Google’s new Gemini 4 Carbon model cuts the rate at which it invents false information nearly in half — and that still isn’t enough to make it a good coding assistant. In the most detailed independent review to date, developer Theo Browne found Carbon’s false-claim rate fell from 36% for the previous Gemini model to 18%, a number that exactly matched what DeepMind engineers privately told him they had measured before launch, speaking on the t3.gg podcast. The problem: most competing models, including Anthropic’s Opus, fabricate at 0–1%.
That single gap explains why Browne spends an hour and a half dissecting a model he ultimately won’t use for real work, before reversing course in the final third over one specific workflow. The story of Carbon is a lesson in how AI models are now sold — on benchmark crowns and introductory pricing — versus how they’re actually used, where latency, cache behavior, and whether the model invents a database column name determine whether a tool is worth the monthly bill.
The hallucination problem: better, still bad
Browne’s core complaint is fabrication in codebase analysis and code review. On a Rust architecture review, the previous Gemini model (called Argan) made 36 claims, 24 of them false. Carbon made 79 claims with 23 false — more opinions, delivered with more confidence, but still 23 fabrications where most competing models produce zero to one, and a handful produce three to four. On his T3 codebase analysis, Argan failed 47 of 92 claims; Carbon failed 23 of 86.
“He knows the vibe of what this repository does,” Browne says of Carbon. “He’s just not going to tell you with perfect accuracy what’s in it.”
| Model (Browne’s evals) | False-claim rate |
|---|---|
| Gemini 4 Argan | 36% |
| Gemini 4 Carbon | 18% |
| Muse-line “Wonder” | 14.5% |
| Most other models | 0–1% (a few at 3–4%) |
He attributes part of the earlier hallucination problem to bugs in the old ACP bindings — the integration layer between his tool and Google’s model — which weren’t passing data such as line numbers to the model the way the official Anti-Gravity CLI does. DeepMind fixed that. But even after those fixes, Carbon still invents column names, line numbers for code it didn’t properly read, and architectural details that aren’t true.
His evals also don’t fully capture the damage. Browne points out that official hallucination benchmarks don’t measure this failure mode at all — the thing that makes a coding model untrustworthy in practice is exactly what the leaderboards skip.
“Fast” is a number, not a feeling
Browne rejects the “fast model” framing outright, arguing tokens-per-second is a metric users don’t experience. In Artificial Analysis data, Carbon’s average task time was 22 minutes on low reasoning effort and 27 minutes on high. Opus ranged from 3 minutes 1 second on low to 25 minutes at max. Carbon is consistently the slowest model, and reasoning-effort settings barely change the round-trip time.
“Everybody’s saying, ‘Well, it’s a fast model.’ Shut up. You have no idea what you’re talking about. Like, I’m sorry if you think TPS means speed.”
He calls using it “genuinely obnoxious” and “miserable to use.”
The price looks reasonable until you calculate what you actually pay
Pricing compounds the latency problem. Carbon launches at $2 per million input tokens and $10 per million output, with a 10-cent context-cache read — but that’s an “introductory price” that doubles on December 1, 2026, to $4 in and $20 out with 20-cent cache reads.
| Gemini 4 Carbon pricing | Through Nov. 30, 2026 | From Dec. 1, 2026 |
|---|---|---|
| Input (per M tokens) | $2 | $4 |
| Output (per M tokens) | $10 | $20 |
| Cache read (per M tokens) | $0.10 | $0.20 |
Browne’s read: the discount window is under a month, versus the two-to-three-month timelines Google previously gave. He says the post-promotion price is Opus-level pricing for a much less pleasant model.
“Normally intelligence is supposed to get cheaper over time but Google truly innovates on shitty things.”
The real cost leak isn’t the sticker price. It’s caching. Unlike OpenAI (explicit cache writes) and Anthropic’s Claude (explicit writes plus configurable time-to-live), Gemini caches “when they feel like it.” In his tests, a 5-second window carried a 20% chance of a cache miss; a 30-second window 34%; a 120-second window more still — producing up to a 50% increase in requests in the worst case.
Browne’s standing argument cuts through the pricing table entirely: token price is irrelevant. What matters is cost per task, cache efficiency, and how many tokens the model burns getting there. Carbon, he says, “burns tokens like nothing I’ve ever seen.”
“The token cost does not matter. The cost per task is what matters, especially when you consider how many tokens are being used in a given task.”
He also notes subscription subsidies on the $200 Gemini plan are historically worse than Anthropic’s $200 plan, which he says feels near-unlimited for real work — a gap that decides budgets for individual developers as much as enterprises.
The reversal: one security fix where Carbon beat every rival
The episode’s turning point is a vulnerability report filed against T3 Code by an independent developer, Brandon (“Pseudo Brandino”), who obtained Anthropic’s Mythos access through a security program and audited the project. He found two niche issues — one nearly impossible to trigger without a targeted attack, one somewhat easier to induce by tricking an agent.
Browne’s attempts to fix them with other models went badly. Fable 52 was handed the report and asked how important the fix was. Opus 4.8, which he says “wasn’t even good when it came out,” produced nothing. Mythos diagnosed the problem correctly but proposed a fix he considered unacceptable: requiring the user to walk to the second machine and generate a six-digit code to approve a link. Browne says he hasn’t physically looked at his machines in months — the fix would destroy the feature’s value. He went back and forth until Mythos agreed to treat trusted T3 Code links through the company’s own auth layer differently.
Carbon, tried almost as an afterthought, described the issue correctly and proposed exactly the approach Browne wanted: a hard separation between known-machine flows and raw-URL flows, plus tighter URL retention in the peer-link section — with no heavy cryptographic redesign. It also flagged two extra issues the report hadn’t raised.
“You do not need heavy cryptographic redesign to kill this vulnerability cleanly. And doing the separation properly actually fixes two extra issues that the report didn’t call out.”
A council review of Carbon’s first pass rated it well by competing AI models:
| Reviewing model | Score for Carbon’s security fix |
|---|---|
| Opus 55 | 9/10 |
| Muse | 9/10 |
| “61 Soul” | 7.5/10 |
| “Terara” | 7/10 |
A separate Windows remote-update investigation he handed Carbon drew 8–9/10 from reviewing models. A Microsoft engineer had reported the update problem in the first place; Browne says he’d been having trouble reproducing it before handing the work to Carbon.
His reaction was genuine surprise. “The fact that I had the best experience addressing these reports with this model is kind of insane,” he says. “Like genuinely, I did not expect this.”
One caveat he flags himself: his early-access status may mean he isn’t flagged as aggressively as other users, which could undercut the incident-response result. And the model still took roughly 25 minutes on a task he had running.
UI and 3D: templated and visibly broken
Google’s launch claims about full-stack, 3D, and UI design did not hold up in Browne’s tests. With Anthropic’s design skill enabled, Carbon’s front ends were worse than its native output. Natively, it defaults to a recurring template — pill-shaped section, two-line centered text, logo, a “begin” button, four links in the middle, a cropped app screenshot — repeated across generations with only the color changing. He counted roughly two of the samples as genuinely different from each other.
On a fish model 3D demo, the first version was “embarrassing” next to Fable 52’s output. Told to make it beautiful, Carbon produced good-looking water and broke everything else. After a roughly 25-minute run, the result was “less bad, still not good”: a questionable propeller design and rings floating off the fish instead of around it.
His conclusion: Google is not ahead on 3D despite his prior impression, and the model is “nearly exclusively useful for cyber security dev work if you have access to other things.”
DeepMind’s responsiveness is the strongest lab relationship he has
Browne devotes substantial airtime to praising DeepMind’s engineering support, which he ranks first among labs for fixing issues he reports — ahead of OpenAI and SpaceX AI (which now owns Cursor). When he complained about unclear policy on using a Gemini subscription in third-party tools, they fixed the policy the next day. When he reported absurd hallucination rates in T3 Code, they audited the ACP bindings, found the data-passing bugs, and shipped fixes. They added a special T3 Code identifier so his tool gets the same treatment as Zed, the popular code editor.
The mechanism matters commercially: Google doesn’t permit using a Gemini subscription for inference in arbitrary harnesses like PI, but it does allow another tool to control Anti-Gravity through ACP bindings. Those bindings were previously a separate binary T3 Code offered as an install option; DeepMind has since folded them into the Anti-Gravity CLI. Browne notes only two labs have landed code directly in T3 Code — SpaceX (the original Grok build implementation) and DeepMind.
He also complains that the Anti-Gravity CLI is no longer open source like the separate Gemini CLI, and that Anti-Gravity’s interface still exposes his email address everywhere.
The verdict: right direction, wrong default
Browne’s overall assessment is that Carbon is a large step up in day-to-day feel — he says the gap from Gemini 3.8 Flash to Argan was smaller than Argan to Carbon — and he’s considering promoting it out of “jester” status in his model council so it can track review feedback over time. But he wouldn’t pick it as a coding default.
“Would I ever pick this model as my default for coding? No. Absolutely not. It’s just Opus is unbeatable still.”
An irony he flags: a model made by the company with one of the biggest codebases in the world struggles on big codebases, and none of the other models do.
“It’s kind of crazy that a model that Google, the company with one of the biggest code bases in the world, made struggles so much in big code bases, but it does and none of the other models do.”
He argues Google should have withheld Argan and shipped Carbon first — a sign, he says, that the company is learning the right lessons and may release a genuinely competitive model in the not-too-distant future. He also notes Google co-built a new benchmark with Artificial Analysis, the multimodal knowledge agent, and placed only second on it. And he’s skeptical of endorsements from the Open Code team, who also praised Argan — a model Browne says he had to revert entirely.
“I’m not going to say Google cooked, but it’s significantly less bad than before.”
What this means for anyone paying for AI
The launch strategy — discount hard for under a month, then double input, output, and cache-read prices on December 1, 2026 — shapes what enterprises will actually pay. Browne’s earlier read was that such discounts existed to steer users off older models while Google reprovisioned capacity, with two-to-three-month windows. The compressed timeline changes that calculus. Combined with unreliable implicit caching and high token burn, the effective cost per task is likely to exceed the sticker comparison with Opus even before the increase.
Google’s positioning, in its own launch materials, is a two-tier product line: Argon for volume work, Carbon for high-precision tasks in legal and financial analysis, full-stack development, and 3D design. Carbon is also the engine behind the “Carbon for cyber” edition, gated behind an application process, which Google-owned Wiz used to surface previously unknown zero-day vulnerabilities in two large software vendors in under two hours. That’s the same security strength Browne stumbled onto — and the one place where his review and Google’s pitch converge.
The unresolved question is whether Carbon’s security edge is a durable differentiator or an artifact of his privileged access. What’s not in doubt is the direction of travel: hallucination rates halved, task adherence improved, and DeepMind is fixing integration bugs within days. For investors and developers weighing where to spend, the signal is that Google is now a credible second supplier for security-sensitive work — but not yet a reason to switch defaults, and not a bargain once the December price step lands.
