Tech • AI • Robotics • Game
Gemini 4 Carbon Is the Best Model Ever?! Fable 5.5 Benchmarks, Gemini 4 Ultra, & Claude Ultrathink!
9.4/10TL;DR
Google has unveiled Gemini 4 Carbon, a new frontier model aimed first at cybersecurity users, while fresh leaks and policy moves across the AI sector point to intensifying competition, rising infrastructure demand and new regulatory friction.
KEY POINTS
Google launches Gemini 4 Carbon
Gemini 4 Carbon is positioned as Google’s newest top-tier model and is being introduced first through the Fairwind program for cybersecurity defenders before a broader rollout in the coming weeks. It is expected to be priced similarly to Gemini 4 Argon at about $2 per million input tokens and $10 per million output tokens, with heavily discounted cached inputs.
Benchmark claims put Carbon at the top
Google says Carbon ranks No. 1 on Artificial Analysis Intelligence Index v5.0 and sets state-of-the-art marks in cybersecurity and long-horizon tasks. Reported security scores include 88.2% on SecBench Pro, 54.8% on Exploit Gym, and 71% on CWE Bench, alongside internal testing that suggests it generates secure code at more than twice the success rate of Argon.
Focus on reliability and agentic performance
Carbon is presented as more cautious about asking clarifying questions instead of guessing, a direct attempt to reduce hallucinations. Google also claims 0% success on prompt-injection testing, while the model reportedly takes the top spot on the long-running Vending-Bench, which measures whether an AI system can manage a simulated business through many linked decisions over time.
Coding and creative output show mixed strengths
Early previews suggest Carbon is stronger than Argon in coding, 3D asset generation and design-heavy frontend work, including intricate animated scenes and detailed interiors generated through code. In web development, however, it appears roughly competitive with Claude Opus 5.5 rather than decisively ahead.
Hints emerge of an even larger Gemini release
Signals from Google DeepMind staff suggest Carbon may not be the company’s final move in this cycle. Speculation is building around a more powerful unreleased model, possibly a Gemini 4 Ultra-style system or a heavier post-trained successor.
Fable 5.5 leak points to only marginal gains
A mysterious model named Obsidian 3 surfaced in Cognition’s Frontier Code 2.0 benchmark and is being linked to Fable 5.5. The leaked result shows 46.8% at high reasoning effort versus 45.5% for Claude Opus 5.5, but no gain at maximum effort and an estimated cost of roughly $47 per question, about 3.6 times Opus at similar effort.
Trump administration pushes “super intelligence” label
President Donald Trump is again promoting “super intelligence” or SI in place of AI for federal terminology. A September 29 executive order reportedly directed agencies to replace the term in official communications, and departments have already begun renaming programs and export initiatives, even though the scientific meanings of artificial intelligence and superintelligence are not the same.
Chinese labs face distillation allegations
Moonshot AI and DeepSeek are facing allegations that they used outputs from OpenAI’s GPT-6 Astra to train newer systems. The report links Kimi K3.1 Pro and DeepSeek V4.1 Flash to this type of distillation, a technique that transfers capabilities from a stronger model into another system and has become more contentious as labs tighten safeguards.
OpenAI and Anthropic expand product offerings
OpenAI is rolling out messaging for personal AI agents through iMessage and RCS, allowing users to send text, photos and voice notes and receive proactive updates. It is also offering eligible ChatGPT Pro and business users up to $500 a month in API credits, though the $200 Pro plan’s Codex allowance is reportedly being cut from 20 times standard limits to 10 times.
Anthropic adds legal tools and restores ultrathink
Anthropic has added a library of ready-made artifacts for Claude for Legal, including workflows for contract review, monitoring and citation checks. It is also restoring Ultrathink inside Claude Code CLI for harder coding tasks, while reports from a Chinese AI community claim the company uses a hidden account scoring system that can raise ban risks based on behavior such as frequent IP changes or abrupt subscription activity.
AI spending boom lifts Google’s financials
In Alphabet’s Q3 results, annualized revenue reportedly reached $500 billion, with Google Cloud up 103% year over year. Google said its model APIs now process nearly 32 billion tokens per minute, up 40% from the prior quarter, and demand is so strong that the company is turning to third-party compute and increasing capital expenditure guidance from $195 billion to $205 billion.
Nvidia and Apple push new AI hardware and tools
Nvidia introduced Neotron 3.5 Transcribe, an open-weight speech model supporting 40 languages, speaker identification and translation, and said it reached a 5.33% word error rate on an English leaderboard. It also unveiled Motion Bricks, which can generate up to 15,000 frames of motion per second at 2 milliseconds latency on an RTX 5090. Apple, meanwhile, introduced its first touchscreen MacBook Pro, a shift that could support more direct interaction with AI-generated apps, creative tools and local models.
CONCLUSION
The latest product launches, benchmark leaks and policy changes show an AI market moving on several fronts at once: stronger models, rising compute costs and growing political influence over language and deployment. Gemini 4 Carbon is the clearest sign that Google intends to compete aggressively at the top end of the field.
Ask a question