Gemini 4 Carbon arrives just in time to feel like yesterday's news in the fastest model cycle yet

Google has released Gemini 4 Carbon to general availability for paying developers and business customers, following a two-week testing phase limited to selected security experts. The launch lands in a crowded stretch of model releases: Anthropic introduced a new top model, Fable 5.2, which edges out Carbon on the Artificial Analysis index. Google positions Carbon not as a replacement for its earlier Gemini 4 Argon but as a companion for tasks that push hardest on model intelligence.
Both models carry the same token price, which makes the segmentation worth reading carefully. Google describes Argon as the better pick for most software engineering and knowledge work, while Carbon is aimed at complex multi-step coding, debugging and open-ended scientific research. Carbon's advantage is depth of reasoning, and it pays for that depth in tokens: Google says it consumes more tokens than Argon to run longer verification loops and reason further on hard debugging or research problems. In practice, this suggests a cheaper sticker price per token can still translate into a larger bill per finished task, which is exactly the gap per-task cost benchmarks are meant to expose.
On cost, Artificial Analysis measured Carbon at $7.76 per solved task, computed at the standard rate of $4 per million input tokens and $20 per million output tokens. That figure puts Carbon below the older budget model Flash 3.8 while performing better. For comparison, the same test placed Opus 5.5 Max at $10.04 per task and GPT-6 Astra Max at $6.25. Fable 5.2 runs as high as $26.50 at its maximum tier; at its High tier it costs roughly $9 per task and still leads the overall index by two points, 62 to 60. Google is offering an introductory rate of $2 per million input tokens and $10 per million output tokens, which is half the standard input price and half the standard output price.
Access runs through the Gemini API in Google AI Studio, in Android Studio, and through Gemini for Business. The article also flags a Flash variant of Gemini 4 as a likely near-term addition, though no date is given.
Google published new benchmark claims alongside the launch. Carbon takes first place on the coding benchmark CursorBench 4.0 with a score just under 60 percent, ahead of Fable 5.2 and Gemini 4 Argon, with GPT-5.6 Sol and Grok 4.7 further back. Google says the result comes at less than half the per-task cost of the next-best model, with average costs in the roughly $8 to $10 range while more expensive configurations of rivals score lower. Carbon also ranks first on Perplexity's WANDR and on Zapier's AutomationBench. On a new benchmark built with Artificial Analysis, the Multimodal Knowledge Agent test for dense documents containing charts, blueprints and tables, Carbon finishes second behind Fable 5.2 and just ahead of Opus 5.5, at less than one-sixth of the top performer's average cost.
The most concrete results come from security work during the Fairwind Program early-access phase. Working with security firm Wiz's Scan for Good program, Carbon found real zero-day vulnerabilities while scanning two large SaaS providers in under two hours. Google's own cloud security team used the model to write working proof-of-concept exploits for two known bugs in a popular open-source project, which the company says was needed to prove the bugs were critically exploitable, comparable to the PHP2Shell vulnerability. Inside Google's autonomous security system Project Fortify, Carbon raised the detection rate on open-source codebases by more than 25 percent over Argon. Google says Carbon leads the SEC-bench Pro, ExploitGym and CWE-bench v1 benchmarks for finding, exploiting and patching vulnerabilities, but those claims apply only to the Fairwind version and come without specific numbers.
Google also published a full safety report and model card. On an internal honesty test of 375 tasks, Carbon passes 92 percent, against 70 percent for Argon, and declines to fabricate missing experimental data in 98 percent of cases. In simulated cyberattacks it blocked 99 percent of attack paths from adaptive red-teaming agents and cut the per-attempt breakthrough rate from 3.1 percent with Argon to 0.8 percent. External testers at Trajectory Labs ran 132 hours and more than 13,000 queries without finding a single universal jailbreak. On the Gray Swan IPI benchmark, Google reports a prompt injection success rate of zero percent, which it describes as a first for any model. That milestone deserves a caveat: a benchmark where every model eventually scores zero tells you little about new attacks, so it may simply be saturated. The honest reading is that this particular test no longer separates models, not that prompt injection as an attack class is closed.
Separately, Google is previewing a "/swarm" feature in Antigravity, its counterpart to Claude Code. A single prompt becomes a coordinated team of agents whose work can be tracked in real time. Carbon takes the roles that need deep reasoning, such as architecture design, coordination and auditing, while faster, lighter models run in parallel isolated workspaces and act as adversarial verifiers. Google's demos include a 3D game exploring the Mountain View campus as a pelican on a GBike, an interactive 3D reconstruction of the Eiffel Tower's construction from historical photos, and a playable 18th-century fortepiano with seat-specific spatial acoustics. The economics of such swarms are contested: OpenAI and Anthropic have recently struck cautious notes, and researchers are skeptical too. The apparent benefit is speed through parallelization, but orchestrating many agents reliably is complex, and token spend multiplies with every worker and verifier. A likely trade-off is that swarms buy latency reductions at the cost of predictability and bill size.
Why it matters: Teams choosing a frontier model now face a genuinely close race on capability, with Carbon competitive on coding and automation benchmarks and substantially cheaper per solved task than Fable 5.2 at its top tier, while Anthropic still holds a narrow lead on the overall index. The token-consumption caveat means buyers should evaluate cost per completed task rather than per-token pricing, since Carbon's deeper reasoning can erase its headline discount. The swarm preview, if it holds up outside demos, points toward a workflow where a strong model plans and cheap models execute in parallel, which is an inference about direction rather than a reported result.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Google is making Gemini 4 Carbon generally available after a two-week testing phase. The model ranks first in several benchmarks, in some cases ahead of Anthropic's new Fable 5.2, at less than half the per-task cost. Google is also previewing coordinated agent teams through the "/swarm" feature, which split up complex tasks in real time. The article Gemini 4 Carbon arrives just in time to feel like yesterday's news in the fastest model cycle yet appeared first on The Decoder .