No. 1 on the leaderboard, market cap plunges! Behind Gemini 4 Carbon’s “rapid rise” lies a high-stakes gamble by Google

Gemini 4 CarbonTops the leaderboard, hallucination rate rises: What’s Google in such a rush to release models?

Author | Gao Yunyi

Editor | Cen Feng

Yesterday, Google unveiled the world model Genie 4, prompting people to exclaim, “This is absolutely crazy.” This morning, Google quickly followed up with Gemini 4 Carbon, immediately topping the leaderboards and “reclaiming the global No. 1 spot.” On top of that, with Gemini 4 Argon released two weeks ago, Google is refreshing its model product lineup at a nearly “one model per week” pace.

This frequency of iteration is unprecedented in Google’s history. In the past, the company was more accustomed to releasing major versions on a “two years to hone a sword” cycle, but now it has suddenly switched to the Internet-style approach of “making quick strides in small steps.”

Behind this sudden change in pace is a comprehensive shift in Google’s choices of technology pathways, bets on its R&D model, breakthroughs in compute infrastructure, and internal management adjustments. It is also a concentrated counterattack after falling behind the pace of the industry.

01

Tops the leaderboard vs. $400 billion in market value wiped out,

Google Falls into the “Illusion of Benchmark Scores”

After the release of Gemini 4 Carbon, the tech world’s reaction could only be described as surreal.

The most eye-catching label in Google’s public announcement was“Back to No. 1 in the World”.It took the top score in 15 of 17 benchmark tests and even overtook Claude Opus 5.5 (58 points) and GPT-6 Astra (57 points) in the latest Artificial Analysis (AA) Intelligence Index evaluation, reclaiming the world’s No. 1 spot.

Caption: Gemini 4 Carbon takes first place in its 17 publicly disclosed benchmark evaluations

But in the capital markets, following the release of the new model and Q3 earnings,Alphabet’s shares fell as much as roughly 10% after hours, wiping out $300 billion to $400 billion in market capitalization.If the decline continued into the next day’s opening, it would set the record for the largest single-day loss in market value in Google’s history.

Caption: After the model’s release, shares of Google’s parent company Alphabet instead plunged

Why has Wall Street reacted so differently to the “perfect scorecard” Google delivered this time? Or what painful price did Google pay to cope with this “exam”?

Google’s latest model update centers on“safety capabilities” and “long-running task capabilities”, particularly inidentifying security vulnerabilities, where it particularly excels. According to the technical data published by Google, Gemini 4 Carbon achieved leading results at every stage of vulnerability discovery, exploit validation, and patch remediation.

Beyond security, Gemini 4 Carbon isbetter at writing code, with a clear breakthrough especially on complex, long-running tasks. It can run continuously for several days, overcome one hurdle after another, and ultimately solve complex, multi-step engineering tasks, such as finding and fixing subtle bugs in large codebases and low-level systems.

The leaderboard data makes it clear that, in the SWE-sweep evaluation (agent-level code-repair capability), Gemini 4 Carbon has doubled Gemini 4 Argon’s 4.3% score to 9.6%; in Terminal-Bench (autonomous operation and tool-use capability), Gemini 4 Carbon improved by 8.1%, reaching 65.2%.

In GraphWalks (ultra-long-text processing capability), Gemini 4 Carbon scored 96.2%, while the leading competitor reached only 71.8%, a staggering gap.

Much like in capital markets, netizens seem unconvinced by the “world No. 1” label.

The public has long been inured to benchmark score-gaming and even cheating. Some netizens directly asked:Have the evaluation results reached their peak, or has Google truly returned to the top?

Some said they had already been burned by Gemini 4 Argon.

Back then, in the official announcement and in Google CEO Sundar Pichai’s marketing, Gemini 4 Argon was billed as “ushering in a new era of long-horizon, high-difficulty professional tasks,” claiming the industry’s lowest hallucination rate and the ability to completely reverse the Gemini family’s shortcomings in agentic capabilities and hands-on coding.

However, Bloomberg subsequently reported that, according to internal employees, Gemini 4 Argon looked good on the leaderboard but fell apart the moment it was put to work, struggling particularly with front-end design.

This time, Carbon has further cemented Google’s image as a “small-town exam-prep prodigy” that studies for the test in pursuit of high scores.

In terms of coding ability, Gemini 4 Argon’s score of 57% on Terminal-Bench 4.0 (a hands-on command-line test) left it far behind the industry benchmark, Claude Opus 5.5 (66.4%). On the broader aggregate intelligence index, it also failed to surpass Claude and merely managed to tie with an old OpenAI model released long ago.

Netizens said,If Gemini 4 Carbon is overhyped again, Google will be humiliated.

Many others also complained thataccess to this Gemini 4 Carbon release was also restricted, leaving everyone able only to stare helplessly at the leaderboard scores without being able to try it directly. By the time it becomes available, it will already be far too late.

All these voices pouring cold water reflect the public’s genuine attitude toward Google’s models.

In terms of its positioning, Gemini 4 Carbon is not a next-generation flagship pretrained from scratch, but a specialized enhanced version built on the foundation of Gemini 4 Argon. This is also the key reason it could be launched so quickly, within two weeks.

Google's perennial problem isstrong benchmark performance but a poor user experience. More of its strengths are concentrated in vertical areas where Google has invested heavily, such as long-context processing, safety, and scientific computing, rather than in the general-purpose coding and everyday agent experiences that users perceive most strongly.

Moreover, Gemini 4 Carbon still has clear shortcomings in core general-purpose areas.

In the FrontierSWE 2 software engineering test, Gemini 4 Carbon scored 61.2%, trailing GPT-6 Astra's 65.5% and Claude Opus 5.5's 62.3%; in the OSWorld-2.1 computer-use test, Claude Opus 5.5 scored 87.2%, while Gemini 4 Carbon scored 81.5%, placing only second; in professional knowledge-work evaluations, Gemini 4 Carbon also failed to establish a clear lead over its competitors and even fell behind in some scenarios.

On the other hand,improved capabilities have also brought a higher probability of “making things up”.In the AA-Omniscience knowledge benchmark subset from Artificial Analysis, Gemini 4 Carbon's hallucination rate rose from 4% for Gemini 4 Argon to 11%. This is also one of the reasons it has "strong benchmark scores but an average experience."

A model that frequently “makes things up” cannot give enterprise customers the confidence to use it, even if it scores full marks in hacking and cyber defense.

02

The inside story behind the shift from discontinuing 3.5 Pro

to a flurry of releases

Since Gemini slipped from the top of the rankings, Google has also faced considerable pressure due to delays to its flagship product, the loss of key talent, and rivals defining the coding-agent direction first. Sundar Pichai's remarks across several podcast interviews fully laid bare the company's current anxieties, trade-offs, and strategic bets.

In an interview with Alex Heath, Pichai spoke for the first time about a key technology-roadmap decision made in midyear:voluntarily abandoning the originally planned Gemini 3.5 Pro.

At the beginning of the 3.5 series' development, the team had originally been advancing both the Flash and Pro tracks in parallel.But at the midyear internal retrospective, the team reached a critical consensus: agentic and coding capabilities are the key factors deciding success in frontier competition today, and Google still has clear gaps in both areas that must be closed as quickly as possible.

The team ultimately concluded that using lightweight Flash models was the most efficient way to ramp up capabilities: their small size, low compute costs, and fast iteration cycles made it possible to run a large number of experiments and rapidly build up capabilities in a short period. Based on this conclusion, Google launched four successive generations of Flash—3.5, 3.6, 3.7, and 3.8—within three months, focusing squarely on agentic and coding capabilities and proving out the entire end-to-end rapid-iteration workflow.

By the time the Flash lineup had been fully refined, pretraining for Gemini 4 Argon had already wrapped up. Releasing 3.5 Pro separately at that point would have given it an extremely short market lifespan, wasting compute and R&D resources instead. The team ultimately decided to skip this version and integrate all the capabilities accumulated during Flash iteration into Argon through post-training.

This marked a shift in Google’s release cadence from “taking two years to forge a single sword” to “periodic incremental releases”: rather than remaining conservative and passive in pursuit of across-the-board capability maturity, Google split the work by task domain and released whichever area was mature first.

When Demis Hassabis announced Carbon on X, he specifically emphasized: “We’ll be releasing more models soon, including more efficient versions.” Carbon was the first clear result of this new mechanism, and the release cadence in the Gemini 4.0 era will only accelerate from here on. Hassabis also pointed out, however, that specialized optimized versions such as Carbon and the distilled version of Argon aimed at broad adoption are two parallel technical tracks that do not overlap.

Behind this iteration speed is one key technical path: Google’s all-in bet on recursive self-improvement (RSI).

In a podcast, Hassabis further disclosed that Google had previously reported that roughly 75% of newly added internal code was AI-generated, that the code rollback rate had fallen by one-third, and that AI now affected virtually every piece of code; researchers were already collaborating extensively with AI agents to build the next generation of models, and the R&D process had officially entered the exploratory stage of RSI.

Google’s RSI approach is highly pragmatic: rather than directly venturing into the high-risk deep waters of modifying model weights, it starts with peripheral areas that can be quantified and easily verified, gradually forming a closed loop.

For example, AlphaEvolve, launched in 2025 as an evolutionary coding agent, has been fully deployed across Google’s production systems, extending into data center scheduling, chip design, and AI training workflows, and has officially been made commercially available through Google Cloud. In September 2026, Google also announced a series of RSI-related studies in quick succession. Among them, Dream-RSI addressed the problem of inefficient exploration paths for agents by constructing historical “dreams,” while RRSI introduced regularization to solve the problem of how to enable AI to autonomously optimize agentic workflows without going off course. As Google co-founder Sergey Brin put it, researchers and models are working hand in hand to build the next generation of models.

Even AI is deeply involved in Google’s internal allocation of computing power..Pichai said that he has his own dedicated internal agent to help him process information.

Generally speaking, the full model-development pipeline includes pretraining clusters, mid-training clusters, post-training data clusters, and so on. He has to spend an hour each day keeping up with 24 hours’ worth of progress. Now, agents produce summaries and roundups, allowing him to quickly grasp the big picture and dig into the details when needed. Similarly, agents monitor computing-power usage, efficiency, and utilization across each business line, enabling him to make the right decisions.

Since Gemini 4 Argon went live, Google’s internal productivity has seen a step-change improvement. Pichai gave a specific example: they used it on a large codebase to resolve known issues.

For example, the Chromium project wanted to implement a feature that allows users to drag a tab out of a window and move it without refreshing the page. In the past, when dealing with a codebase of millions of lines, the team would typically have to work through every layer of Chromium to accomplish this. Now, the task can simply be handed to Argon, which will autonomously keep pushing forward until the feature is fully implemented.

Improved model capabilities feed back into R&D efficiency, and greater R&D efficiency in turn produces stronger models. Once this cycle gets going, the pace of iteration will continue to accelerate. Pichai predicts thatover the next 12 months, the acceleration brought by RSI will become even more pronounced. The year 2027 will be a golden age at the frontier of productivity, with fields such as mathematics, science, and engineering experiencing a wave of “breakthroughs that leapfrog existing limits.”

For a long time, Google has struggled with insufficient computing power. Within 72 hours of its launch, Gemini 4 Argon had processed nearly 25 trillion tokens, with usage three times that of the previous-generation Gemini 3.0 over the same period, setting a Google model launch-usage record.

At the same time, because of computing-power constraints, Gemini 4 Argon currently gives priority access to API developers and Ultra/Pro subscribers and is not yet available to all ordinary users. Although Google is trying to create a lower-cost alternative through “model distillation,” computing power remains the key bottleneck in this survival game.

Pichai said on the podcast that computing-power shortages are a challenge facing the entire industry. The decision-making cycle for computing-power infrastructure generally takes as long as two years, and the computing capacity that supports Google today is the result of his forward planning 24 years ago.

To lock in its position in this prolonged war of attrition, Google has aggressively raised its 2026 capital-expenditure guidance to between $195 billion and $205 billion and has already begun planning ahead for land, power, and infrastructure in 2028 and 2029. This massive spending has directly resulted in Alphabet recording its first stretch of consecutive quarters with negative free cash flow since going public. This is the severe short-term financial cost paid in exchange for long-term technological dominance.

In this no-retreat trade-off over computing power, Pichai has made an exceptionally decisive choice:Google is actively giving up short-term revenue from its cloud business and directing its extremely limited computing resources, without reservation, toward developing internal frontier models.On this issue, there is absolutely no room for compromise.

Ultimately, Google’s rapid stream of new releases is the result of both external competitive pressure and internal strategic realignment.

Externally, OpenAI and Anthropic are updating their models at a rapid pace, and tools like Claude Code have already embedded models into developers’ daily workflows. Google lacks an equivalent breakout product and can only sustain market expectations through continuous model releases and benchmark results.

Internally, it is a restructuring of the “R&D paradigm.” Google has overturned its previous delivery model while implementing RSI throughout the entire R&D workflow. With limited computing power, it is using faster R&D cycles to close the gap in product entry points.

Once this closed loop is working, Google’s former weaknesses—its “sprawling organization and complex chain”—will, with the help of RSI, instead become an advantage no one can match: AI can handle chip design, data center scheduling, browser optimization, and more all in one go. This kind of “full-suite” self-improvement is a hard moat that no startup can copy.

03

Next: Can Gemini Redacted Save the Day?

TAG590@@

Interestingly, shortly after Gemini 4 Carbon was released, Google’s next model had already leaked.

In response to netizens joking that “Fable 5.5 is about to be released and will replace Gemini 4 Carbon,” a DeepMind employee proudly stated, “Gemini [redacted] isn’t here yet, though.” Netizens quickly turned the remark into a meme, treating the redacted term as the code name for the next-generation model: Gemini Redacted.

It remains to be seen whether the rumored Redacted is Google’s ace up its sleeve in its “three-pronged strategy,” or merely another all-sizzle-no-substance PR stunt.

Reference links: https://www.youtube.com/watch?v=fJBSNEFdm8c

https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/?utm_source=x&utm_medium=social&utm_campaign=&utm_content=

Hop on board with Lei Feng Network(WeChat official account: Lei Feng Network)and explore the highlights of the world’s top AI conferences

Enjoy exclusive access to:

Expert presentation slides

Full conference reports

Analysis of popular papers

Interviews with rising stars in academia

Scan the QR code above

or click“Read the original article”to visit the dedicated section.

This article was originally created by Lei Feng Network. Unauthorized reproduction is prohibited. See theReposting Notice for details.

(empty)