Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

Google's Strongest Gemini Is Now for Everyone: Carbon Goes Wide — but It's Not for Every Job

Google opened Gemini 4 Carbon to everyone on November 10, two weeks after limiting it to cyber defenders in its Fairwind Program. The #1 ranks on CursorBench, WANDR, and AutomationBench, the /swarm multi-agent teams, and the 99% attack-block rate are real — but all of them are Google-reported, and the introductory $2/$10 price has no end date in the post.

Google's official Gemini 4 Carbon key art: the Gemini star and the words 'Gemini 4 Carbon' on a deep blue background, beside a translucent ice-like crystal.

Google's strongest Gemini is finally open to everyone

On November 10, Google made Gemini 4 Carbon broadly available: developers, enterprises, scientists, and consumers can all use it starting today. The blog post puts it plainly — this is "our strongest model," strong in cybersecurity capabilities and in long-horizon, multi-step agentic work. Read only halfway, that sounds like another routine model launch. But Carbon's story did not start on November 10. Two weeks earlier, it went to a small group of cybersecurity defenders through a program called Fairwind. What actually happened on November 10 was not a launch but an opening up — from a list of defenders into the hands of every paying developer and AI Ultra subscriber.

That gap in timing is worth noting first, because it explains what Carbon is for. Google gave it to the people who understand security best, collected feedback, hardened the safeguards, and only then opened it up. A model that goes to security teams first is a model whose capability is itself part of the risk — Carbon is good at finding vulnerabilities, proving they are exploitable, and writing verified root-cause patches. Handing that to everyone means handing out the spear and the shield at the same time. Letting defenders use it for two weeks first means letting the shield get familiar with the spear. The blog also says that with each new release, safety mitigations, agentic misuse prevention, alignment, and harm reduction are strengthened in proportion to capability. That sounds like boilerplate, but for Carbon it points at something concrete, which the safety section below takes apart.

There is one more line of positioning in the blog that deserves to be quoted in spirit: Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model's intelligence — complex, multi-step coding and debugging, and open-ended scientific research and discovery. For those tasks, Carbon spends more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops. In plain language: Argon is the daily driver, Carbon is the crew you call for the hard jobs. Pick wrong and you either waste the big model on small work or stretch the small model past its limits.

Google's official Gemini 4 Carbon key art: the Gemini star and the words 'Gemini 4 Carbon' on a deep blue background, beside a translucent ice-like crystal.

Not a launch, an opening: defenders first, everyone second

Here is the timeline. Around October 28, Google announced Gemini 4 Carbon but opened it only to a set of cyber defenders in the Fairwind Program. On November 10, Carbon became broadly available: paying developers use it through the paid tier of the Gemini API in Google AI Studio and Android Studio, enterprises through Gemini for business, and Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. Scientists and researchers can apply for a chance to use both Argon and Carbon through Antigravity's highest usage tier free of charge for six months — applications opened four weeks ago, thousands of scientists have already been onboarded, and Carbon has now been added for current and future participants.

This "defenders first, everyone second" order is unusual. The normal pattern is the reverse: open up to everyone first, then patch the safety measures as problems appear. Google could afford to do it backwards because half of Carbon's capability is safety capability — finding flaws, verifying they can be exploited, and generating verified root-cause patches. Opening that up to everyone at once would mean shipping the spear and the shield together. Two weeks with defenders first means the shield gets to learn the spear. The blog's claim that safeguards scale with capability is, for Carbon, a testable statement rather than a slogan, and the numbers are in the safety section below.

On price, Carbon matches Argon's introductory price: $2 per million input tokens and $10 per million output tokens. Google labels this an introductory price in the post, which means it is a time-limited offer — but the post does not say when the introductory period ends or what the price becomes afterwards. That is worth recording honestly: the price itself is public, its expiry is not stated in the post. Anyone planning to run heavy workloads on Carbon long-term should check the current pricing page rather than budgeting on today's number.

Three groups of benchmarks, each to be read separately

The blog's benchmark results come in three groups, and each group needs its own reading of what was actually measured. Copying the rankings alone would miss the point.

The first group is agentic workflows. Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, at more than 2X lower cost per task compared to the next best external model, according to Google. This is the weightiest set of numbers in the post, because these benchmarks test doing real work rather than answering questions: CursorBench tests coding agents, WANDR tests research agents, AutomationBench tests automation tasks. Three first places plus lower cost adds up to one claim — Carbon does not just win on benchmarks, it wins at getting work done. But note the qualifiers: the post does not name the "next best external model," and it does not spell out how cost per task is calculated. That is standard vendor-blog writing. The safe reading is that Carbon is genuinely strong on these three benchmarks, not that it is #1 on every agentic task.

The second group is high-precision knowledge work. Google partnered with Artificial Analysis to launch a new benchmark, the Multimodal Knowledge Agent (MKA), which tests how a model reasons natively over dense files with charts, images, blueprints, schematics, and tables in agentic workflows. Carbon ranks #2 on it, while being more than 6X lower in average cost per task compared to the top performing model. Two financial evaluations are also named: on Hebbia's E2E Financial Agentic evaluation, Carbon matched the performance of the best model tested on deep financial research; on Balyasny Asset Management's General Financial Intelligence benchmark, it ranked at the top out of 23 frontier model configurations, spanning legal expertise, agentic tool use and execution, coding and data engineering, and perception and context understanding. The judgment here is that Carbon's strength is not only writing code but also chewing through big, dense documents and producing polished spreadsheets, presentations, and reports in strict formats — the daily work of legal, finance, and enterprise operations. The detail that Carbon ranks #2 on MKA is actually reassuring: Google launched a new benchmark with a partner and took second place on it, which reads as more credible than sweeping every category.

The third group is cybersecurity, Carbon's home turf. Carbon for cyber, accessed through the Fairwind Program, leads on key benchmarks across finding (SEC-bench Pro), exploiting (ExploitGym), and patching (CWE-bench v1) vulnerabilities. There are three reported results from Google's own systems and partners. First, Project Fortify, Google's autonomous system spanning scanning, triage, and remediation: Carbon hit 100% recall on scanning and 95% precision on triage on an internal benchmark, and in open-source code repository testing it showed over a 25% relative improvement on Argon's already strong recall rate, surfacing previously undetected, real-world logic flaws. Second, Google's Cloud Vulnerability Research team used Carbon to prove that two bugs in a popular open-source software were in fact exploitable vulnerabilities, comparable in severity to the recent PHP2Shell exploit; using Carbon to build a proof of concept, the team helped maintainers re-classify the bug as critical and swiftly issue a patch, with details to be shared after the responsible disclosure process completes. Third, as part of Wiz's Scan for Good program, the team used Carbon to assess vulnerabilities in public digital infrastructure: scanning two major SaaS providers, the model uncovered 0-day vulnerabilities in under two hours, compressing weeks of vulnerability research to a single afternoon, and partnered with the impacted organizations to remediate in record time, helping to mitigate potential impact on tens of thousands of companies.

What these three results share is that they are about real vulnerabilities in real systems, not benchmark scores. The 100% recall and 95% precision figures come from an internal benchmark and should be read with a discount; but "two bugs re-classified as critical" and "0-days found in under two hours" are, if accurate, evidence that no benchmark score can replace. The other side must also be said: all of this is reported by Google and its partners, with no independent third-party reproduction in the post. Readers should treat these as vendor-reported field records, not industry consensus.

"/swarm": one prompt becomes a team

Released the same day as Carbon is "/swarm" in Google Antigravity (in preview). It turns a single prompt into a coordinated team of agents you can track in real time. According to the Antigravity blog, a single agent works well for scoped tasks but breaks down when a change spans dozens of files and test suites; dynamic subagents upgraded that into a primary agent spinning off isolated helper agents, but the pattern is still bottlenecked by a single coordinator context reviewing, synthesizing, and routing every step. "/swarm" changes the structure: a coordinator spins up specialized parallel workers in isolated git worktrees so edits never collide, plus adversarial verifiers dedicated to running tests and challenging diffs. Every agent collaborates directly on a shared, append-only message board, parking at zero token cost when idle and activating the instant an agent needs input.

Two details are worth unpacking. First, isolated git worktrees: the biggest engineering headache of parallel agents is two agents editing the same code and overwriting each other's changes, and worktree isolation is the standard fix — Google has now built it into the product. Second, adversarial verifiers: the verifiers are not there to rubber-stamp a test run, they are there to challenge the change — the same thinking as Carbon's own "run longer verification loops." Antigravity also ships a built-in Swarm UI: in the side pane you can watch the agent team debate trade-offs in real time on the Board feed, inspect live tool activity or jump into any agent's trajectory from the Members roster, and scrub through the interactive Graph timeline to see how the team converged from the initial brief to verified code. A live full-screen terminal board is available via "/swarm board" or Alt+S in the CLI.

On roles, Google's advice is explicit: assign Carbon to the roles that demand the deepest reasoning — architectural design, coordination, and auditing — while pairing it with faster, lighter models across parallel isolated workspaces. The Antigravity blog also includes an unusually honest line: inside Google, in five weeks of internal dogfooding of "/swarm," engineers ran swarms to tackle multi-package migrations, cross-subsystem root-cause investigations, and complex architectural redesigns; but even with zero-token idle waits and dynamic model hot-swapping, parallel multi-agent systems consume quota much faster than a standard chat. That is the most useful warning in the whole announcement for developers: the bigger the team, the faster the bill grows. Carbon already spends more tokens than Argon, and multiplying that by a whole team means the cost model has to be recalculated.

The science side: from formal proofs to proteins

The longest and most easily skipped part of the blog is Carbon's use in scientific research. It is worth reading closely, because it answers a real question: what exactly is a "strongest model" strong at?

First, code verification. Google researchers built a new agentic framework for formal verification called AlphaProof Code and put it to work on a major security challenge: mathematically proving the actual code underlying the cryptographic protocol of Signal. Using Carbon alongside other models, the researchers translated security goals to mathematical theorems. AlphaProof Code agents then used Carbon to generate millions of lines of code in Lean, a computer language for writing mathematical proofs, helping them prove key security properties in weeks — a task that historically takes expert teams years. The blog's wording is careful, saying the result "indicates Carbon's potential to unlock scalable formal verification," not that it has been achieved. That caution is right: formal verification is the gold standard for building secure software, and also the hardest kind of work to make routine. One success does not mean the method is mature.

Second, enzyme assay development. Provided with only lab notebooks and raw data, Carbon went beyond standard analysis to surface hard-to-interpret data anomalies, which it used to diagnose subtle wet lab experimental artifacts such as incomplete automated mixing in deep wells, leading to substantially improved reproducibility. It also improved methods to deconvolute chromatography peaks and caught trace sample carryover in research instruments, significantly accelerating the development of new assays and tightening the feedback loop between computation and the bench. What matters here is "finding anomalies" rather than "running the analysis": the model was not following a procedure, it was seeing problems in the data that people had missed.

Third, protein structure prediction. Structure prediction models like AlphaFold 3 use Multiple Sequence Alignments (MSAs) of evolutionary relatives to help infer how amino acid chains fold. The problem is that understudied viral, microbial, and eukaryotic proteins often have very few matches in standard databases. Working in a multi-agent setup, Google researchers used Carbon to autonomously discover remote homologs across petabase-scale unannotated sequence archives, environmental and viral sequence catalogs, and 3D structure databases. For example, across 7 low-homolog protein families from the CASP17 blind test set, Carbon increased the number of homologs retrieved by 40-fold on average, greatly increasing the information available to structure prediction models without need for retraining. The blog notes that broader benchmarking is still underway. Taken together, these three cases pin down Carbon's scientific positioning: it is not "better at answering questions," it is "better at finding answers where nobody knows them yet."

The safety side: what 99% and 0% are worth

The stronger Carbon gets, the harder the safety questions become. The blog devotes a full section to model safety, and the numbers there deserve a closer look than the benchmark scores.

The first set of numbers comes from simulated cyberattack testing. In simulated cyberattack testing environments seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent that can iteratively revisit its attack plans after each block. The safeguards lowered the per-attempt breach rate to just 0.8%, down from 3.1% on Argon. They also lowered the automated cyberattack query success rate to 2.17% in Gray Swan's testing, compared to 9.86% on Argon. And external testing by Trajectory Labs found zero new universal jailbreaks across 132 hours and over 13,000 requests. These numbers should be read like this: 0.8% is not 0%, and 2.17% is not 0% either — the safeguards are "sharply reducing" risk, not eliminating it. Google itself admits in the blog that prompt injection robustness remains an area for continued improvement, and that advanced automated red teaming has shown models could be further improved against sophisticated attacks in some distributions.

The second set of numbers is about honesty. On a 375-task agentic honesty benchmark, Carbon passed 92% of tasks, compared to 70% for Argon. Broken down: it declined to fill in missing experimental data 98% of the time and flagged methodological flaws 68% of the time. This matters most for people doing science and data analysis: a model's worst habit is not "not knowing" but "not knowing and making something up anyway." A model that refuses to fabricate data is more trustworthy than one that confidently invents it. The blog also notes that when Carbon was announced two weeks ago, it achieved a 0% prompt injection attack success rate on the Gray Swan IPI benchmark — the first time that milestone has been reached. The same paragraph says 0% does not mean the problem is solved, and red teaming continues to push the limits.

Put together, Carbon's safety story is complete: as capability grows, safeguards grow with it, and the effect is measurable — not a sentence saying "we take safety seriously," but a breach rate down from 3.1% to 0.8%, an attack query success rate down from 9.86% to 2.17%, and an honesty score up from 70% to 92%. But all of these are numbers reported by Google or by testing firms it commissioned, and independent verification is still missing. Readers should remember the direction (safeguards are getting stronger), not treat any single number as settled.

What it means for developers: Argon or Carbon

If you are an independent developer or a small team, Carbon is probably not your default choice today — Google says so itself: Argon remains the ideal choice for the majority of software engineering and knowledge work. Carbon costs more in token usage: it spends more tokens per task on exploration, reasoning, and verification. For everyday coding, bug fixing, documentation, and knowledge work, Argon is the better deal. The cases where Carbon is worth considering are the ones the blog already lists: multi-package migrations, cross-subsystem root-cause investigations, complex architectural redesigns, open-ended scientific research, and cybersecurity auditing — jobs where getting it wrong once costs a lot, and spending extra tokens on verification is worth it.

A practical way to choose is by verification cost. If a task is cheap to verify — run a test, look at the rendered result, and you know whether it is right — use Argon or an even lighter model; if it is wrong, fix it and move on. If a task is expensive to verify — one architecture decision shapes the next six months, one missed vulnerability in a security audit becomes an incident, one wrong scientific conclusion wastes months of experiments — then it is worth letting Carbon spend more tokens on verification loops. Carbon's value is not in being "smarter" but in being "more willing to spend effort verifying." That distinction decides where it belongs in your workflow: not replacing Argon, but sitting after Argon, handling the high-leverage steps that deserve repeated verification.

One more angle is worth adding. On the same day, Anthropic's Claude Fable 5.2 had just taken first place on the Artificial Analysis Intelligence Index (covered in an earlier article on this site), and Google responded by opening Carbon up to everyone. That is not a coincidence, it is how frontier model competition works: every lab uses the word "strongest," but the yardstick differs each time — Fable 5.2 leads on a general intelligence index, Carbon leads on agentic workflows and cybersecurity. For developers, the right move is not to pick a side but to pick per task: coding, agents, and security auditing each have their own best model. Carbon going broadly available at least makes that choice a real one, with a usable option instead of a name on a Fairwind list.

Three things the blog did not say

First, how long the introductory price lasts. The blog only says Carbon shares Argon's introductory price ($2 per million input tokens and $10 per million output tokens), labeled as introductory, without saying when it ends or what comes after. Before budgeting long-term on this number, check the current pricing page. The number is right today and may not be tomorrow.

Second, the free tier. Every channel the blog lists is paid: the paid API tier, Gemini for business, and AI Ultra subscriptions. Whether free users and AI Pro users can use Carbon is not stated — nor does the blog say they are excluded. That is a blank, not a "no." If you are an ordinary user who wants to try Carbon, the most direct route is to look at AI Ultra, not to hunt for it in the free tier.

Third, independent verification. The #1 rankings on CursorBench, WANDR, and AutomationBench, the #2 on MKA, the leads on SEC-bench Pro, ExploitGym, and CWE-bench, and the 99% block rate and 0.8% breach rate — all of it is reported by Google itself or by firms it commissioned. Artificial Analysis is independent, but MKA is a benchmark Google launched together with Artificial Analysis, and that partnership itself is worth noting. How much stronger Carbon really is than Argon, and than its competitors, will only be known when independent teams run it on their own tasks. Until then, the blog's numbers should be read as vendor-reported figures, not industry consensus.

Back to the phrase "strongest model." What matters most about Carbon going broadly available is not any single benchmark score: a model that two weeks ago was only in the hands of cybersecurity defenders is now in the hands of every paying developer, and a model designed for "tasks that test the limits of intelligence" can now be tried by anyone. The real test is not on launch day but in the months that follow — how many teams actually use it for jobs Argon cannot handle, how many security teams actually find vulnerabilities with it that they had missed before, and how many scientists actually make discoveries with it that they could not make before. Those answers are not in the blog. They are in the usage records.

Views 6Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
Against a black background, a conductor seen from behind raises a baton, with swirling trails of light circling beneath him like an orchestra being conducted
News
One API for 45 Models, and It Switches Providers When One Fails: Perplexity's Orchestrator Wants to Be the Switchboard

On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.

Developer WorkflowProduct LaunchModel Updates
In a dim office, a developer wearing glasses sits with his back to the camera before several monitors showing code, lit by blue and purple light
News
The Fourth Gemini 4 in Seven Weeks: Neon Sits Between Carbon and Hydrogen, Selling Speed With Enough Smarts

On 18 November 2026, Google released Gemini 4 Neon: third overall on the Artificial Analysis Intelligence Index (61) and first on the Coding Agent Index (58.6), priced at $3 per million input tokens and $15 per million output. It sits between the heaviest Carbon and the lightest Hydrogen, conducting teams of Hydrogen subagents in Antigravity — yet per token it is November's priciest Gemini 4, only becoming the cheaper middle option after 1 December.

Product LaunchModel UpdatesAI Coding