On Oct 28 US time, about three hours before Alphabet’s Q3 call, Google launched Gemini 4 Carbon, topping Artificial Analysis’s Intelligence Index at 59—ahead of Claude Opus 5.5 and GPT-6 Astra. Gemini 4 Argon, opened to paid users only two weeks earlier, scored 50. Carbon’s jump is real, but rollout stays gated, early users report much higher latency, and same-day earnings showed strong Cloud growth alongside a second straight quarter of negative free cash flow—with after-hours shares down more than 10% at one point. A leaderboard win does not automatically “save face” with markets.
Benchmarks: Back on Top After ~Half a Year
Gemini last shared the crown around April (3.1 Pro with GPT-5.4 and Opus 4.7). Carbon’s 59 on Intelligence Index v5.0 is +9 vs Argon, ~1 above Opus 5.5 and ~2 above GPT-6 Astra, with four firsts across nine components—strong on hard reasoning, science, QA, and automation.
Coding: Coding Agent Index 59 (Opus 5.5 at 62); Terminal-Bench 4.0 57.1%→65.2%; WebDev Arena ~1768 Elo (+~89 vs Argon), fourth place. GraphWalks long-context 84.2%→96.2% vs Astra’s 71.8%. Google says it tuned for multi-day agent workflows over large codebases. Claude/GPT still often lead on coding agents, some SWE, and computer-use—see our GPT-6.1 Terra and Gemini 4 Pro leak pieces.
Safety, Honesty, Pricing—Defense Orgs First
Per Google: SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench fix 71.0%; internal secure-codegen success more than 2× Argon. Gray Swan indirect prompt injection: 0% attack success for Carbon (up to 15 tries) vs ~0.7% Argon, ~1% Opus 5.5, ~8.5% Astra. It also cites 0% Terminal-Bench cheat success, ~0.4% cheat attempts, ~half the unsupported assertions, and ~80% fewer high-severity behavioral alerts.
Promo pricing matches Argon: $2 / $10 per million input/output tokens (95% cache discount), then $4 / $20. At list rates, ~$7.76 per composite eval task—near Argon, ~23% below Opus, ~20% above Astra. Access starts with select cyber-defense orgs this week, then Fairwind trusted partners; no public GA date. More: AI News.
Stronger and Slower: Latency and Hallucination Trade-offs
Early Arena/preview feedback: better 3D/voxel generation than Argon, but ~2–4× slower; UI/frontend still weak, less “lazy” on task follow-through. Vals AI: finance/tax agents ~2.25× / 2.12× Argon’s time, finance accuracy slightly down; legal research #1 but ~52 minutes/task (~20 minutes more than Argon). AA-Omniscience: correctness 48%→58%, hallucination 4%→11% (wrong answers instead of abstaining)—a different lens from Google’s “fewer unsupported claims.” Developers also dislike announce-first, GA-later cadence.
Same-Day Earnings: Cloud Doubles, Cash Flow and Ads Weigh on the Stock
Alphabet Q3 revenue ~$127.58B (+25%); Google Cloud ~$30.71B (+103%), operating profit ~$11.12B at 36.2% margin. Gemini App ~1B users; first-party model API ~32B tokens/minute (4×+ YoY). Ads ~$82.01B (+11%) missed by ~$1.41B; AI Mode queries grow but monetize below classic Search. Capex ~$51.36B; FCF ~−$9.07B (second negative quarter; includes ~€5.2B EU fine cash and TPU stocking). 2026 capex guided $195–205B; third-party compute may pressure Cloud margins near term. Pichai said a distilled Carbon is in progress for cost/latency/quality at billion-user scale; Argon already lifted some internal teams’ code delivery +35% / bugs −30%, with Carbon rolling inside Google.
A #1 model proves the tech round; putting that model in a billion users’ hands is still a cost, latency, and monetization problem.
FAQ
Q: Where does Carbon beat Argon? A: Intelligence Index 50→59, plus coding and GraphWalks gains, longer agent runs, and stronger safety benches—with much higher latency and no broad GA yet.
Q: When can regular users access it? A: No public GA date; priority goes to cyber-defense orgs and Fairwind partners first.
Q: Did the #1 ranking help the stock? A: Alphabet rose in regular hours before earnings, then fell more than 10% after hours as investors focused on the ads miss and heavy capex driving negative FCF.
相关文章
