6 Luna (Max) has the lowest cost per task on MKA, at $0.18, followed by Claude Haiku 5.5 (Max) at $0.32 and MiMo-V2.6-Pro at $0.32" ; raw value mkaClimbWsr 0.5125916894412726 (gemini-4-argon)	https://artificialanalysis.ai/evaluations/mka
R1↑1↑2↑3↑4↑5↑6+6	Our calculations from the raw values in S3 and S4: cost per task × 1,000: $7,755.55 (Carbon), $9,946.61 (Argon); line items × 1,000, Carbon / Argon: output tokens 2,455.36 / 2,327.36; cache writes 3,214.02 / 5,642.40; cache re-reading 2,039.15 / 1,952.31; other input tokens 47.03 / 24.54; gaps Argon − Carbon: 2,428.38 − 128.00 − 86.84 − 22.48 = $2,191.06; per test × 1,000, Carbon / Argon: FrontierCode 2.0 12,018.23 / 18,736.97; HLE-Diamond 1,627.60 / 483.56; Carbon cheaper on AA-Briefcase, Terminal-Bench 4.0, FrontierCode 2.0, Terminal-Bench-Science, more expensive on AutomationBench-AA, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt; output tokens over the exam: 204,759,670 / 113,329,224 = 1.81; index weights (weighted cost ÷ cost per task): 0.15 AA-Briefcase, FrontierCode 2.0, AA-Omniscience, Terminal-Bench-Science; 0.10 Terminal-Bench 4.0, HLE-Diamond, CritPt; 0.05 AutomationBench-AA, GDP.pdf; output tokens per weighted average task × 1,000: 122,767,796 / 116,368,029 (+5.5%); output tokens per task, Carbon / Argon: AA-Briefcase 137,925 / 142,374, Terminal-Bench 4.0 160,695 / 165,330, FrontierCode 2.0 127,768 / 156,345, Terminal-Bench-Science 276,017 / 290,818 (Carbon writes less, costs less); AutomationBench-AA 52,006 / 32,595, AA-Omniscience 11,900 / 883, GDP.pdf 23,384 / 8,432, HLE-Diamond 67,974 / 20,698, CritPt 130,898 / 71,508 (Carbon writes more, costs more); cache-write price derived from AA's totals: tokens written to cache = input tokens − cache re-reads (cacheRead ÷ $0.20) − uncached (nonCacheInput ÷ $4); Argon 7,553.86M − 6,542.58M − 26.56M = 984.72M; $3,938.87 ÷ 984.72M = $4.00 per million; Carbon 7,622.68M − 6,933.21M − 82.75M = 606.71M; $2,426.85 ÷ 606.71M = $4.00 (= displayed price); output: $4,095.19 ÷ 204.76M = $20.00 (full price).	—