DeepSeek ships 552B multimodal model with 8x smaller cache and lower API prices
DeepSeek just dropped V4.1-Flash, and it is quietly replacing everything in their lineup.
Here is the setup: their previous Flash model was decent, but V4-Pro was the “smart” option you paid more for. Now V4.1-Flash beats V4-Pro on speed, cost, and performance. So they are retiring V4-Pro entirely.
The architecture is the interesting part. It packs 552 billion parameters total, but only activates 8B while reading your input and 16B while writing output. Think of it like a huge library where you only pull the books you actually need. More knowledge, less energy spent.
The biggest win for anyone building AI agents: the memory cache (the scratch pad the model keeps between calls) now needs 4x less RAM and 8x less storage. That directly cuts your API costs.
- Sees images natively, no extra setup
- #1 on Terminal-Bench 2.1, ahead of Claude Opus 5 and GPT-5.6
- Off-peak pricing is 50% cheaper, schedule batch jobs accordingly
To use it: set your model to deepseek-flash. Old model names still route there automatically.
Anthropic publishes its most detailed report on Claude misuse and how it stopped them
Anthropic just dropped its most detailed look yet at how people tried to weaponize Claude, and every single operation got caught and shut down.
So what were people actually trying to do? Seven harm areas: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, weapons development, and model theft.
Here are the standout findings:
- Attackers used Claude not just as an advisor but to execute attacks directly. A suspected Chinese state group used it to infiltrate roughly thirty global targets.
- AI has removed the skill advantage that once set state-sponsored hackers apart from lone criminals.
- In one case, Claude scored social media posts by political sensitivity and flagged people for government control.
- Anthropic tightened controls on newer models, restricting a broader range of dual-use biological queries.
Every operation was disrupted, safeguards were strengthened, and intelligence was shared with authorities. Protect your API keys. Seriously.
OpenAI ships ChatGPT for financial services with GPT-6 and live market data
OpenAI just launched ChatGPT for Financial Services, a dedicated workspace powered by GPT-6 Astra, built with Morgan Stanley and Evercore.
The big deal? No more stitching together data sources. You can start working with these datasets immediately, with no separate contracts to negotiate or connectors to set up.
Here’s what’s included out of the box:
- Built-in data from Daloopa, PitchBook, LSEG News, and Crunchbase covering earnings transcripts, financial statements, and private company info
- Editable research notes, financial models, and pitchbooks using your firm’s own Excel, Word, and PowerPoint templates
- Granular citations that let you trace every figure back to the exact paragraph or table it came from
GPT-6 Astra has been tuned specifically for retrieval, financial reasoning, and artifact generation. That means the AI actually understands financial documents, not just text.
Eligible financial institutions can contact OpenAI sales or their account team to get access. OpenAI has not published pricing.
OpenAI made a ChatGPT specifically for finance
ChatGPT for Financial Services is a version of ChatGPT Work built around GPT-6 Astra, with financial data from Daloopa, PitchBook, LSEG News and Crunchbase already baked in. It’s built for research, financial models, pitchbooks and client materials.
Morgan Stanley and Evercore helped shape it, and OpenAI has integrations with S&P Capital IQ, MSCI, Moody’s and others so firms can also bring in data they already pay for.
The useful thing is Astra can dig through financial statements, tables and notes, do the analysis, then turn it into an Excel model, research note, deck or chart with the sources still there to check.
More OpenAI stuff:
- OpenAI also launched GPT-Live-1 in the API. It can listen and talk at the same time while models like Astra or Codex do the actual work behind it. Speak says it saw almost 80% fewer interruptions, and it costs $0.05/minute before the backend model costs. So it roughly costs $72/day to run it 24/7 (much cheaper than i thought).
- And apparently OpenAI is thinking about slowing down frontier model development. Sam Altman reportedly brought up pacing development in an all-hands this week, and OpenAI has even asked Congress whether coordinating an AI slowdown could run into antitrust laws. The hope, obviously, is that everyone else slows down too.
Anthropic says it blocked possible efforts to build biological weapons
Anthropic published its latest Threat Report covering misuse it shut down over the last eight months, including possible bioweapon research, espionage, surveillance and weapons development.
One case involved a scientist asking Claude for help with gain-of-function research on the chikungunya virus, including mutations that could make it more harmful. Anthropic says the work was apparently going to happen at a military research institute, which was enough for them to shut it down.
It also named seven Chinese AI labs, including Alibaba, DeepSeek, Moonshot and Xiaomi, that tried to distill Claude using thousands of fake accounts. Moonshot and DeepSeek were quietly relaying customer prompts to Claude through fraudulent accounts. Users thought they were using Chinese AI, but were actually getting answers from Claude.
Other cases include someone in Yemen using Claude Code to build rocket guidance software, a system designed to monitor 25M phone lines in Mali, 4,700 dating-app personas and malware rewritten to dodge antivirus software.
And this is with today’s models.