DeepSeek V4.1 Flash, GPT‑6 Astra, and 10 more AI updates you should know about

AI just had another wild week. DeepSeek dropped a model so cheap and fast it’s putting serious pressure on GPT and Claude, while OpenAI claims its internal systems may have cracked a 90‑year‑old math problem. Apple finally gave Siri a real brain, Meta launched a true AI assistant that can act for you, and Google, Anthropic, and Perplexity all rolled out major upgrades.

DeepSeek V4.1 Flash: ultra‑cheap, fast, and surprisingly powerful

DeepSeek’s new model, V4.1 Flash, is designed to be both powerful and extremely affordable. It’s a massive 552‑billion‑parameter model, but it doesn’t fire up the whole thing for every request. Instead, it uses a technique where only a small part of the model is activated at a time—around 8 billion parameters to understand your input and 16 billion to generate its answer.

Think of it like a huge company where you only bring in the specialists you need for each task instead of the entire staff. That keeps it fast and dramatically cuts compute costs.

On coding benchmarks, V4.1 Flash is reported to beat models like GPT‑5.6 “Soul” and Claude Opus 5 in many tests, while running faster and cheaper. This continues the same aggressive strategy that has already shaken up pricing across the AI industry, as seen with earlier DeepSeek releases like V4 and its massive price cuts. For more context on that shift, you can check out how DeepSeek’s pricing is impacting AI valuations.

How DeepSeek V4.1 Flash pricing works

Because of its efficient architecture, V4.1 Flash can offer extremely low API pricing. During off‑peak hours, rates can drop to around $0.15 per million input tokens and $0.60 per million output tokens—far below many premium models on the market.

You can also download the model weights for free from Hugging Face, but you’ll still need to pay for the hardware to run it yourself. For many people, using it via API will be simpler and cheaper than managing your own infrastructure.

Hands‑on tests: what DeepSeek V4.1 Flash can actually do

Benchmarks are one thing; actual usage is another. In testing, V4.1 Flash handled a mix of writing, design, coding, and even 3D workflows.

Content and copywriting

Using a local workspace called DeepSeek Harness (similar to tools like Claude Code), V4.1 Flash was asked to draft a LinkedIn post explaining the model. The run took under a minute and produced a clear post plus an alternative hook. The interface also showed detailed run metrics: how many tokens were used, how long the model took to respond, and a full “trajectory” log of the reasoning steps and context.

Improving slide decks

Next, the model was given a basic presentation and asked to redesign it. It applied a new style and color theme, tightened the layout, and produced a more polished version of the same content. This test focused less on writing and more on how well the model could interpret structure and improve visual communication.

Building interactive websites and animations

The model was then used to help build a small product website—for an “iPhone Duo”—with scrolling animations. As you scroll, the design and motion work together, showing that the model can assist with front‑end structure, styling, and animation logic, not just plain HTML.

Coding a Rubik’s cube simulator

For a more complex coding challenge, V4.1 Flash helped create a Rubik’s cube simulator. You can:

• Scramble the cube and have it solve itself again
• Switch between 3×3, 4×4, 5×5, 6×6, and 7×7 cubes
• Use on‑screen controls to perform manual moves

There are still a few bugs that need fixing, but the core logic—scrambling, solving, and manual control—works. That’s a strong sign the model can handle multi‑file projects and non‑trivial algorithms, not just code snippets.

From a room photo to a 3D scene

Finally, V4.1 Flash was used in a workflow that turned a photo of a room into a 3D scene inside Blender. The result wasn’t just a 2D mockup—it produced actual 3D geometry. In wireframe view, you can see the mesh lines for objects; in solid view, you see the shapes and layout. This shows the model can help bridge the gap between images and 3D tools.

When you combine this kind of capability with the model’s low cost, you get a system that’s not just competitive on benchmarks but also very practical for real projects. For a deeper look at how DeepSeek’s latest models compare with OpenAI’s stack, see this breakdown of DeepSeek V4 vs GPT‑5.5.

How to use DeepSeek V4.1 Flash with DeepSeek Harness

To try the model in a developer‑friendly environment, you can use DeepSeek Harness, a local workspace that runs on your own machine and connects to models via API.

The basic setup flow is:

1. Open the DeepSeek Harness page and copy the quick‑start command.
2. Paste it into your terminal and run it to launch the interface locally.
3. Add your API key (for example, via OpenRouter) to access DeepSeek V4.1 Flash.
4. Add a workspace folder—this is simply the project directory you want the AI to work on.

Harness offers several modes:

• Standard mode: lets the AI edit files, run commands, search the web, plan tasks, and collaborate with other agents.
• PTC mode: handles tools differently and skips the workflow tool for a leaner setup.
• Minimal mode: keeps things simple with a persistent terminal session and fewer extras.
• Creator mode: adds tools for building custom agents and monitoring how they run.

Once configured, you can choose how much effort the model should spend on a task—low, high, or max—depending on whether you want a quick draft or a more thorough solution.

OpenAI’s 90‑year‑old math breakthrough with GPT‑6 Astra

OpenAI says its internal systems may have cracked a famous unsolved math problem called Navier–Stokes, which deals with how fluids like water, air, and blood flow. The problem is so important that it’s one of the Clay Millennium Prize Problems, each worth $1 million to solve.

According to OpenAI, they used around 10,000 AI agents powered by an internal model that is even more capable than GPT‑6 Astra. Over about 88 hours, one approach produced a potential proof. GPT‑6 Astra then converted that proof into a fully formal version that computers can verify step by step, which took another 17 hours.

The proof still needs to be checked by the broader math community. But if it holds, it would mark a major shift: AI moving from answering questions humans already know how to solve to tackling problems that have resisted human effort for decades.

Apple’s new Siri AI powered by Apple Intelligence

Apple is finally giving Siri the upgrade people have been asking for. The new Siri AI is powered by Apple Intelligence and is designed to feel more like a conversational assistant than a rigid voice interface.

Key changes include:

• Natural conversations: You can ask follow‑up questions and brainstorm ideas similar to how you’d use a modern chatbot.
• Personal context: Siri can understand what’s on your device and in your apps. You can ask it to find an old photo, surface a buried email, or pull details from a note you wrote months ago.
• In‑app actions: It can take actions inside apps like Messages, Music, and Reminders—such as sending a message, adding a reminder, or controlling playback.
• Dedicated app: Apple is launching a standalone Siri app so your conversations can continue seamlessly across iPhone and iPad.

The upgraded Siri AI will roll out in English later this year, bringing Apple much closer to the experience people have come to expect from tools like ChatGPT.

Meta Muse: an AI that actually acts on your behalf

While most chatbots only talk, Meta’s new AI agent, Muse, is built to take real actions for you—while keeping you in control.

Here’s what Muse can do:

• Handle school or work logistics: For example, it can read a school email, extract the supply list, find those items online, add them to your cart, and then ask for your approval before placing the order.
• Plan and register for events: Muse can look at your fitness data, find a local race, and ask if you want to sign up. Say “yes” and it fills in the registration form.
• Manage travel changes: If your flight is delayed, Muse can check other options in the background and ask if you want to switch. With your confirmation, it updates the booking.
• Everyday life management: Drafting replies, tracking goals, analyzing sleep, planning trips, sending gifts, and finding ways to save money.

Muse can also connect with apps like Instagram, Facebook, Messenger, Threads, and Shopify. The idea is to let AI do the legwork—searching, comparing, filling forms—while you still approve important decisions.

Investing in the AI “shovels”: chips, power, and data centers

When people think about investing in AI, they often focus on the companies building the models: OpenAI, Anthropic, Google, Meta, and others. But there’s another side to the boom: the infrastructure that makes AI possible.

In a gold rush, the people selling shovels often do as well as the miners. In AI, those “shovels” are:

• Chip designers and manufacturers
• Lithography machine makers
• Rare earth metal suppliers
• Power generators and grid equipment
• Data center builders and operators

Some investment platforms now offer themed baskets (or “smallcases”) that group together companies across three major AI themes: the physical AI supply chain, frontier model providers (US and Chinese), and the electricity and data center infrastructure behind it all. As always, any investment carries risk, so it’s important to research carefully before putting money in.

ChatGPT Images 2.5: faster, more consistent AI image editing

OpenAI has upgraded its image system with ChatGPT Images 2.5. The focus this time is on speed, realism, and consistency when you keep editing the same image.

Some examples of what you can do:

• Upload a photo of your fish tank and ask it to design a tiny home office inside it. Then say “make it more modern” or “add a bookshelf”—the model keeps the original layout and only updates what you asked for.
• Upload a photo of your cat and turn it into a tattoo design.
• Upload a selfie and try out new hairstyles before you commit in real life.

There’s also a new sketch feature. Type @sketch, draw a rough idea—like a basic candle holder—and ChatGPT can turn that into a polished design. OpenAI has added ready‑made templates for posters, merch, and more, making it easier to iterate from rough idea to final design in one place.

Anthropic’s scenarios: how AI could reshape the economy

Anthropic recently laid out three scenarios for how AI might affect the US economy by 2030. These are not predictions, but possible futures:

1. Internet‑level impact: AI boosts productivity in a way similar to the early internet, improving workflows but not completely transforming work.
2. AI does about half of knowledge work: AI systems handle roughly 50% of tasks in jobs like writing, coding, analysis, and research. The economy grows faster, but wages for many workers barely move because AI takes over a big chunk of what they do.
3. AI does most knowledge work: AI becomes better than humans at most knowledge work and performs nearly all of it. Anthropic estimates GDP growth could reach 15% per year—but many workers could face wage pressure or job losses.

The big open question: who actually benefits from that growth—workers, companies, or mainly the owners of AI infrastructure?

Misuse of Claude: model theft, weapons, and cyber operations

Anthropic also released a report on how its Claude models are being misused around the world. It highlighted several worrying patterns:

• Model scraping for training: Anthropic says some Chinese AI labs, including DeepSeek and Moonshot, used thousands of fake accounts to harvest Claude outputs and train their own models.
• Weapons development: A group in Yemen allegedly used Claude as a kind of virtual engineering team to help build guidance software for rockets and missiles.
• Cyber and surveillance work: In other cases, a single person or very small team used Claude to do work that would normally require a much larger group, including surveillance operations and cyber attacks.

Anthropic says it has taken steps to block these abuses, but the report suggests this is likely just the beginning of a broader challenge around AI misuse and security.

ChatGPT Work for Data: a built‑in data analyst for your company

OpenAI has launched “ChatGPT Work for Data,” a version of ChatGPT tuned for company analytics. Instead of digging through dashboards and spreadsheets, you can ask questions in plain language.

For example, you could say: “Traffic is high but sales are flat—what’s going on?” ChatGPT can:

• Look through your company data
• Compare metrics over time
• Explain what changed and why
• Generate a full report with charts, key findings, and recommended actions

It can also push that analysis directly into tools you already use. You can ask it to add slides to a Google Slides deck, create or update live dashboards, and refresh them on a schedule (for example, every Monday morning) and send summaries to Slack.

Instead of just showing numbers, ChatGPT can now analyze them, explain them, and help you act on them.

Gemini Live: deep research from your voice on mobile

Google has made its deep research features easier to use on phones with Gemini Live. You can now start a full research project just by talking.

For example, you might say: “I’m studying abroad in Vienna and want to understand its influence on classical music. Can you run a deep research report?” Gemini then:

• Starts the research in the background
• Lets you lock your phone or switch apps while it works
• Sends a notification when the report is ready

Once it’s done, you can open the report and keep asking follow‑up questions using your voice—like “Explain how Vienna’s political history shaped its music scene”—and Gemini will answer directly from the research it compiled.

ChatGPT for financial services: AI tuned for finance workflows

OpenAI also announced ChatGPT for financial services, a version of GPT‑6 Astra tailored for banks, asset managers, and other finance teams.

Here’s what it can do:

• Client prep: Before a meeting, you can ask it to research a company using sources like financial statements, earnings calls, market prices, analyst estimates, investor presentations, and news.
• Verified insights: Every key claim comes with a citation so you can click through and verify the exact source.
• Integrated outputs: It can turn analysis into PowerPoint slides using your firm’s template and work with Excel and Word, as well as platforms like PitchBook, Dalupa, and LSEG.

The goal is to make ChatGPT a single workspace for research, analysis, and creating client‑ready materials, without constantly jumping between tools.

Google Lyria 3.5: from idea to full AI‑generated song

Google has upgraded its AI music generator with Lyria 3.5. Creating a song now starts with a simple description.

You can:

• Choose whether you want AI vocals or just an instrumental track
• Pick from genres like K‑pop, Latin, hip‑hop, rock, and more
• Turn short clips into full‑length songs

Lyria 3.5 also includes templates for common use cases such as business promos, brand jingles, birthday songs, and even rap battles. It’s designed to help non‑musicians turn ideas into complete tracks quickly.

Perplexity Portable Computer: AI that runs locally on Windows

Perplexity has introduced a feature called Portable Computer, which lets its AI work directly on your Windows PC instead of sending everything to the cloud by default.

Once set up, Portable Computer can:

• Connect different AI tools and let them work together locally
• Run models directly on your machine, including Perplexity’s own pplx‑27B model
• Keep your files and data on your computer by default

If your PC isn’t powerful enough for a specific task, Perplexity will ask for permission to use a stronger cloud model. You can approve that task once, and only that job will go to the cloud. It supports more than 15 cloud models when needed.

You can also schedule tasks to run automatically, so AI can repeat recurring work for you in the background.

Why DeepSeek V4.1 Flash matters in the bigger AI picture

DeepSeek V4.1 Flash isn’t just another model with good benchmarks. Its combination of strong coding ability, solid multimodal support (text and images), high speed, and extremely low cost makes it a serious contender in the AI stack wars.

As more companies and developers look for ways to keep AI costs under control while still building ambitious products, models like V4.1 Flash could change which platforms they choose as their default. And when you pair it with higher‑end systems like GPT‑6 Astra in a hybrid setup, you can route tasks to the cheapest capable model and only use the most powerful (and expensive) options when truly necessary.

Between DeepSeek’s price pressure, OpenAI’s research breakthroughs, Apple and Meta’s new assistants, and Google and Anthropic’s steady advances, the AI landscape is moving fast. The upside is clear: more capable tools, lower prices, and smarter assistants. The challenge will be making sure the benefits—and the risks—are managed in a way that works for everyone, not just a handful of companies.