Quick answer
Gemini 4 Carbon is Google’s newest frontier AI model, and Google says it ranks first on Artificial Analysis’s Intelligence Index (version 5) while setting new records in cyber security and long, multi-step tasks. It is not open to everyone yet: it goes first to cyber security defenders, with wider access promised within a few weeks. WorldofAI expects it to launch at Gemini 4 Argon’s price of $2 per million input tokens and $10 per million output tokens, which would make it cheaper than Claude Opus 5.5. The same round-up covers a disappointing Fable 5.5 leak, a cut to Codex usage limits, Claude Code’s new dynamic workflows and the return of “ultrathink”.
Key takeaways
- Gemini 4 Carbon is Google’s new top model, aimed at cyber security, coding and long-running agent tasks
- Early access is for security defenders first; everyone else should get it in the next few weeks
- Reported pricing (if it matches Argon) is $2 / $10 per million tokens, well below Opus 5.5
- A leaked model called Obsidian 3, possibly Fable 5.5, beat Opus 5.5 by only 1.3 points on one coding benchmark
- OpenAI is reportedly halving Codex usage on the $200 Pro plan from 30 October
- Claude Code gets reusable dynamic workflows and the ultrathink keyword is back
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- What to Do When Your AI Plan’s Usage Limits Are Cut — measure your real usage, cut waste and keep your key automations running when a plan’s allowance drops
- How to Get an Accurate AI Transcript When Several People Are Talking — record, transcribe and check meeting or call transcripts so every speaker is labelled correctly
In this guide
- What is Gemini 4 Carbon?
- When can you use Gemini 4 Carbon, and what will it cost?
- How does Gemini 4 Carbon score on benchmarks?
- Is Gemini 4 Carbon better than Claude Opus 5.5 in real use?
- Is a Gemini 4 Ultra model coming next?
- What does the Fable 5.5 “Obsidian 3” leak show?
- What changed at OpenAI this week?
- What is new in Claude and Claude Code?
- What do Google’s latest results say about AI demand?
- What else happened in AI this week?
- What is still unconfirmed?
- Watch: the source video
- Frequently asked questions
- The bottom line
- Related posts
Google had a big week. It released a model it says is the most capable it has built, then posted quarterly results showing just how much AI is driving its growth. Meanwhile, leaks about Anthropic’s next model looked underwhelming, and OpenAI gave with one hand (free API credits) and took with the other (lower Codex limits).
WorldofAI’s latest video packs in more than 20 stories. Below we pull out what matters for people who use AI tools to run a business, keep the numbers as reported, and clearly flag what is a leak, a rumour or the creator’s own opinion.
What is Gemini 4 Carbon?
Gemini 4 Carbon is the successor to Gemini 4 Argon and, according to Google’s announcement as described in the video, its most powerful model so far. Google is pitching it on three strengths: cyber security, long-horizon tasks (jobs that need many connected steps over a long time) and reliability.
Google also says Carbon is better at asking for clarification instead of guessing, has stronger world knowledge and hallucinates less. It even claims 0% success for prompt injection attacks in its testing, where hidden instructions try to hijack the model. That is a bold claim, and the creator rightly says to wait for real-world use before believing it.
One welcome change: a common complaint about Argon was long, padded answers. Early reports suggest Carbon is more concise and better at sticking to the constraints you give it.
When can you use Gemini 4 Carbon, and what will it cost?
Not quite yet for most of us. Google is following the same approach it used for Argon: Carbon goes first to cyber security defenders through an early-access programme starting this week, then expands to all users over the next couple of weeks.
Google has not confirmed the price in the video. The creator expects it to match Argon’s launch pricing:
| Item | Expected price (per million tokens) |
|---|---|
| Input tokens | $2 |
| Output tokens | $10 |
| Cached input tokens | 95% off the input price |
If that holds, Carbon would be noticeably cheaper than Claude Opus 5.5 for similar work. Cached input is the part of a prompt that repeats between requests, such as long instructions; our guide How to Use Prompt Caching to Cut Your AI Agent Costs explains how to take advantage of it.
⚠️ Warning: Do not move client work or automations to Carbon on the expected price alone. Wait for Google’s published pricing and test it on your own tasks first.
How does Gemini 4 Carbon score on benchmarks?
Very well on paper. Here are the scores reported in the video:
| Benchmark | What it tests | Carbon’s reported score |
|---|---|---|
| SEC Bench Pro | Finding real security vulnerabilities | 88.2% |
| ExploitGym | Proving vulnerabilities by exploiting them | 54.8% |
| CWE Bench (Colinear) | Finding and fixing the root cause of security issues | 71% |
| Vending Bench | Running a simulated vending machine business over a long period | #1, the first new leader in over a year |
| Artificial Analysis Intelligence Index v5 | Broad overall intelligence | #1 |
Google’s internal testing also reportedly shows Carbon is more than twice as successful at writing secure code as Argon.
Vending Bench is the most interesting one for business owners. The model has to manage stock, order products, set prices and track money over a long stretch, making hundreds of connected decisions. Most models eventually lose the plot, so a strong score here suggests Carbon can stay on track in longer agent tasks.
💡 Tip: Benchmarks tell you what a model can do in a test, not what it will do for your business. Before switching, run the same three or four real tasks through your current model and the new one and compare the results and the cost.
Is Gemini 4 Carbon better than Claude Opus 5.5 in real use?
It looks close, and cheaper, but not better at everything. The creator points out that Argon also looked great on benchmarks and then felt weaker than Opus and GPT-6 Astra in daily use. With Carbon, his early previews look more convincing.
- 3D and visual work: a big step up from Argon, with detailed demos such as a mechanical flower and a coded luxury kitchen scene, though with small quirks
- Front-end design: strong design taste in a scrolling web page demo
- Web development: roughly level with Opus 5.5 but slightly behind
- Other coding: competitive with Opus on many real-world tasks
His verdict is that near-Opus results at a much lower price could make Carbon a sensible daily model. He would also like Google to add higher reasoning modes, so the model can spend longer on hard problems. That is his opinion, not a Google plan.
Is a Gemini 4 Ultra model coming next?
🔎 Unconfirmed / rumour: A Google DeepMind pre-training researcher replied to a discussion on X with a comment suggesting an even more capable, unnamed Gemini model is still to come. The creator guesses it could be a Gemini 4 Ultra or a heavily tuned Gemini 4 Pro. Google has announced nothing, and we also previously covered the Gemini 4 Hydrogen leak.
What does the Fable 5.5 “Obsidian 3” leak show?
If the leak is real, Anthropic’s next model may be only a small step up on coding. A model called Obsidian 3 appeared in Cognition’s FrontierCode 2.0 benchmark data, which tests AI on 132 realistic software engineering tasks. It reportedly runs through the Claude Code harness, which is why some people think it is Fable 5.5 being tested quietly.
| Model | High reasoning effort | Max reasoning effort |
|---|---|---|
| Obsidian 3 (reported) | 46.8% | 45.5% |
| Claude Opus 5.5 | 45.5% | Not stated in the video |
That is a gain of just 1.3 points at high effort and nothing extra at max effort. The reported cost is about $47 per question, around 3.6 times Opus 5.5 at the same effort level. For background on why effort levels matter when comparing models, see Claude Effort Levels Explained: How to Compare Models Fairly.
🔎 Unconfirmed / rumour: Nobody has confirmed that Obsidian 3 is Fable 5.5, and earlier Fable checkpoints looked much stronger (see our Fable 5.5 leaks round-up). One benchmark under one setup is not the full picture.
What changed at OpenAI this week?
- Text your ChatGPT agent: Pro users in the US can now message their personal “Dot” agent through iMessage or RCS with text, photos and voice notes, and it can message you back with updates
- Free API credits: eligible ChatGPT Pro subscribers can reportedly claim $100, $200 or $500 a month in API credits depending on tier, and Business workspaces up to $500 based on paid seats
- Lower Codex limits: from 30 October, OpenAI is reportedly cutting Codex usage on the $200 Pro plan from 20 times to 10 times the standard limit, at the same price, with temporary promotional credits to soften the change
- Coming soon: a combined ChatGPT Chat and Work experience, and an image model update reported as GPT Image 2.6
The Codex change matters if you run daily automations or coding work through it. Halving an allowance overnight can break a routine that used to fit comfortably.
📘 Main article: If your plan’s limits are changing, our guide What to Do When Your AI Plan’s Usage Limits Are Cut walks you through measuring your real usage, cutting waste and keeping your most important jobs running.
What is new in Claude and Claude Code?
Three updates stand out.
- Ultrathink is back in Claude Code: type the keyword to ask Claude for extra thinking on one difficult task, without changing your default effort level. Useful if you normally run Opus 5.5 on medium
- Dynamic workflows: Claude Code can now run repeatable multi-step jobs with several sub-agents, in order or in parallel. Each workflow is an editable JavaScript file, but you can simply ask Claude to write one, save it with the
/workflowcommand and rerun it by name as a slash command - Claude for Legal: a gallery of ready-made legal tools for contract review, regulatory monitoring and citation checks that you can customise and share across a team
Dynamic workflows are the most practical change for non-coders. Instead of re-explaining the same complicated job every time, you describe it once and rerun it the same way each time.
🔎 Unconfirmed / rumour: Members of a Chinese AI community claim they have reverse engineered a hidden points system Anthropic uses to decide which Claude accounts get banned, with factors such as changing IP addresses often or maxing out usage counting against you. Anthropic has not confirmed any of this.
What do Google’s latest results say about AI demand?
Demand is enormous. According to the figures covered in the video, Alphabet grew revenue 25% year on year to around a $500 billion annual run rate, Google Cloud revenue grew 103%, and Google’s own model APIs now process nearly 32 billion tokens a minute, up 40% on the previous quarter.
Google also said demand is outstripping its own infrastructure, so it is buying third-party computing capacity and raising planned spending from $195 billion to $205 billion. Search query growth was reportedly at its highest level in five years.
For small businesses, the takeaway is simple: AI providers are short of computing power. That is one reason usage limits keep changing, so it pays to have a backup plan for the tools you rely on.
What else happened in AI this week?
- NVIDIA Nemotron 3.5 Transcribe: an open-weight speech model that transcribes 40 languages, labels speakers even when they talk over each other, translates and redacts personal data. NVIDIA says it tops Hugging Face’s open English speech recognition leaderboard with a 5.33% average word error rate
- NVIDIA MotionBricks: generates realistic character movement at up to 15,000 frames per second, trained on over 350,000 motion clips, for games and robots
- Gemini notebooks connectors: Google Drive is reportedly the first connector, letting Gemini search your documents without changing them
- Gemini immersive exploration: interactive 3D lessons on topics like jet engines, currently in the Gemini Student Hub
- Touchscreen MacBook Pro: Apple’s first, with a tap, swipe and pinch display; Apple reports a 25% gain on AI-assisted tasks versus M5
- Distillation claims: a report says Moonshot AI and DeepSeek used GPT-6 Astra outputs to train Kimi K3.1 Pro and DeepSeek V4.1 Flash. These are allegations
- “SI” instead of “AI”: the US State Department has reportedly started renaming AI programmes to “SI” (superintelligence) following an executive order. This applies to government use, not to the public or AI companies
📘 Main article: Nemotron’s overlapping-speaker feature is a reminder that transcripts of meetings and calls are still often messy. Our guide How to Get an Accurate AI Transcript When Several People Are Talking shows how to get a clean, labelled transcript with the tools you already have.
Relying on one AI plan for your daily work? Take ten minutes with our usage-limits guide so a sudden change does not stop your automations.
What is still unconfirmed?
🔎 Unconfirmed / rumour: Treat these as unconfirmed: Carbon’s final pricing and its 0% prompt injection claim; a Gemini 4 Ultra; Obsidian 3 being Fable 5.5 and its cost per question; the Claude account points system; the distillation allegations; the Codex limit details until you see them in your own account; and GPT Image 2.6 timing. The creator’s comparisons with Opus are his own impressions from early previews. The video also includes a sponsored segment about selling print-on-demand wrapping paper, which we have not covered.
Watch: the source video
What this video covers: Gemini 4 Carbon’s launch, access, pricing and benchmarks; early demos; hints of a Gemini 4 Ultra; the Obsidian 3 / Fable 5.5 leak; the US “SI” renaming; distillation allegations; ChatGPT Dots texting, free API credits and Codex limit changes; Claude for Legal, the account-ban claims, ultrathink and dynamic workflows; Google’s Q3 results; NVIDIA Nemotron 3.5 Transcribe and MotionBricks; Gemini notebook connectors and immersive exploration; and Apple’s touchscreen MacBook Pro.
Video: WorldofAI (@intheworldofai). This summary and the opinions in it are our own. Benchmark scores, prices, leaks and reports are as described in the video unless we say otherwise.
Frequently asked questions
What is Gemini 4 Carbon?
It is Google’s newest frontier AI model and the successor to Gemini 4 Argon. Google says it ranks first on Artificial Analysis’s Intelligence Index v5 and sets new records in cyber security and long, multi-step tasks.
Can I use Gemini 4 Carbon now?
Most people cannot yet. It is going to cyber security defenders first, with access for all users expected over the next couple of weeks.
How much will Gemini 4 Carbon cost?
Google has not confirmed it in the video. WorldofAI expects Argon’s pricing: $2 per million input tokens, $10 per million output tokens and 95% off cached input.
Is Fable 5.5 out?
No. A model called Obsidian 3, which some people think is Fable 5.5, appeared in benchmark data. Anthropic has not announced Fable 5.5.
Are Codex usage limits going down?
According to the video, from 30 October OpenAI is cutting Codex usage on the $200 Pro plan from 20 times to 10 times the standard limit, with temporary promotional credits. Check your own account for the details.
What is ultrathink in Claude Code?
It is a keyword you add to a request to ask Claude for extra thinking effort on one hard task, without changing your usual effort setting.
The bottom line
Gemini 4 Carbon could give near-top performance at a much lower price, but wait for wider access and test it on your own work before switching. The bigger practical lesson this week is about dependence: limits and prices are moving because providers are short of computing power. Your next step: check how much of your AI plan you actually use, and set up a fallback for the jobs that matter most.
Related posts
- Genie 4, Claude’s Drug Discovery Work and the Gemini 4 Hydrogen Leak: This Week’s AI News
- Claude Fable 5.5 Leaks: Is It Really “Baby AGI”? Plus GLM 5.5 and GPT-6 Astra Pro
- GPT-6.1 Terra: OpenAI’s Cheaper Fast Model, Anthropic’s IPO Filing and This Week’s AI News
[…] guide supports our main article, Gemini 4 Carbon: Google’s New Frontier Model, the Fable 5.5 Leak and This Week’s AI News. In it, we covered NVIDIA’s new Nemotron 3.5 Transcribe model, which NVIDIA says can tell […]
[…] it looks cheaper than most of Anthropic’s models but more expensive than most of OpenAI’s. Our Gemini 4 Carbon article has more detail on early […]