Gemini 4 Carbon vs DeepSeek V4 Pro: What 37 Index Points and 4.4x the Cost Per Task Actually Buy

Put G​emini 4 Carbon and DeepSeek V4 Pro on the same scale and the result is not close: 60 against 23 on Artificial Analysis Intelligence Index v5.0.1, a 37-point gap between the fourth-best model on the board and one sitting below the median. Put them on the cost scale and it is not close either, in the other direction — $7.76 per index task against $1.77. The interesting question is not which is better. It is which of those two numbers you are actually paying for in the work you have, because for a large share of production traffic the 37 points go unused while the 4.4x bill arrives every month.

Both models are callable today, which makes this a real comparison rather than a roadmap exercise. G​emini 4 Carbon opened to paid API developers on 10 November 2026 after a twelve-day gated rollout. DeepSeek V4 Pro has been generally available since 13 August 2026 as the 0813 checkpoint, with open weights under MIT.

Two models built on opposite assumptions

The specs explain most of the behaviour difference before any benchmark does.

• Price, standard rate — G​emini 4 Carbon: $2 in / $10 out per million until 30 November 2026, then $4 / $20. DeepSeek V4 Pro: $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak, straight from D​eepSeek's own rate card.

• Cached input — Carbon $0.10 per million now, $0.20 after 1 December, implicit caching only. D​eepSeek $0.022 off-peak and $0.044 at peak, a 97% discount on the fresh-token rate.

• Context and output — both carry a 1M-token window. Carbon caps synchronous output at 65,536 tokens and allows up to 1,048,576 with long decode; V4 Pro caps at 384,000.

• Input modality — Carbon takes text, image, audio, video and PDF. V4 Pro is text only, and D​eepSeek's own pricing page says vision is not supported on v4-pro.

• Weights — Carbon is closed. V4 Pro ships under MIT, with a 1.6-trillion-parameter mixture-of-experts and 49 billion active parameters, so you can self-host, fine-tune and redistribute subject to the licence.

• Effort control — Carbon exposes thinking at low, medium and high. V4 Pro exposes a reasoning dial at low, high and max.

The modality line is not a footnote. If your pipeline ingests screenshots, scanned PDFs or call recordings, V4 Pro cannot be the model in that part of the stack at any price, and the comparison collapses to a single candidate. If your traffic is code, logs and text documents, the gap is genuinely a choice.

The 37 points, decomposed

A composite gap that size is usually a few evaluations doing all the work, and that is what the independent boards show for this pair. Artificial Analysis measures Carbon at high effort and V4 Pro 0813 at max effort; both figures are that evaluator's, not the vendors'.

• Cost per Intelligence Index task — Carbon $7.76, V4 Pro $1.77. That ratio already folds in verbosity, so it is the fairest single economic number between them.

• Output tokens burned on the index — Carbon 205 million, V4 Pro 119 million, against a 72.8 million median. Both models are talkative; Carbon is roughly 1.7x more so, which is where a chunk of the price gap comes from rather than from the per-token rate alone.

• Speed and latency — Carbon 80 output tokens per second with 66.85 seconds to first answer token; V4 Pro 67 tokens per second with 67.71 seconds. On this axis they are near-identical, and both are slow enough that neither belongs in an interactive chat path without streaming and patience.

• Effort sensitivity — Carbon scores 60 at high, 58 at medium and 54 at low. V4 Pro's non-reasoning configuration scores 13 against 23 at max. The cheaper model loses far more when you turn reasoning down, so "run it cheap" is not a free switch on either side.

On the vendor side the numbers are loud and unaudited, and they do not overlap enough to compare directly. D​eepSeek's own card for the 0813 build reports Terminal-Bench 2.1 at 87.9, DeepSWE at 62.7, CyberGym at 83.3, SWE-bench Verified at 80.6 and a Codeforces rating of 3206 — all vendor-reported, none reproduced by a neutral party at the time of writing. G​oogle's card for Carbon reports SEC-bench Pro at 88.2%, ExploitGym at 54.8% and CWE-bench at 71.0%. Note what happens if you try to line those up: CyberGym and ExploitGym are different harnesses, Terminal-Bench 2.1 and Terminal-Bench 4.0 are different versions of a moving target, and only one of the two vendors published a cyber score on a benchmark the other also ran. Any chart setting 83.3 beside 54.8 is comparing a scale to a different scale.

The billing clock nobody budgets for

D​eepSeek's price is not one price, and this is the most under-reported difference between the two models. V4 Pro bills at half rate outside a defined peak window: peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else — nights, weekends, holidays — is off-peak.

Seven hours a weekday out of 24, five days a week, is about 21% of the calendar week. If your batch jobs are scheduled by a cron in UTC, that is easy to align. If your traffic is a live product with a US or European daytime curve, a meaningful share of it lands inside the peak window and your real blended rate sits between $0.66 and $1.32, not at the number on the marketing page.

Carbon's clock runs the other way. It is flat until 30 November and then doubles. Two models, two completely different sources of bill variance: one varies by hour of day, the other by calendar date. Neither is visible from a single per-token figure, and a comparison that quotes $4/$20 against $0.66/$1.98 is describing a worst case on one side and a best case on the other.

The honest version: at Carbon's introductory rate against V4 Pro's peak rate, Carbon is 1.5x on input and 2.5x on output. After 1 December against off-peak, it is 6x and 10x. The 4.4x cost-per-task figure sits between those and is the number worth planning with, because it accounts for how many tokens each model actually needs to finish.

Where the gap is worth paying for

Rather than a verdict, four splits that hold up against how the two models are actually shaped.

Cyber and vulnerability work. This is Carbon's home ground and the reason G​oogle gated it to defenders first. Carbon reads code, images and PDFs in one call, which is what a vulnerability triage pipeline needs, and G​oogle's own reported numbers put it at the top of the cyber boards it chose to publish. V4 Pro has respectable self-reported cyber scores and cannot see a screenshot. For a security team, Carbon at 6x the input price is not obviously the wrong call.

High-volume classification, extraction and routing. Send this to V4 Pro. A 37-point composite gap says almost nothing about whether a model can tag a support ticket or pull fields from an invoice, and at $0.66 per million input the cheap model wins on arithmetic long before it loses on quality. Most teams over-buy here.

Long-context document work. Both windows are 1M tokens, so raw capacity ties. The differentiator is modality and output ceiling: Carbon ingests the scanned PDF and writes up to 1M tokens; V4 Pro needs the text already extracted and stops at 384K. If your pipeline is already OCR'd, V4 Pro carries the same 1M window for a tenth of the output price.

Anything you might have to self-host. Only V4 Pro is an option. MIT weights mean no vendor can withdraw the model, reprice it or gate it behind a programme — which is not a hypothetical concern in a year when Carbon's access conditions changed twice in fourteen days and its price changes again on 1 December.

The routing answer, with the honest caveat

The mature version of this decision is not picking one. It is running V4 Pro on the bulk of traffic and escalating the hard cases to a frontier model, which is exactly the composition our routing DSL exists for: a cheap branch that handles the volume, an escalation condition that fires on low confidence or failed validation, and one OpenAI-compatible endpoint in front of both.

What is on that endpoint today is deepseek/deepseek-v4-pro at D​eepSeek's own $0.66/$1.98, alongside google/gemini-4-argon at $2/$10 — Argon being the G​emini 4 model that carries the same 1M window and a 50 on the same index, at half Carbon's post-December output price. G​emini 4 Carbon is not in our catalogue; it is reached through G​oogle's paid API, and a Carbon branch in a pipeline is a direct integration. Because provider list price is passed through with 0% markup across the 200-plus models on the key, a vendor rate change lands the day it happens, which for Carbon means 1 December.

Automatic failover is the part that changes the calculus on the cheap model. The reason teams over-buy frontier capacity is fear of a bad answer on the 5% of requests that matter. Failover and escalation let the cheap model take the 95% it handles fine, with the expensive model reachable on one endpoint when a request falls outside it — rather than paying frontier rates on every token to insure against a tail you could route instead.

What would change this recommendation

Two developments, both plausible within a quarter. If a neutral evaluator runs Carbon on a cyber benchmark that D​eepSeek also published, the two scorecards become comparable for the first time and the gap stops being an inference. And if D​eepSeek ships the V4.1-Pro that its own migration note pointed at, the cheap side of this comparison gets stronger without the expensive side moving — which would make the escalation split even more attractive than it is now.

Until then, the defensible position is unglamorous: V4 Pro for volume, Carbon only where multimodal input or cyber capability is load-bearing, and no long-term contract signed on either side of a price that doubles in twenty days.