Today's thread
This issue comes down to one question: which model actually does your work, and where is the key that calls it? Grok Bot says it will pick other companies' models per task, Claude hands risky-looking requests to an older model, and Cursor has added a Meta model. The choice of model is moving from the user to the tool.
That also means handing more to the tool. The OpenCode Console breach showed what happens when a service holding keys from many providers is compromised. A cheap DeepSeek model and the Cursor plugin pstack widen your options, but checking which model answered and where your keys are stored is still on you.
OpenCode Console breach may have exposed API keys of 1.4% of users
On October 17, the AI coding tool OpenCode said 1.4% of users of its online service, OpenCode Console, were affected by a breach. Keys from outside providers such as OpenAI and Anthropic saved in their workspaces may have been exposed. Whoever holds a key can run up charges on your account. OpenCode emailed those affected and said users who got no email were not impacted.
OpenCode attributed it to unauthorized access to a legacy system it was preparing to shut down, and tied it to a large-scale distillation attack, meaning mass queries to someone else's model to train your own on its answers. Anthropic revoked the Claude API keys it believes were affected and told users who stored keys in OpenCode to create new ones. OpenCode said it is also working with OpenAI and Google.
Using several AI tools often means pasting company keys into each one. If a service holding your key is breached, changing your password does not stop the key. You revoke it in the dashboard of the provider that issued it, and using a separate key per tool lets you cut off only the compromised one.
How attackers got in, how long they stayed and how many keys were actually used have not been disclosed. OpenCode said the incident is ongoing. Since the damage is not settled, if you ever stored keys there, review your usage history whether or not you got an email.
Claude Opus 5.5 hands offensive security requests to Opus 4.8
Per Anthropic's help center, Opus 5 and Opus 5.5 run safety classifiers, separate checkers that judge whether a request is risky, on every request. Offensive security requests such as exploit writing or penetration testing fall back to Opus 4.8, and dual-use biology requests such as virology or toxicology fall back from Opus 5.5 to Opus 5. Content the model reads, like memory, web results and files, is checked too.
When it switches, a notice appears and the reply is labeled with the model that answered. The conversation then stays on the less capable model until it ends, though you can switch back. If the original request is still in the thread it may trigger again, so edit that message before retrying. Defensive work such as secure coding stays on Opus 5.5.
You can turn automatic switching off in the Claude apps under Settings > Capabilities ('Switch models when a message is flagged') and in Claude Code under Config > MODEL & OUTPUT. With it off, a flagged request pauses and you choose to edit and retry or send it to a lower model. On the API it is off by default. If quality suddenly drops, check the model name on the reply first.
There is a Cyber Verification Program for defensive security work, but per the article Opus 5.5 is not yet included. The exact criteria for what trips the classifiers are not public.
Grok Bot to route tasks to outside models, including Claude Opus 5.5
In an October 7 post on X, Elon Musk said Grok Bot will use the best back-end model for any given task. A back-end model is the AI behind the app that actually writes the answer. He named Claude Opus 5.5, MidJourney, Suno and 'other leading APIs'.
Until now, AI apps have defaulted to their own company's models. What changed is a Grok product saying openly that it will mix in rivals' models. Cursor's pricing docs also say Grok Bot shares a usage pool with the Cursor editor and CLI, so the two products sit under one billing structure.
For users, one app could hand writing, images and music to whichever model does each best. The trade-off is that it gets harder to know which company receives your conversation. Teams putting company material into Grok Bot should set rules on what goes in, assuming it may reach an outside model.
The announcement is a single post, with no official product documentation yet. It did not say which tasks go to which model, whether users can see or choose, or what is shared with outside companies. When MidJourney and Suno will actually be used is also unconfirmed.
Meta's Muse Wonder lands in Cursor at $4 per million input tokens
Muse Wonder, released by Meta Superintelligence Labs on October 12, is now in Cursor. Cursor's docs call it Meta's flagship and the successor to Muse Spark 1.3. Pricing per million tokens is $4 input, $20 output and $0.20 for cache reads, which reuse text already read. Tokens are the word-fragment units AI counts when reading and writing.
It reads 300k tokens at once by default and up to 1M in Max Mode, with no price increase for long prompts. Thinking depth has six levels from minimal to max, defaulting to high. Cursor says it scores 46.1% at default on its own coding test, CursorBench, the highest of any Muse model. It reads text and images, including screenshots.
Cursor's pricing docs say the usage pool is shared by the in-editor agent, the CLI (Cursor in the terminal) and cloud agents. An agent is an AI that chains several steps on its own. Using Muse Wonder in the CLI draws from the same pool, and on legacy plans from the 'Other Models' share, so check what is left first.
If team or enterprise admins restrict model access, they must enable the Meta provider before members see it. CursorBench is Cursor's own test, so scores are hard to compare with other companies' tests. The larger sibling, Muse Wonder Full, is reportedly not sold via API.
DeepSeek V4.1-Flash-1017 raises scores at unchanged prices
On October 17, DeepSeek upgraded its Flash model to V4.1-Flash-1017. The API name is still deepseek-flash, and pricing and calling method are unchanged. DeepSeek reports 77.0 on the coding test DeepSWE v1.1 and 42.3 on the terminal-task test Terminal-Bench 4.0. The model files are on Hugging Face under the MIT license.
DeepSeek says scores rose over the previous V4.1 while task time and output length fell. Prices per million tokens are $0.30 input and $1.20 output, halved outside peak hours (weekdays 01:00-04:00 and 06:00-10:00 UTC). The same day it opened its in-house search engine as Search and Fetch APIs at $0.75 per 1,000 successful requests.
Teams that put off AI automation over cost now have a reason to redo the math. The open model files can run on your own servers, though their size calls for high-end hardware. Either way, run it side by side with your current model on the same material and check quality yourself.
All scores are DeepSeek's own, with no independent results yet. Search calls made when a DeepSeek model is connected to Claude Code or Codex and uses its web search tool are not billed for now, but per-request billing starts at 00:00 Beijing time on November 1. Data going to a Chinese company's servers should be weighed against your security policy.
Cursor's official pstack plugin picks a playbook via /poteto-mode
pstack is listed in Cursor's official plugin collection. A plugin is an add-on bundle for an AI coding tool, and pstack is a set of skills meant to make AI write less but better code. A skill is a work manual the AI follows for a specific job. It is MIT-licensed and made by poteto, who has worked at Cursor and is on the React core team.
The core command is /poteto-mode. Describe the result you want and it picks one of 23 playbooks, such as bug fix, feature, performance or refactoring, copies the steps into a task list and works through them. /setup-pstack sets thinking depth and which model handles each role; the default pairing is Opus 5.5 and Grok.
It also works in Claude Code: enter /plugin marketplace add cursor/plugins, then /plugin install pstack@cursor-plugins, and the command becomes /pstack:poteto-mode. Instead of writing long instructions each time, you call a set procedure, which helps a team align how it hands work to AI.
There is no official figure for how much the plugin improves results. Its readme presents /poteto-mode as the default entry point for non-trivial tasks, so try it first on something like a bug whose cause needs finding rather than a one-line edit.
Try today
If you have used OpenCode Console, revoke the keys you stored there in the OpenAI, Anthropic and Google dashboards and create new ones. Check recent usage and billing for anything unfamiliar.
If Claude's output suddenly seems off, check the model name on the reply first. If your team does security testing, decide whether to keep automatic switching on or off.
Pick one recurring AI task and run the same material through DeepSeek V4.1-Flash-1017 and your current model, comparing results and cost. Test only with material that is allowed to leave your company.
If you use Cursor or Claude Code, install pstack and hand one small bug whose cause needs finding to /poteto-mode.