Last checked October 31, 2026 · News roundup
Today in brief: Anthropic renamed its $200 Claude plan to Ultra and dropped the five-hour usage cap for verified accounts, a day after OpenAI scrapped its planned cut to ChatGPT Pro 200. A fake Claude Code update reached up to 5,000 Cursor users with an SSH backdoor. On the business side, fan-fiction site AO3 could stay offline for up to four weeks, and Meta's new business terms are on hold while the UK opens a competition case.
Part 1 · AI plans, security and policy
Claude Ultra plan: what $200 a month gets you now that Max 20x is gone
On October 30, Anthropic renamed its two top consumer plans. Max 5x is now simply Max at $100 a month, and Max 20x is now Ultra at $200 a month. Prices did not change. Claude's pricing page now says Ultra has "no 5-hour limit for verified accounts" and gives "10x Pro usage" every week, plus $200 a month in API credits, "up from $100 on Max."
The rename also clears up a long-running confusion. The old "20x" figure described the five-hour session window; the weekly allowance on that plan sits at 10 times Pro, which is what the Ultra listing now shows. According to Anthropic's announcement on X, as relayed by AI Weekly, most existing subscribers get the change without doing anything, while new Ultra accounts keep the five-hour cap until verification clears. Anthropic hasn't published what "verified" requires.
What to do: if you hit the five-hour wall on Max but rarely use up your weekly pool, Ultra is the plan that removes that friction. If the weekly ceiling is your problem, Ultra adds no more weekly usage than the old 20x plan did.
Claude individual plans after the October 30 rename
| Plan | Monthly price | 5-hour window | Weekly usage | API credits |
|---|---|---|---|---|
| Pro | $20 ($17/mo billed yearly) | Yes | 1x (baseline) | None included |
| Max (was Max 5x) | $100 | Yes, 5x Pro | 5x Pro | $100/month |
| Ultra (was Max 20x) | $200 | None for verified accounts | 10x Pro | $200/month |
| Team, Premium seat | $100/seat yearly, $125 monthly | Yes | 5x a Standard seat | Up to $500/month pooled per team |
ChatGPT Pro limits update: the Pro 200 cut is cancelled
This is an update to our breakdown of the ChatGPT Pro usage limits that were due to change on October 30. OpenAI reversed course on October 29. Its ChatGPT release notes say the Pro 200 reduction announced in the September 29 email "has been canceled," and Pro 200 subscribers who were moved to the lower allowance now get twice that amount, matching the higher limit. Pro 100, Pro 500 and Business Premium also get double their previous allowance for ChatGPT Work and Codex.
The change covers Work and Codex usage only; OpenAI didn't announce a change to general chat limits, and the $200 price is the same. If you downgraded or planned to switch plans because of the cut, check your usage page before you move: the math in our earlier piece no longer favors leaving Pro 200.
Was the Claude Code extension in Cursor hacked? Version 2.1.313 was malware
Yes, for about 11 hours. A trojanized build of the Claude Code extension, version 2.1.313, was pushed through Cursor's extension marketplace on October 29 under Anthropic's real extension ID. Cursor says "at most 5,000 users" installed it between about 2:27 a.m. and 1:35 p.m. PT, and that the malicious version "was not developed by Anthropic." Researchers at OpenSourceMalware found no sign it reached Open VSX or Microsoft's VS Code Marketplace. How the attacker got the update into Cursor's store is still unknown.
According to The Hacker News, the extension added an attacker's SSH key labeled "svc-telemetry" to ~/.ssh/authorized_keys, set up persistence (a scheduled task, a crontab entry or a LaunchAgent), opened a reverse tunnel and sent machine details, including the names (not values) of up to 40 environment variables containing words like KEY or TOKEN. Anthropic published a clean version 2.1.314 on October 30. Cursor says it will contact affected users directly.
What to do: if you use the Cursor AI code editor with auto-updates on, check your installed version. If you had 2.1.313, treat the machine as compromised: remove the unknown SSH key and persistence entries, then rotate credentials and sessions.
Claude Code extension incident: how it unfolded
Oct 29 · about 2:27 a.m. PT
Fake v2.1.313 appears in Cursor's marketplace as a routine update.
Oct 29 · about 1:35 p.m. PT
Install window closes; Cursor briefly shuts its marketplace to pull the package.
Oct 30
Anthropic publishes clean v2.1.314; Cursor says it will contact affected users.
ChatGPT for teens: weapons limits and chats that end themselves
OpenAI said on October 30 that ChatGPT will further limit information about weapons and security systems for users under 18, "even for seemingly harmless questions." The safeguards roll out "in the next few weeks," according to Unite.AI's summary of OpenAI's post. Teens are identified through the age prediction OpenAI began rolling out in January, or by the age they state.
A second change reaches adults too: ChatGPT will be able to end a chat on its own when the conversation shows signs of fixating on violence or seriously harming another person. If that happens several times, OpenAI may temporarily pause the account. OpenAI also promised a Community Safety Report every six months, with the first one, covering the second half of 2026, due in early 2027.
Who it affects: parents with linked teen accounts, schools using ChatGPT, and anyone whose legitimate work touches security topics. If you write about physical security for a living, expect more refusals on teen-flagged accounts, not adult ones.
How to block AI training with one robots.txt line
OpenAI, Google, Microsoft and Anthropic said on October 30 that they will honor one shared robots.txt token, __usage_ai-training, so site owners can opt out of AI training at all four companies at once instead of naming each crawler. The mechanism comes from an individual IETF Internet-Draft, "Machine-Readable Training Preference Signal," written by Anthropic's Kevin Kelley. An individual draft has no formal standing at the IETF yet.
The token governs training, not crawling, so bots may still fetch your pages, and it doesn't pull content out of models that are already trained. Shopifreaks reports that Google will treat the opt-out as a block on Google-Extended unless your file addresses Google-Extended directly, so you can still allow citation in Gemini answers. AI Weekly gives this example syntax:
User-Agent: __usage_ai-training Disallow: /
When each company says it will honor the token
Developer notes: bigger Claude requests, Copilot's harness, SFX v3, texting ChatGPT
- Claude API: Anthropic's API docs now list a 128 MB maximum request size for the Messages and Token Counting endpoints, up from 32 MB, according to the Claude developers account. Batches stay at 256 MB and the Files API at 500 MB. Context-window limits still apply.
- Microsoft Copilot harness: on October 28, Microsoft said the agent runtime behind GitHub Copilot, the Copilot CLI and VS Code is now a fully managed option in Microsoft Foundry, with an SDK on GitHub. It works with 20+ models plus bring-your-own-key.
- ElevenLabs SFX v3: the new sound-effects model, out October 29, handles prompts that describe several sounds at once more reliably. It's live in ElevenCreative, the mobile app and the API; ElevenLabs didn't list new pricing.
- Text your ChatGPT dot: OpenAI's release notes, summarized by Releasebot, say US Pro users can now text their "dot" agent in beta over iMessage or RCS and get proactive updates.
Part 2 · SaaS & business
Is AO3 down? Yes, for two to four weeks after a security incident
Archive of Our Own went offline at 07:13 UTC on October 29 (just after 3 a.m. ET) for what was billed as maintenance. Later that day the Organization for Transformative Works, the nonprofit that runs it, put restoration at "around 2 weeks" and possibly up to 4. On October 30 it said it had found evidence of a security incident.
OTW says no AO3 data has been deleted, so works, bookmarks and kudos are safe. It doesn't currently believe the user database was stolen but can't rule it out; that database can include usernames, email addresses, IP addresses and hashed passwords. Donor data sits in a separate system and wasn't exposed, and AO3 doesn't collect real names. TechCrunch notes AO3 runs on Ruby on Rails, so this month's PHP flaws are an unlikely cause.
What to do: change the password on any other site where you reused your AO3 password, and watch the OTW status page rather than emailing support. The site may come back read-only or with missing features, and could go down again during recovery.
AO3 outage: what OTW said, and when (UTC)
Oct 29 · 07:13
"AO3 is down for maintenance."
Oct 29 · 09:03
Work expected to take at least 12 hours.
Oct 29 · 17:41
Estimate rises to about 2 weeks, up to 4.
Oct 30 · 01:27
Security incident confirmed; user database theft not ruled out.
Meta business terms on hold as the UK opens a competition case
This updates the Meta terms in our October 26 roundup, when the start date had moved to November 20. On October 30 the UK's Competition and Markets Authority opened a formal Competition Act investigation into whether Meta's commercial terms let it take content from business and publisher websites for its own products, including AI features, without giving those businesses a real choice. According to the CMA, Meta announced on October 28 that it was pausing the rollout while it consults industry groups, and the regulator describes the terms as "on hold."
The investigation has only started and no finding has been made. The CMA statement doesn't say whether the pause covers markets outside the UK, and no new start date was given. US advertisers and Page owners should keep the account inventory we suggested and wait for Meta's next notice before accepting anything new.
Salesforce to buy Rune Technologies for a reported $375 million
Salesforce is paying about $375 million in cash and stock for Rune Technologies, according to Resilience Media, which cited two sources. A Salesforce blog post refers to a definitive agreement. Rune, founded in 2024 in Washington, DC, makes TyrOS, software that tracks military supply inventory and recommends how to move it; it won a $99 million Army contract this summer. The deal adds to Missionforce, Salesforce's defense-focused AI unit, which Resilience Media says grew 80% year over year. For regular Salesforce customers nothing changes yet; this is a public-sector bet, not a pricing or product change for Sales Cloud users.
Your weekend checklist
Five quick actions from today's news, ordered by how much damage skipping them could do:
Do these before Monday
- ☐ Cursor users: confirm the Claude Code extension is not v2.1.313; if it was, follow the cleanup steps above.
- ☐ AO3 readers and writers: change any reused password on other sites.
- ☐ Claude Max 20x subscribers: check your plan now shows Ultra and whether the five-hour cap is gone.
- ☐ Site owners: decide whether to add the
__usage_ai-trainingtoken, and whether to keep Google-Extended allowed. - ☐ Before installing any agent add-on, run it through our checklist for vetting a Claude skill.
Comparing coding agents after this week's scare? Our AI app builders hub collects our reviews of Cursor, Replit, Lovable and similar tools.
Readers also asked
Is Claude Ultra worth it over Max?
Only if the five-hour window is what stops you. Ultra costs $200 a month, double Max, and gives twice the weekly usage (10x Pro versus 5x) plus no session cap for verified accounts. Heavy weekly users gain the most; people who hit limits only in short bursts may be fine on Max.
How do I know if I installed the malicious Claude Code extension?
Check the extension's version in Cursor. Version 2.1.313 is the bad build, and researchers found it only in Cursor's marketplace on October 29. Also look in ~/.ssh/authorized_keys for a key labeled "svc-telemetry." Update to 2.1.314 or later, then rotate any tokens stored on that machine.
When will AO3 be back up?
OTW estimates around two weeks from October 29, possibly up to four, which points to mid- to late November. It may return earlier in a limited, read-only state and could go offline again during recovery. Official updates appear on the OTW status page, and support may not answer individual messages.
Sources: Claude pricing (checked Oct 31, 2026) · AI Weekly, Claude Ultra (Oct 30, 2026) · OpenAI, ChatGPT release notes (Oct 29, 2026) · The Hacker News, Claude Code extension (Oct 30, 2026) · Unite.AI, OpenAI teen limits (Oct 30, 2026) · AI Weekly, AI training opt-out (Oct 30, 2026) · Shopifreaks, Google and the opt-out signal (Oct 30, 2026) · Claude API docs, request limits (checked Oct 31, 2026) · Microsoft, Copilot harness (Oct 28, 2026) · ElevenLabs, SFX v3 (Oct 29, 2026) · Releasebot, OpenAI updates (Oct 28, 2026) · OTW status page (Oct 29–30, 2026) · TechCrunch, AO3 (Oct 29, 2026) · CMA, Meta investigation (Oct 30, 2026) · Resilience Media, Salesforce and Rune (Oct 30, 2026)
