TL;DR: Google pulled Gemini 4 Carbon out of its gated cybersecurity-only preview on November 10 and opened it to everyone, from AI Studio developers to $99.99-a-month Gemini app subscribers. The headline number is the price: Carbon’s introductory API rate of $2 per million input tokens and $10 per million output tokens runs through November 30, a quarter of what Anthropic charges for Claude Fable 5.2. Even the standard post-promo rate, $4/$20, matches Claude Opus 5.5 instead of beating it, and Carbon still trails Fable 5.2 on the main intelligence benchmark that matters most to developers.
From cybersecurity preview to a public model in two weeks
Google first unveiled Gemini 4 Carbon on October 28, but almost nobody outside a small group of invited researchers could touch it. The model lived inside DeepMind’s “Fairwind” program, a gated environment built for professional cyber defenders who wanted to test an AI model with fewer of the usual safety guardrails. Google said that group used Carbon to find zero-day vulnerabilities and make headway on protein structure prediction, and a specialized cyber-focused version of the model is still only available through that same restricted Fairwind access.
On November 10, Google widened the door considerably. Carbon is now reachable through Google AI Studio and Android Studio for developers, through Gemini for business for enterprise customers, and through the consumer Gemini app, Search’s AI Mode, and Google’s Antigravity coding platform for anyone paying for Google AI Ultra. That subscription runs $99.99 a month for the standard “5x” tier or $199.99 a month for the higher-usage “20x” tier. Google is also giving roughly 10,000 scientists free access to its tools for six months, and the company says thousands have already been onboarded.
Antigravity picks up a new preview feature alongside the rollout, an internal “/swarm” mode where a team of AI agents splits a coding project into pieces and checks each other’s work before handing it back. Google’s materials also point to specialized agents built on Carbon, including the vulnerability-patching tool CWE-bench targets and the scientific-discovery system AlphaEvolve, though the company hasn’t published a country-by-country availability schedule for the paid API outside the US.
The pricing math that actually matters
Here’s where Carbon gets interesting for anyone running production AI workloads instead of just kicking the tires. Google’s introductory API pricing, in effect through November 30, is $2 per million input tokens and $10 per million output tokens. After that window closes, the rate doubles to $4 and $20. Google is also offering a 95% discount on cached input tokens, which would put effective cached pricing at roughly 10 cents per million tokens during the promotional period.
Stack that against Anthropic’s current lineup and the gap is obvious. Claude Fable 5.2 and Mythos 5.2, which Anthropic launched just a day before Carbon’s wide release, cost $8 per million input tokens and $40 per million output tokens at standard rates. Carbon’s introductory price is a quarter of that. Even once the promotion ends and Carbon settles at $4/$20, it’s still half the cost of Fable 5.2 per token. Claude Opus 5.5, Anthropic’s workhorse model for daily tasks, prices in at $4/$20, which means Carbon’s standard rate merely matches it rather than beating it, and Carbon undercuts Opus 5.5 only during the introductory window.
Gemini 4 Argon, Google’s earlier and less capable Gemini 4 release from October 13, carries the exact same pricing tiers as Carbon. That’s a deliberate signal: Google wants developers choosing between its own two models on capability rather than cost, while using price as the lever against Anthropic and OpenAI.
Carbon still isn’t the smartest model on the board
Price is one axis. Raw capability is another, and that’s where Carbon’s story gets more complicated. On Artificial Analysis’s widely-watched Intelligence Index, Fable 5.2 currently sits at 65 running at maximum effort, Carbon trails at 60 on its high-effort setting, Claude Opus 5.5 scores 59, and GPT-6 Astra comes in at 57. Carbon actually debuted in October at the top of that leaderboard with a score of 59, before Artificial Analysis revised its scoring methodology for one benchmark component and re-ran every model, a reshuffle that let Fable 5.2 pass it.
Where Carbon does pull ahead is cost-efficiency per completed task, not just per token. Artificial Analysis found Carbon’s high-effort setting costs about $7.76 per task on average against $26.54 for Fable 5.2 at maximum effort, and Zapier’s AutomationBench scored Carbon’s high-effort mode at 53.88% task completion for roughly $2.12 per task, edging out Fable 5.2’s 51.45% at more than double the price. On raw coding benchmarks the picture is mixed: Carbon leads on Terminal-Bench 4.0 (65.2% versus Fable’s 61.6%) but falls behind on FrontierCode 2.0, where Fable’s best run hits 49.1% against Carbon’s 40.6%.
Google has also made some benchmark claims it hasn’t fully backed up. The company says Carbon took first place on CursorBench 4.0 and Perplexity’s WANDR evaluation at under half the cost of the next-best model, but launch materials didn’t include the actual scores, and the public CursorBench leaderboard doesn’t list a Carbon entry at all. On the cybersecurity side, Google’s own pre-launch testing put Carbon at 88.2% on SEC-bench Pro for finding vulnerabilities, and the security firm Wiz reportedly used the model to turn up previously unknown vulnerabilities while scanning two major SaaS providers in under two hours. Anthropic’s competing cyber numbers for Mythos 5.2, by contrast, claim 89.6% rediscovery of known vulnerabilities in internal testing, a category where Google hasn’t published a directly comparable Carbon figure.
Why Google is racing to ship this now
The timing isn’t a coincidence. Carbon’s wide release lands one day after Anthropic’s Fable 5.2 and Mythos 5.2 launch knocked Carbon off the top of the Artificial Analysis leaderboard, and it arrives while Anthropic is deep in preparations for what could be the largest stock market listing in history, with a price range that implies a valuation near $1 trillion. Every model release from either company right now gets read partly as a technical announcement and partly as a signal to investors and enterprise buyers about who’s actually winning the AI race heading into that listing.
Google’s strategy here looks less like “beat Anthropic on every benchmark” and more like “make the cost-benefit math uncomfortable for Anthropic’s customers.” A model that’s merely competitive on intelligence but dramatically cheaper per token, and noticeably cheaper per finished task, is a real pitch to any company running high-volume AI workloads where token costs compound fast. Whether that’s enough to pull developers away from Fable 5.2 depends on how much raw capability actually matters for a given use case versus how much the invoice does.
The bottom line
Gemini 4 Carbon’s wide release is a genuine shot at Anthropic’s pricing, not just another incremental model drop. Google undercut Claude Fable 5.2 by as much as 75% on paper during the introductory window, and it’s betting that cost-per-task, not just cost-per-token or a single intelligence score, is what will actually move enterprise buying decisions. But “cheaper” and “smarter” aren’t the same claim, and Carbon still isn’t the top model on the leaderboard that matters most to developers choosing where to build. If you’re already deep in the Claude ecosystem, this isn’t a reason to switch overnight. If you’re price-sensitive and your workload tolerates a slightly less capable model, Carbon just became a lot harder to ignore.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
