Locked-up AI15 minOn November 5, OpenAI launched GPT-6 Cyber, a model it rates Critical for cybersecurity, and almost nobody can use it. The bigger story isn't hacking. It's whether the most capable AI is starting to live behind an application form.Read now →

Google Built the Best AI Model in the World for 12 Days. Almost Nobody Could Use It
Gemini 4 Carbon topped the leaderboards, then lost the crown the day before most customers could touch it. Meanwhile Google's share of enterprise AI spend fell from 21% to 9%. Here's the three-part test for whether Google is actually back, or just back on the leaderboard.
For 12 days this autumn, Google had the most intelligent AI model on the planet, at least by one of the industry's most-watched scorecards.
For almost all of those 12 days, you couldn't use it.
And on the day Google finally opened the doors to paying developers and businesses, it had already lost the top spot.
That strange sequence sums up Google's 2026 in AI better than any earnings call. The company that invented the transformer, designs its own chips and reaches billions of users has spent the year trying to answer one awkward question: if Google can build the best model in the world, why are the companies that actually pay for AI moving the other way?
Here's the full picture, and a simple three-part test for whether Google is truly back.
A year of missed deadlines
To understand why Gemini 4 Carbon matters, you have to remember how rough the months before it were.
- February: Gemini 3.1 Pro sat at the top of the Artificial Analysis Intelligence Index. Google looked comfortable.
- May, at Google I/O: Google promised Gemini 3.5 Pro the following month.
- June, July, August: it didn't come. Google kept shipping smaller, cheaper Flash models while Anthropic and OpenAI kept shipping flagships.
- Summer: a wave of senior departures, reportedly including Jeff Dean and Oriol Vinyals, according to one widely shared analysis of Google's year. Demis Hassabis was described as moving up to a chair role, with Koray Kavukcuoglu running Gemini day to day.
- End of September: Google abandoned Gemini 3.5 Pro altogether and skipped straight to a new generation.
That's a strange position for a company with Google's advantages. It has some of the best researchers in the world, its own TPU chips, enormous data centers and products already used by billions of people. Earlier this year, those advantages simply weren't turning into the model performance Google needed.
Then October happened
The comeback came fast:
- September 30: Google announced Gemini 4 Argon. It went to cybersecurity teams first.
- October 13: Argon opened up to developers.
- October 28: Google announced Gemini 4 Carbon, on the same day as Alphabet's quarterly results.
Carbon went straight to the top of the Artificial Analysis Intelligence Index. At launch, it scored 59, against 58 for Anthropic's Claude Opus 5.5 and 57 for OpenAI's GPT-6 Astra. On LMArena, where users vote on anonymous answers, Carbon was first and Argon second by early November. Google held the top two spots.
Then, on November 9, Anthropic released Claude Fable 5.2, which scored 65, several points clear of Carbon. (Artificial Analysis currently lists Carbon's high-effort setting at 60, fourth among the models it tracks, behind Fable 5.2's 65.)
So Carbon is no longer number one. But it's right up there, with two standout strengths.
Strength 1: it makes things up far less often
On a test of how often a model invents an answer when it doesn't actually know, Carbon fabricated about 11% of the time, according to figures cited in the analysis. Opus 5.5 and GPT-6 Astra were both around 50%. If you're building anything where a confident wrong answer is costly, that gap is enormous.
Strength 2: it's cheaper to run
On the same suite of tests, Carbon cost roughly $7.76 per task, versus about $10 for Opus 5.5. At enterprise scale, a 20%+ discount on frontier intelligence adds up.
The timing problem
Here's the catch that most launch-day coverage missed.
For the 12 days Carbon was number one, it was available only to a small group of cybersecurity defenders. The model everyone else could actually use was Argon, which scored 50, eight points behind Opus 5.5.
Google opened Carbon to paying developers, businesses and its top consumer subscription tier on November 10, the day after Fable 5.2 took the crown. Google's own announcement lists availability through the paid Gemini API in AI Studio, Gemini for Business and Google AI Ultra.
So Carbon was the best model when no one could use it, and became available once it had lost the lead. It proved Google can still build a model that competes with the very best, and after the summer it had, that wasn't a given. But being at the frontier without adoption isn't the same as being back.
Question 2: is anyone actually choosing Google?
This is where the picture splits in two.
On reach, Google is winning
- The Gemini app has passed 1 billion monthly users, which Sundar Pichai has called the fastest-growing product in Google's history.
- Apple's new Siri in iOS 27 is built on Gemini models (though not yet available in Europe).
- Google says its AI now processes more than four times as many tokens as it did a year ago.
On paid enterprise spend, Google is losing badly
Menlo Ventures' 2026 State of Enterprise AI report, based on a survey of nearly 400 U.S. enterprise decision-makers fielded September 8–26, estimates each provider's share of enterprise LLM API spending:
| Provider | 2025 | 2026 |
|---|---|---|
| Anthropic | 40% | 55% |
| OpenAI | 27% | 31% |
| 21% | 9% |
That survey was fielded before either Gemini 4 model was out, so it's the picture Google is trying to change, not a verdict on Carbon. But Menlo's explanation is telling: Google's API spend now skews toward cheaper Flash models because, for most of the year, it didn't have a top model matching Anthropic's or OpenAI's. (Worth knowing: Menlo discloses it's an investor in Anthropic.)
The lesson is simple. Using something by default isn't the same as choosing it. A billion people touching Gemini through Android, Search and Siri doesn't translate into engineering teams betting their products on the Gemini API.
The bigger market Google is missing
The rest of the Menlo report shows why that 9% stings so much. This isn't a niche market:
- Enterprise AI deployment spend reached roughly $177 billion by September 2026, up from about $33 billion a year earlier.
- Anthropic, OpenAI and Google together command 69% of deployment dollars, up from 50%, so the frontier labs are pulling away from everyone else.
- Coding ($34.0 billion) and chatbots ($25.0 billion) make up 76% of a $77.3 billion application market.
- Claude Code and Codex alone account for about $24 billion, or 70%, of coding spend.
- Open-source models handle about 19% of usage but only 4% of spend, and near-frontier options like xAI's Grok 4.7, Moonshot's Kimi K3, DeepSeek V4 and Z AI's GLM-5.3 together hold under 5% of API spend.
The takeaway from Menlo is blunt: for now, money follows capability. Buyers pay a premium for the most capable models even when much cheaper options exist. That's good news for Google in one sense, because it means a genuinely top model can win spend back. But it also explains the collapse. When Google's best available model trailed the leaders, enterprises didn't stick around to save money. They moved.
It also shows where Google's opening lies. Coding agents are the single biggest pot of money, and Google's own first-party coding tool has, until now, been split across competing products. If Antigravity running Carbon can close a three-point gap on autonomous coding, that's the fastest route back into enterprise budgets.
What investors see: landlord vs. vendor
Wall Street's view depends on which part of Google you look at.
The analysis lays out the quarter like this:
- Google Cloud revenue jumped sharply to about $30.7 billion, with roughly $561 billion in signed contracts still to deliver.
- Search, still around half of Alphabet's revenue, slowed from 17% growth to 12%.
- Google spent roughly $9 billion more cash than it brought in that quarter, partly due to a $5 billion EU fine.
- The day after results, the stock fell more than 12%, which the analysis calls its worst day since Google went public in 2004.
The landlord's dilemma
Here's where it gets interesting. Google isn't just building models. It's renting out the chips and cloud capacity other labs need, including the labs that are beating it. According to the analysis, Anthropic's IPO filings show more than $110 billion committed to Google Cloud.
Google has said it doesn't have enough capacity to meet demand. So every chip it rents to a rival:
- grows Google Cloud,
- is a chip Gemini doesn't get, and
- helps power the models winning the enterprise spend Google wants.
Google is winning as a landlord while losing as a model vendor. And for most of the year, it seems to have leaned toward the landlord business, signing chip deals with Anthropic and Meta, while Oracle said it would offer Google's chips in its own cloud. Compute constraints reportedly left fewer resources for areas like coding, and created tension between Google Cloud and DeepMind over who gets the chips.
Question 3: did Google fix what broke?
Earlier in the year, the case was that Google needed to fix three things to get back in the race: coding, compute and people. Here's the scorecard.
Coding: trending the right way, not a clean pass
Coding is where the enterprise money is. Menlo estimates coding is the largest AI app category, worth about $34 billion. By the accounts of Google's own engineers, it's the capability Google underinvested in.
- On a test of autonomous, real-world coding work, Claude Code running Opus 5.5 scored 62; Carbon running in Google's Antigravity tool scored 59.
- On LMArena's crowd-voted coding rankings, Carbon came first.
- Google is consolidating its tools, shutting down Jules, its separate coding agent, by year-end and folding it into Antigravity. That's a step toward fixing the problem of multiple internal teams building competing products.
Verdict: the gap has narrowed, but Anthropic still leads on the coding work companies actually pay for.
Compute: mixed at best
On paper, Google is spending more on infrastructure than almost anyone, second only to Amazon this year, with more planned for next year. But it's also renting chips to competitors, and it has been short enough on capacity that it started renting computing power from other providers just to keep up. There's no public sign that Gemini now gets a bigger share of Google's own compute.
Verdict: mixed.
People: still a no
This is the clearest of the three. According to the analysis, DeepMind used to hire about 12 people for every one who left; now it's about 2. Anthropic's ratio is cited at roughly 22 to 1.
There has been one notable return: Barret Zoph rejoined Google DeepMind in August as a VP of research, working on improving models after their initial training. That helps. But one hire doesn't reverse a 12:1 ratio turning into 2:1.
Verdict: not fixed yet.
Why this matters
If you build on AI, Google's position affects you more than you might think.
- Competition keeps prices down. A strong third frontier lab puts pressure on Anthropic's and OpenAI's pricing. Carbon's lower cost per task is exactly that pressure in action.
- Reliability is a real differentiator. An 11% vs. ~50% fabrication rate on "I don't know" questions is the kind of gap that matters for legal, medical, finance and customer-facing work.
- Vendor concentration is a risk. When two providers hold 86% of enterprise API spend, a pricing change or outage at either one hits a huge slice of the market. A credible Google lowers that risk.
What to watch next
Three signals will tell you whether Google is back for real:
- Adoption of Carbon. Now that it's generally available, do developers and businesses actually switch workloads to it?
- Chips for Gemini. Does Google keep more of its own TPU capacity for its models rather than renting it out?
- The next enterprise spending survey. Does that 9% share start climbing?
If all three move, the "Google is back" headline will be earned.
Practical takeaways for teams choosing a model
- Benchmark on your own tasks. Leaderboard leads of one to five points change constantly. Your workload is what matters.
- Test Carbon where hallucinations hurt. If your use case punishes confident wrong answers, its reported reliability is worth a serious trial.
- Price per task, not per token. Carbon's per-task cost advantage is the number to compare.
- Design for model portability. With leaders swapping every few weeks, avoid architectures that lock you to one provider.
Counterpoints and risks
- Benchmarks aren't adoption, but they aren't nothing. Leading the index, even briefly, shows Google's research engine works. Talent and compute problems can be fixed faster than raw capability.
- Distribution is a moat. A billion Gemini users and Siri integration give Google data and reach that rivals can't easily buy. "Default" may eventually become "choice."
- The cloud strategy may be the right one. Being the landlord to the winning labs is a very profitable position. Google may be hedging intelligently rather than losing.
- Several figures come from one analysis. Talent ratios, per-task costs and some financial details are as cited in a single widely shared breakdown; treat them as estimates until confirmed in primary filings.
FAQ
What is Gemini 4 Carbon? Google's newest frontier model, announced October 28 and generally available since November 10 via the paid Gemini API, Gemini for Business and Google AI Ultra.
Is Carbon better than Claude Fable 5.2? On the Artificial Analysis Intelligence Index, no. Fable 5.2 scores 65 versus about 60 for Carbon's high-effort setting. Carbon's reported advantages are a lower fabrication rate and lower cost per task.
Why did Google's enterprise share drop to 9%? Menlo Ventures attributes it to Google lacking a flagship that matched Anthropic and OpenAI for most of the year, so spending shifted to cheaper Flash models. The survey predates Gemini 4.
Why couldn't most people use Carbon when it was number one? Google first gave Carbon to a small group of cybersecurity defenders, following the same defenders-first pattern it used for Argon. Broad paid access only opened on November 10. In Google's consumer apps, Carbon currently requires the Google AI Ultra tier, while the cheaper everyday Gemini models are available on lower plans.
Is Google back in AI? On the leaderboard, yes. On enterprise adoption, compute allocation and talent retention, not yet.
The bottom line
Gemini 4 Carbon proved Google can still build a model that leads the field, and after a summer of delays and departures, that mattered. But being back takes more than a leaderboard win. It takes companies choosing your model on purpose, and right up to Carbon's launch, they were choosing Anthropic and OpenAI.
Until Google shows Gemini is getting the chips it needs and stops losing its people, a 12-day reign at the top will be hard to turn into real adoption. Google is back on the leaderboard. Whether it's back is a question the next few quarters will answer.





