· 07:00 · 7 min + 2 min Explained
DeepSeek updates its open Flash model and cuts the cost per task
DeepSeek's new open model scores higher on coding tests and uses fewer words to get there. Also: Claude explains a 25-year-old crypto mystery, standalone Claude Design closes, and Replit picks OpenAI.
Written by AI and checked against our editorial standards. Every claim links to its primary source. Not reviewed by a human editor.
Get it in your inbox at 7am every day ↓
1. DeepSeek’s updated Flash model scores higher and costs less per task
DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October. It is an update to its Flash model, and it already runs in DeepSeek’s app, website and API.
Anyone can download the weights from Hugging Face. They use the MIT license, one of the most permissive open-source licenses. The model reads up to 1 million tokens at once.
DeepSeek says the model finishes tasks faster and writes shorter answers. Shorter answers mean fewer tokens, and you pay per token.
In DeepSeek’s own run of the DeepSWE coding test, it scores 77.0, up from 74.2 for the previous version. The Decoder reads DeepSeek’s chart as just under $0.20 per task, down from about $0.50.
On the same chart, rivals cost $2 or more per task. Claude Opus 5 scores 74.0 on that test.
Read the whole table, though. Opus 5 still scores higher on most of its tests. On Terminal-Bench 4.0, it scores 51.8 against 42.3. Nobody has tested the new model independently yet.
DeepSeek also opened its own web search to developers. The Search and Fetch APIs cost $0.75 per 1,000 calls each.
Search and Fetch calls made through Claude Code, Codex or DeepSeek’s own tool are free until 31 October at 16:00 UTC.
Running it yourself takes serious hardware. The Hugging Face page lists 763 billion parameters.
So what: If you pay for many coding-agent runs, test this model on your own tasks through DeepSeek’s API. Compare the cost per finished task, not the price per token.
Source: api-docs.deepseek.com · Official · huggingface.co · Official · the-decoder.com · Reported
2. Claude explains where the numbers behind widely used encryption curves came from
Anthropic said on 16 October that Claude solved a 25-year-old puzzle in cryptography. It found where the starting numbers for most of the NIST curves came from.
These curves belong to ECDSA, the most widely used standard of its kind. P-256 is the best known.
The NSA supplied the starting numbers, called seeds, in the late 1990s. It never explained them, which fed fears of a hidden backdoor.
Claude found that the seeds came from an ordinary C random-number function. Unix timestamps from 1997 and 1998 set it going.
It took under half an hour. P-256’s seed was the 1,842nd number the method produced.
Anthropic says this is not a break and not a weakness. It makes a backdoor less likely. The $36,864 bounty will go to the nonprofit behind Let’s Encrypt.
So what: Change nothing in your systems. Note it as a sign that AI can now help settle old questions in security research.
Source: anthropic.com · Official · coindesk.com · Reported
3. Claude designed quantum circuits that need far fewer steps
Anthropic also said on 16 October that Claude beat the best published designs for eight quantum circuits. The designs needed fewer qubits and fewer gates (the basic steps).
One circuit traded 12% more qubits for 17 times fewer gates. These are designs for future machines, not runs on real hardware.
So what: Nothing to do today. Expect AI to speed up research where a computer can check each answer automatically.
Source: anthropic.com · Official
4. Standalone Claude Design closes on 14 December, so save your projects
Anthropic will close the standalone Claude Design site on 14 December. After that, claude.ai/design will open Claude itself.
Design systems can move into Claude now as artifacts. Projects can’t move yet. Chats and comments will be deleted in December, and public project links will stop working.
So what: If you use Claude Design, download each project you need as a .zip. Use Share, then Project HTML, and keep “Project archive” selected. Do it before 14 December.
Source: support.claude.com · Official
5. Replit makes OpenAI the default model for its coding agent
Replit’s CEO, Amjad Masad, said onstage at TechCrunch Disrupt that OpenAI now powers Replit Agent by default. Replit is a popular platform for building apps by chatting with an AI.
Masad said Replit was until recently one of Anthropic’s five largest customers. He named cost, the partnership and performance as reasons. “I think ChatGPT is better now,” he said.
Claude is still available on Replit. But most users keep the default, so the change matters. One unnamed insider told TechCrunch the Anthropic partnership is still in place.
So what: If you build on Replit, check which model your agent now uses. Compare results on a real project before you accept the new default.
Source: techcrunch.com · Reported
6. US agencies publish a playbook for when an AI agent goes rogue
NIST, the NSA and CISA published joint guidance on 16 October for AI incidents during training and testing. An incident here means an AI agent takes a big action nobody asked for, against the operator’s rules.
The guide is voluntary. It sets out ten practices in five stages: prepare, detect, respond, recover and report.
It is written for the largest AI labs. The US government uses the term “super intelligence” for such advanced AI.
So what: If you run agents with real access, write down now who stops them and how. Use the five stages as a checklist.
Source: nist.gov · Official
7. Meta drops wording that let it train AI on business websites
Meta has removed a clause from its updated business terms, TechCrunch reported on 17 October. It let Meta use content from the websites of firms with Meta business accounts to train its products.
Legal experts had warned it could cover publishers’ sites. Meta says the terms only applied to its ads and business products.
So what: If your company has a Meta business account, read the updated terms and check what they allow Meta to do with your site.
Source: techcrunch.com · Reported
8. Codex on Mac can quietly fill your disk with crash files
Users of OpenAI’s Codex desktop app on Apple Silicon Macs report a bug that keeps writing crash files. One user found about 323 GB this week, spread over two folders. The bug was first reported in June, and OpenAI has not replied in the thread.
So what: If you use Codex on a Mac, quit the app and check
~/Library/Application Support/Codex/Crashpad/pending. Also check~/Library/Application Support/com.openai.codex/web/Crashpad/pending, and empty both.
Source: github.com · Reported
Open source & local AI
llama.cpp can now find and download models for you
The llama.cpp project added a Discover Models view to its built-in web app on 17 October. It is the free engine behind many tools that run AI models on your own computer.
Until now, you had to find a model file on Hugging Face yourself. The new view lets you search Hugging Face from the app and filter by size. It shows which versions fit in your computer’s memory.
Downloads can be paused and resumed. The app also fetches a small helper model that speeds up answers. That makes plain llama.cpp closer to easier apps like LM Studio and Ollama.
The feature is off by default. You turn it on with the setting “Enable Discover Models”.
So what: If you run llama.cpp’s server, update it and turn the setting on. Pick a version it marks as fitting your memory, and keep your data on your own machine.
Source: github.com · Official
Paper of the day
A small model that asks for help when unsure matches a top model for less
A paper posted on arXiv on 15 October shows how to make computer-use agents cheaper. These agents click and type on a computer to finish tasks. The method is called LEASH.
Running a top model at every step is expensive. Small open models are cheap, but they often fail alone. LEASH lets a small model do the work and pass control to a strong model only when it is unsure.
The trick is how it measures doubt. It compares the small model’s inner state with its earlier wins and failures.
If the current state looks like a past failure, it asks for help. It reuses work the model already does, so it needs no extra pass.
On the OSWorld test, a cheaper Qwen model with LEASH solved 73.7% of tasks. Using Claude Opus 5 at every step solved 72.9%. LEASH did it for 19% less money.
So what: If you run long agent tasks, try a cheap model first and hand off only the hard steps. Measure where it fails before you pay for a top model everywhere.
Source: arxiv.org · Official
Explained: if you have a little more time
3 terms from today's news · 2 minWhat is a benchmark?
A benchmark is a fixed test for AI models. It is a set of tasks, such as coding problems or exam questions, plus a rule for scoring each answer. Because every model takes the same test, you can line up the scores. That is why AI news is full of names like SWE-bench, DeepSWE, Terminal-Bench or HLE. Each one tests a different skill, from fixing real code to answering hard expert questions. > **Think of it like this:** a driving test. Everyone drives the same route, and the examiner uses the same checklist, so the results are comparable. Where the comparison fails: a driving test has an independent examiner. With AI, the company that built the model often runs the test and reports the score. A few things limit what a score tells you. First, who ran it. A lab's own numbers are claims until someone else repeats them. Second, the setup. The same model can score differently with a different tool, prompt or time limit. Third, leaks. Once a test is public, its questions can end up in the data a model learns from. The model may then have seen the answers before. Hugging Face's evaluation guide warns about this directly. Last, a benchmark only measures what it tests. A model can top a coding test and still be weak at the work you need. **Why it matters to you:** use benchmark scores to make a short list, not to decide. Check who ran the test, then try the top few models on your own tasks.Two more from today
What does "open-weight" mean?What you get when a model's weights are public, and what you don't.
An AI model is a huge file of numbers called weights. “Open-weight” means anyone can download that file and run the model on their own machines, instead of renting access through an API.
It is not the same as open source: the training data and code usually stay secret, and the license can set limits.
What is a context window?How much a model can read at once, and why it runs out.
The context window is how much text a model can look at in one go: your question, the documents you paste in and its own answer. It is measured in tokens, roughly three quarters of a word each.
Anything outside the window the model simply cannot see. That is why long chats get summarised before they run out of room.
How sure are we?Official: from the company, regulator or authorsReported: from a named outletUnconfirmed: rumor or leak