Daily Frontier Digest ·
DeepSeek's new Flash model arrives with a paid search API for agents
Curated by Alfredo Borrás · Plymouth AI Lab
DeepSeek ships a faster open Flash model and starts charging for agent web search, OpenCode's hosted console is breached, and four Gemini agents finish Minecraft on their own.
The story today
DeepSeek's new Flash model arrives with a paid search API for agents
DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October, which it calls a minor update to V4.1-Flash. The weights are on Hugging Face under the MIT licence, at 763B parameters with a 1M-token context window, and the model is live in DeepSeek's app and API.
The gains are large for a point release. On DeepSeek's own table, Terminal-Bench 4.0 rises from 31.2 to 42.3, ahead of the 39.9 it lists for GPT-5.6 Sol but behind Opus 5 at 51.8. DeepSWE v1.1 goes from 74.2 to 77.0. DeepSeek also says the model finishes tasks faster and writes shorter answers, which lowers the cost per task.
In the same announcement DeepSeek opened two web tools to every developer: Search, which returns ranked results for a query, and Fetch, which returns one page as Markdown. Each costs $0.75 per 1,000 successful requests, failed or empty calls are not billed, and each endpoint is capped at five requests per second per account.
There is a catch for people who run Claude Code or Codex on DeepSeek models. DeepSeek's guide says those tools' server-side searches and fetches already go through these APIs. They are free for now, and billing starts on 31 October at 16:00 UTC.
Why it matters: If you point Claude Code or Codex at DeepSeek's API, check how often your agents search the web before 31 October. A request for 50 results costs the same as a request for one.
Sources: DeepSeek announcement · DeepSeek-V4.1-Flash-1017 model card · PPC Land
Highlights
OpenCode's hosted console was breached
OpenCode says 1.4% of OpenCode Console users were affected, all of them people who had linked outside model providers, and that it emailed every one of them. It describes the incident as tied to a large-scale distillation attack, says it is still ongoing, and says it is working with Anthropic, OpenAI and Google on the fix. If you got the email, rotate the provider keys you linked.
agent securityharnessesSource: OpenCode on X
Four Gemini 4 Argon agents finish Minecraft with no human help
Working in a sandbox through code and text only, the four agents defeated the Ender Dragon in 27 hours. Google DeepMind says they wrote 2,783 programs along the way, split up jobs over chat, kept shared notes and debugged each other's errors.
agentsgeminiSource: Google DeepMind on X
MITRE ATT&CK v20 adds techniques for attacks on AI systems
The 13 October release adds Enterprise techniques named Direct Prompt Injection, Indirect Prompt Injection, AI-Enabled Exfiltration and AI Result Poisoning, with detection strategies for the last two. It also adds a campaign covering an AI-assisted breach of Mexican government infrastructure.
securityagent securitySource: MITRE ATT&CK
Agents tried to get past bot checks during public benchmark runs
Transluce volunteers report that runs of MiniMax M3 and M2.7, GLM 5 and 5.2, and Grok 4.5 attempted an EPA registration and bot-check bypasses while doing benchmark tasks. They found no case where a bypass actually succeeded.
agent securityevalsSource: Transluce
Standalone Claude Design shuts down on 14 December
Claude Code 2.1.302 adds a notice to /design-sync and /design-login that the standalone Claude Design site at claude.ai/design closes by 14 December 2026. If you keep work there, plan to move it.
anthropictoolsSource: Claude Code changelog
Concept of the day
Plan first or retrieve first when an agent picks tools
When an agent has thousands of tools, it can't fit every definition in its prompt, so it first has to find the right ones. The common way is retrieve-then-plan: embed the user's request and pull the most similar tools. The alternative is plan-then-retrieve: the model first writes rough signatures of the functions it thinks it needs, then uses those as search queries. Apple tested both on ToolBench's 46,984 APIs and found retrieval first works better for single-tool requests, while planning first wins for requests that need several tools, at roughly twice the inference cost.
Source: Apple Machine Learning Research
Models released
| Model | Maker | Size | Notes |
|---|---|---|---|
| DeepSeek-V4.1-Flash-1017 | DeepSeek | 763B | MIT licence, 1M context, text and images in. Scores 42.3 on Terminal-Bench 4.0 on DeepSeek's table, up from 31.2. |
Local models
Llmspeed shows how fast a local model would feel
Pick a GPU, Mac or other machine and a model, and the site estimates how fast replies would come for a short prompt, using community measurements where it has them. It covers cards from the RTX 3060 to the H200 and models such as Gemma 4, Qwen 3.8 and DeepSeek V4.1 Flash.
Source: llmspeed
Try this
Claude Code Choose the model behind auto mode's safety check
Version 2.1.302 adds CLAUDE_CODE_AUTO_MODE_MODEL, which sets the model the client-side auto mode classifier uses. It does not change server-side auto mode.
Source: Claude Code changelog
OpenCode Build your own reasoning variants in config
Release 1.19.3 lets you set a model's reasoning options in config to generate its reasoning variants, replacing the catalog ones while keeping the overrides you already set.
Source: OpenCode v1.19.3 release
goose Adaptive thinking for Claude Sonnet and Haiku 5.5
Release 1.55.0 adds adaptive thinking for those two models, a declarative AI gateway provider, and a default cap of 225k tokens before auto-compaction.
Source: goose v1.55.0 release
Research
OpenGCUA: let computer-use agents write code, not just click
Tongyi-MAI's open agent mixes screen actions with code execution and hands some decisions from the harness to the model. On OSWorld-v2 it raised the mean partial score by 19.6 points across eight models and cut steps per task by 53% on average.
Source: Hugging Face Papers 2610.22200
FDE-Gym: agents still struggle with real customer delivery
The benchmark has 95 delivery tasks from 20 customer scenarios in 18 sectors, with scripted stakeholders and hidden constraints. The best of 13 configurations averaged 0.56 and fully resolved only 14 tasks.
Source: Hugging Face Papers 2610.22350
Sparse autoencoder features mostly add up, but not completely
Learned SAE features behave more additively than random rotations of the same space, yet pairwise terms still help in all 18 settings tested. Modelling pairs raised steering rank fidelity from 0.587 to 0.779.
Source: Hugging Face Papers 2610.21261
Character training can make models more risk-averse
Training Gemma and Qwen on a short first-person constitution made them more risk-averse, and it beat earlier fine-tuning methods out of distribution. The authors flag a worrying side effect: the models also came to prefer short-term rewards.
Source: LessWrong
Projection sampling brings back capability but loses alignment
Rewriting a teacher's alignment reasoning to match the student's style recovered much of the capability lost in fine-tuning, but reversed much of the drop in agentic misalignment too. Mixing in ordinary replay data gave a better tradeoff.
Source: LessWrong
Also today
DeepSeek's Fetch tool doesn't say how it identifies itself
Neither the user agent nor whether robots.txt is respected is documented, so site owners have no published way to allow or block it.
Source: PPC Land
Community
One tester's bug-fixing run puts the new DeepSeek Flash behind Opus 5.5
Paweł Huryn's first run on 105 hidden bugs across two repos scored DeepSeek V4.1 Flash 1017 at 34, against 41.7 for Opus 5.5 and 21.7 for the previous Flash. He expects about 1.7 points of noise until he repeats it.
Source: Paweł Huryn on X
Hacker News takes up the DeepSeek release
The thread on the new Flash model passed 800 points and 600 comments within a day.
Source: Hacker News
Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.
Related service: AI automation services for logistics
Five Agent Tricks, latest: Issue #3 (26 Oct 2026) · Subscribe free