Google Restricts Gemini 4 Carbon Launch Despite Top Artificial Analysis Ranking

<p>Google’s new <strong>Gemini 4 Carbon</strong> model took first place on the Artificial Analysis Intelligence Index with a score of <strong>59</strong>, edging past Anthropic’s <strong>Claude Opus 5.5</strong> (<strong>58</strong>) and OpenAI’s <strong>GPT-6 Astra</strong> (<strong>57</strong>), though Alphabet (<strong>GOOGL</strong>) is restricting its immediate commercial release. Shares of Alphabet gained <strong>1.5%</strong> following the announcement on Wednesday.</p><h2>Restricted Release and Cybersecurity Benchmarks</h2><p>Google is initially limiting access to <strong>Gemini 4 Carbon</strong> to vetted cyber defenders through its <strong>Fairwind Program</strong>, offering no firm date for broader availability to developers or consumers. The phased rollout stems from safety precautions, particularly given Carbon's performance on security-focused benchmarks.</p><p>On the <strong>ExploitGym</strong> test, which evaluates a model's capacity to verify usable software vulnerabilities, <strong>Gemini 4 Carbon</strong> achieved a score of <strong>54.8%</strong> with safety filters disabled. In comparison, OpenAI’s <strong>GPT-6 Astra</strong> scored <strong>42.4%</strong> and Anthropic’s <strong>Claude Opus 5.5</strong> registered <strong>34.5%</strong>. Google is participating in voluntary pre-release testing with the U.S. government, mirroring the deployment schedule of its prior model, <strong>Gemini 4 Argon</strong>, which launched for paying developers on <strong>October 13</strong>.</p><h2>Pricing Structures and Competitive Performance</h2><p>Despite leading in security metrics, Carbon trails competitors in software engineering benchmarks. On the <strong>FrontierSWE 2</strong> coding test, <strong>GPT-6 Astra</strong> earned <strong>65.5%</strong> and <strong>Claude Opus 5.5</strong> scored <strong>62.3%</strong>, outperforming Carbon's <strong>61.2%</strong>.</p><p>Pricing for <strong>Gemini 4 Carbon</strong> starts at <strong>$2</strong> per million input tokens and <strong>$10</strong> per million output tokens, though rates are scheduled to double to <strong>$4</strong> and <strong>$20</strong> respectively. The pricing comes as OpenAI reduced <strong>GPT-6.1 Terra</strong> costs to <strong>$1</strong> input and <strong>$5</strong> output on <strong>October 26</strong>, while Anthropic prepares for a public listing following its recent <a href="/news/anthropic-discloses-1317b-cash-and-587b-commitments-in-s-1-ipo-filing">S-1 IPO filing</a>. The competitive atmosphere follows broader tech market dynamics where major firms face questions over <a href="/news/oracle-and-openai-avoid-capital-questions-despite-664b-backlog-and-cash-burn">capital expenditures and infrastructure commitments</a>.</p><ul><li><strong>Benchmark Performance</strong>: <strong>Gemini 4 Carbon</strong> scored <strong>59</strong> on the Artificial Analysis Intelligence Index, topping <strong>Opus 5.5</strong> (<strong>58</strong>) and <strong>GPT-6 Astra</strong> (<strong>57</strong>).</li><li><strong>Cybersecurity Lead</strong>: Carbon scored <strong>54.8%</strong> on <strong>ExploitGym</strong>, beating <strong>GPT-6 Astra</strong> (<strong>42.4%</strong>) and <strong>Opus 5.5</strong> (<strong>34.5%</strong>).</li><li><strong>Tiered Pricing</strong>: Initial token fees of <strong>$2</strong> input / <strong>$10</strong> output will later double to <strong>$4</strong> / <strong>$20</strong>.</li></ul><h2>Why It Matters</h2><p>Google's decision to gate Gemini 4 Carbon behind the Fairwind Program highlights growing regulatory and national security sensitivities around offensive AI capabilities. By scoring high on vulnerability testing while lagging in standard coding tasks like FrontierSWE 2, Carbon represents a specialized asset that carries potential dual-use risks. As frontier labs compete on price and performance, safety-driven launch delays could test developer patience against lower-cost alternatives like OpenAI's GPT-6.1 Terra.</p>