DeepSeek Updates V4.1 Flash for Faster Agent Work

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026. It is an update to its low-cost model, and DeepSeek says it finishes agent tasks faster and gives shorter answers. On DeepSeek's own scores, a test of work done in a computer's command line rose from 31.2% to 42.3%. You don't need to switch anything. Anyone using the "deepseek-flash" API name already runs the new version. DeepSeek also opened its web search and page-reading tools to developers at $0.75 per 1,000 calls, about $7.50 for 10,000 lookups. For DeepSeek API users, calls made from Claude Code, Codex or DeepSeek Harness are free until October 31 at 16:00 UTC. model card ↗

OpenCode Reports Breach Affecting 1.4% of Console Users

OpenCode said on October 17, 2026, that an overnight attack reached 1.4% of users of OpenCode Console, its hosted service. Everyone affected had linked an outside AI provider. OpenCode emailed those users and says anyone without an email was not affected. Coverage of that email says saved keys for services such as OpenAI and Anthropic may have been exposed. Those keys work like passwords for a paid AI account. If you got the email, cancel the keys at each provider and check your bill. OpenCode said the incident was still ongoing when it posted. OpenCode's post ↗

Anthropic Speeds Up Claude Code by Compressing Requests

Anthropic's ClaudeDevs account said on October 16, 2026, that Claude Code, its coding assistant, now compresses the conversation it sends on every turn so it reaches the model sooner. There is no new model. Replies just start sooner, on the same plan. Cloud sessions also start up while you are still typing. Anthropic's changelog has one caveat. Some company networks inspect traffic and can break compressed requests, and setting CLAUDE_CODE_GZIP_REQUEST_BODIES=0 turns the feature off. ClaudeDevs post ↗

Cursor's Cloud Agents Now Build Shareable Dashboards

Cursor, the AI coding app, said on October 16, 2026, that its cloud agents can now build interactive canvases, such as a dashboard or a report. You can share a finished canvas with your team as a link. So an agent's work can now be a page your coworkers open, not just changed code. Cursor also says large project setups load about five times faster. Its launch post and its changelog give different limits for how many code projects one setup holds: 500 and 300. Cursor changelog ↗

NVIDIA Releases Dory, a Small Model That Trades Speed for Accuracy

NVIDIA posted Dory1.0-4B on Hugging Face on October 17, 2026. With 4 billion parameters, it is small enough to run on a personal computer. One copy can run its core step up to four times. More passes take longer and fix more bugs. In NVIDIA's own table, one pass solves 50.3% of SWE-bench Verified, a test built from real GitHub bug reports, and four passes solve 61.8%. No one outside NVIDIA has rerun those numbers yet. Dory reads English only, and its OpenMDW license is worth checking before business use. model card ↗

GitHub Will Remove Three AI Models From Copilot on November 16

GitHub said on October 15, 2026, that it will remove Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra from most parts of Copilot, its coding assistant, on November 16. GitHub suggests Claude Opus 5.5, GPT-6.1 Sol or GPT-6 Luna instead. People on individual annual plans keep Opus 4.8. GitHub changelog ↗

Anthropic Says Claude Mythos 5 Designed Smaller Quantum Circuits

Anthropic said on October 16, 2026, that Claude Mythos 5 designed quantum circuits from plain-language goals, and that some beat the best published designs by ten times or more. Coverage of the post says 7 of 8 designs improved on earlier work. The circuits were checked by computer calculation. None has run on a quantum computer. Anthropic's post ↗

Cloudflare Shares a Security Checklist for Coding Assistants

Cloudflare's security-audit-skill is on GitHub's trending list this month. It gives a coding assistant a routine for checking code for security holes in several rounds. Each problem it finds is checked again separately, then saved in a format other tools can read. repo ↗