DeepMind Releases Gemini 4 Carbon With 42% Lower Energy per Thousand Tokens

DeepMind Releases Gemini 4 Carbon With 42% Lower Energy per Thousand Tokens

On October 28, 2026, Google DeepMind released Gemini 4 Carbon, the first frontier model in the Gemini 4 family, claiming roughly 42% lower energy use per thousand tokens than Gemini 3 Pro while expanding the context window from one million to four million tokens; it is available the same day across the Gemini app, AI Studio and Vertex AI. It is the first time a leading lab has put intelligence-per-unit-of-compute, rather than a benchmark score, in the headline of a launch. Liu Gong judges that

At 17:00 on October 28, 2026, Google DeepMind published Gemini 4 Carbon on its official blog, the first member of the Gemini 4 family, with a name pointing directly at carbon intensity and energy efficiency. DeepMind CEO Demis Hassabis and Google CEO Sundar Pichai confirmed the launch simultaneously on social platforms. The post lists three headline figures: about 42% lower energy per thousand tokens versus Gemini 3 Pro; a four-million-token input window with a 128,000-token output ceiling; and, on Google's own comparison basis, roughly 11 percentage points higher on SWE-bench Verified and 18% higher accuracy on one-hour video understanding. The lineup ships at once: Carbon Pro targets complex reasoning and agentic work, sustaining tool-call chains of about 30 minutes; Carbon Flash targets low latency at $0.60 per million output tokens; Carbon Nano is an on-device mixture-of-experts

Why now: over the past 18 months, marginal capability gains at the frontier have slowed while power, not silicon, has become the binding constraint. Grid interconnection queues for North American data centers now stretch for years, single-cluster power density has climbed from tens of megawatts to hundreds, and electricity rather than GPUs gates expansion. Google already surfaced cost-per-token in the previous two Gemini launches, but as a footnote; promoting efficiency to the headline signals a change in internal weighting. Competitive pressure is real: OpenAI and Anthropic both shipped agentic products in 2026 pitched on task completion rather than parameter counts, and enterprise buyers

Liu Gong's take: within six months, at least two other frontier labs will adopt the efficiency narrative, and by the first half of 2027 most flagship model launches will ship an energy-per-million-tokens line alongside benchmark tables.

From a technical standpoint, the real variable here is not the four-million-token window but the decoupling of capability from compute. In the Gemini 3 era, extending context beyond one million tokens yielded diminishing returns, with retrieval accuracy visibly decaying toward the tail of long documents. Quadrupling the window while cutting energy per token by 42% implies the gains came from sparse routing and attention scheduling rather than brute force. The claim of 18% higher accuracy on one-hour video is more informative than any aggregate score, because long video is a memory-bandwidth problem first and a reasoning problem second; progress there suggests the inference stack was rewritten, not merely scaled. In industrial quality inspection, the trade-off is familiar: a production line generates terabytes of image data daily, and two points of model accuracy are usually worth less than a 30% cut in per-frame inference cost, because the former can be absorbed by human re-inspection while the latter decides whether the line goes live at all.

On the business side, pricing is the sharpest part of the release. Carbon Pro output at $10 per million tokens, roughly 35% below its predecessor, and Flash at $0.60, sits close to the marginal cost of self-hosted inference for capable open models. The enterprise arithmetic is simple: a support system handling 200,000 calls a day at about 1,200 output tokens each saves enough in a year at the new rate to fund a small additional team. Google can afford to cut because it also sells TPUs, cloud capacity and the model itself, recovering margin downstream; pure-play model vendors have no such hedge. That asymmetry means API prices should fall faster in 2027 than in prior cycles.

For the industry, efficiency is becoming an entry barrier rather than a marketing line. On published parameters, Carbon Pro's output price is roughly one-third to one-half of comparable frontier listings while its context window is four times larger, which means the same budget buys close to an order of magnitude more usable context. For teams building repository-scale code understanding or multi-stream video analysis, that gap changes product shape, not just the income statement. Liu Gong predicts that the first quarter of 2027 will bring a wave of applications pitched on whole-codebase comprehension and full-season surveillance video search, viable precisely because of this price-and-window combination rather than because models became smarter.

  • Energy: about 42% lower energy per thousand tokens than Gemini 3 Pro on Google's published basis, with methodology and sampling conditions disclosed in the model card.
  • Context: input window expanded from one million to four million tokens, four times the prior generation, with a 128,000-token output ceiling.
  • Pricing: Carbon Pro at $10 per million output tokens and $1.25 input, roughly 35% below the previous generation; Carbon Flash output at $0.
  • – Google's own comparisons: about 11 percentage points higher on SWE-bench Verified, 18% higher accuracy on one-hour video understanding, and roughly 6 points higher on GPQA Diamond versus Gemini 3 Pro.

For the industry, the first three months will be dominated by price-sheet rewrites. Companies whose products center on long documents, code comprehension and video analysis will migrate traffic to Carbon Flash first, and gross-margin repair will show up in earnings before any new feature does. Smaller model providers that differentiated on cheap-but-small windows will be forced to follow on price. Over twelve months, energy efficiency is likely to enter enterprise procurement scorecards alongside accuracy and latency, shifting selection from who is strongest to who delivers the most output per dollar on a given task.

For ordinary users, the near-term change is a Gemini app that can read

Among competitors, neither OpenAI nor Anthropic issued a formal response on launch day, but both developer accounts emphasized long-horizon task completion for their agent products within the same week, read by the community as sidestepping a direct efficiency comparison while defending the quality narrative. Supporters of Meta's open-weight路线 argued that closed-model price cuts depend on subsidized integration across cloud and silicon, and questioned whether the pricing holds if compute costs rebound.

Expert opinion split. Systems and computer-architecture researchers broadly welcomed the 42% figure for its methodology, arguing that disclosing measurement conditions matters more than the

Questions worth exploring:

  • How energy efficiency moves from marketing claim to cross-verifiable standard, and why third-party replication is still missing
  • The retrieval-accuracy decay curve across a four-million-token window, which decides whether long-document products actually hold up
  • Whether cloud-plus-silicon-plus-model subsidy can persist, and where the floor for this round of API price cuts sits

Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.

This item comes from DeepMind Blog (published 2026-10-28 17:00:58). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.


Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private

Leave a Reply

Your email address will not be published. Required fields are marked *

中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn