Skip to content

← X AI Daily

X AI Daily

Date25 October 2026 · SundayWindow2026-10-24 02:00 to 2026-10-25 07:59CoverageComplete

Translation

English translation of the Chinese original.

1. AI Models, Products & Applications

Two Claude Code leads leave Anthropic the same day: Thariq Shihipar and Anthony Morris

  • What happened: Claude Code team member Thariq Shihipar posted at 10-24 12:04 confirming his departure from Anthropic, saying he spent more than a year on Claude Code and watched it scale beyond imagination, and that he will take a break before starting something new. He gave no reason, and named @addyosmani, @lydiahallie, @_catwu, @ClaudeDevs and @bcherny as accounts worth following. About eight and a half hours later, Anthony Morris, creator of the Claude Code desktop app, posted at 10-24 20:33 that "yesterday was my last day at Anthropic," also without explanation.
  • Why it matters: Thariq was one of Claude Code's most active official faces on X, long responsible for explaining features and answering developer questions. Two departures within a day, just as The Information reported Anthropic is preparing to list next month, quickly moved discussion from "who left" to "why now." Both posts read as positive farewells, and neither side hinted at internal problems.
  • Posts & engagement: Thariq's departure post (10-24 12:04, 4.12M views, 27.6K likes, 950 reposts, 2013 replies) was the most-engaged AI post of this cycle; he later joked under @lydiahallie's post asking "is it just me or has Opus 5.5 gotten dumber," and that reply (12:23, 750K views, 9944 likes, 514 replies) was also widely shared; Anthony Morris's departure post (10-24 20:33, 610K views, 7331 likes, 204 reposts, 701 replies). Before Thariq posted, @ArfurGrok (07:54, 626K views, 1780 likes) had already reported it. @JasonBotterill (12:50, 140K views, 1706 likes) said someone found out early by checking whether his Google Workspace company email still existed; that claim is unverified. On the listing timeline, see The Information's post (10-24 04:30, 38K views, 88 likes).
  • Verification: The fact and timing of the departures rest on the two men's own posts, whose text was checked; Morris's role as creator of the Claude Code desktop app comes from his own post. There is no first-hand information on reasons or destinations; claims such as "leaving ahead of the IPO" or "starting a new lab" are speculation.
  • Take: In the short term this is a staffing change, not a product change. What is worth watching is who takes over Claude Code's external communications and whether more core engineers leave around the listing. Two cases do not add up to an exodus.

Anthropic's single-agent guide spreads widely on X: fewer subagents, more Skills and on-demand tools

  • What happened: An Anthropic article from 10-22, "Building agents: When to fold subagents into your main agent loop," circulated widely in this window. It says that after comparing deployments across enterprises, the Applied AI team found a single agent with Skills often beats both "one big prompt does everything" and "multiple subagents" on quality, usually at lower cost and lower per-task latency. It recommends turning subagent prompts into Skills and their API calls into MCP tools, merging them one at a time and comparing with evals. Subagents still fit three cases: context bloat causing measurable performance degradation, independent and disposable contexts that must run in parallel, and a single agent choosing the wrong tools or facing conflicting system prompt instructions.
  • Why it matters: This is the first time a model vendor has systematically told enterprises to delete structure rather than add orchestration. Specific signals in the article: one customer's single-agent loop ran about 2x faster in a comparison; Anthropic cut Claude Code's system prompt by more than 80% for stronger models; and tool selection degrades once the system prompt defines more than 30 to 50 tools. The article also names "architecture becoming a copy of the org chart" as a common failure mode.
  • Posts & engagement: @cherry_mx_reds's explainer (10-24 20:21, 320K views, 1672 likes, 104 reposts, 113 replies) was the main driver; @omarsar0's addendum (10-25 02:13, 13K views, 176 likes) argues subagents remain useful for parallel research and similar tasks, which is his own main use now.
  • Verification: Checked against the original Anthropic article, confirming the "often" wins wording for single agent plus Skills, the single-customer case of roughly 2x speed, the 80%+ cut to Claude Code's system prompt, the 30 to 50 tool-definition threshold, the three conditions for subagents, and the migration toolkit and /claude-api Skills series. The roughly 2x figure is a vendor's account of a single customer, not a general conclusion. The article was published 10-22, before this window; what is recorded here is its spread in this window.
  • Take: For teams that already built "a router plus a pile of specialist agents," the article offers an executable way back: build a small end-to-end eval first, then merge one subagent at a time and keep the merge if scores do not drop. It is not "subagents are obsolete" but "start from a single loop by default."

Two disputes after the Kimi K3.1 Pro launch: LiveBench below DeepSeek Flash, and promo posts without a paid-partnership label (Follow-up)

  • What happened: The previous daily covered the Kimi K3.1 Pro launch. Two new threads emerged in this window. First, Bindu Reddy posted that Kimi K3.1 scores below DeepSeek Flash on LiveBench, arguing it is tuned for public benchmarks, that per-task pricing is competitive, but that it is not in the Fable or Astra tier, and that DeepSeek is still the top open model. Second, multiple creators alleged that Moonshot's promotional campaign sent them a uniform script whose screenshots instruct contributors to submit a draft for review, post only after approval, and not use the paid-partnership label.
  • Why it matters: The first is a gap between third-party and vendor-reported benchmarks; the second directly affects the credibility of K3.1 demo content on X. @birdabo's point is that plenty of creators would have shown off the model for free, and the campaign now makes every demo look suspect.
  • Posts & engagement: @bindureddy's LiveBench post (10-24 07:37, 53K views, 786 likes, 50 reposts, 86 replies); @birdabo's comment on the promo dispute (10-24 11:42, 13.9K views, 222 likes); the creator-guidance screenshot relayed by @buildwithrajath (10-24 17:54, 753 views, 4 likes).
  • Verification: For the LiveBench ranking, only Bindu Reddy's post and screenshot were obtained; no snapshot of the LiveBench leaderboard page was retrieved, so it is a sourced claim. The authenticity and origin of the paid-promo guidance screenshot were not verified. A check of Moonshot-related account @crystalsssup's posts in the window found only 3D modeling demos and ranking promotion, with no response to the dispute.
  • Take: The benchmark gap needs the LiveBench page and more independent evaluations to confirm. If the promo dispute holds, the damage is not to the model but to the community's trust in K3.1 user feedback going forward; independent tests after the weights drop before 11-04 will matter more.

Other

  • Perceptions of Claude limits swung back: @JustinGorya (10-24 18:52, 21.6K views, 374 likes) said Claude's 5-hour session limit is "back to normal" and that the $20 plan can use Opus 5.5 normally; this is user perception only, and no official Anthropic statement on limit changes was found.
  • The Information reported that a hacker group targeting banks used Grok to gather intelligence, Claude to plan attacks, and Kimi for tasks Claude refused, combining models to bypass safety limits: The Information's post (10-25 01:30, 4468 views, 23 likes). The original is paywalled and was not read in full; this is a sourced claim.

2. AI Compute & Industry

OpenAI uses steep discounts to poach Anthropic's coding customers: Rippling share flips, Replit migrates

  • What happened: According to The Information, OpenAI is courting Anthropic's large coding customers with steep discounts. Key figures as relayed: in coding work at HR software company Rippling, Anthropic's model share fell from 57% in May to 18% while OpenAI rose from 16% to 52%; Replit moved most of its AI spending to OpenAI after getting a discount above 30% and a contract worth more than $100M a year, prompted by Anthropic not renewing its previous discount unless Replit committed to more spending; OpenAI also offered DataCamp a 40% discount on condition it use Codex rather than rival coding tools.
  • Why it matters: Coding is where Anthropic's revenue is most concentrated, and this shift is driven by pricing and contract terms, not by a capability gap. The relayed account also notes OpenAI had projected compute spending above revenue this year, meaning the discount war is buying share at a loss. @kimmonismus's view is that Anthropic leading again on OpenRouter shows a better model can win users back quickly.
  • Posts & engagement: @kimmonismus (10-25 02:50, 18K views, 296 likes, 75 replies); @choblin29's detailed relay (10-25 01:07, 7330 views, 71 likes); Techmeme (10-25 03:45, 5313 views, 33 likes) summarized the above-30% Replit discount and the contract above $100M a year.
  • Verification: The Information's original is paywalled and was not read in full. The Rippling share figures are attributed in Subvolts' relay to Rippling CEO Parker Conrad, consistent with the relay on X. Replit making OpenAI the default has traceable background: a TechCrunch report on 10-17 recorded Replit CEO Amjad Masad's remarks at TechCrunch Disrupt and his later clarification that users can still choose Claude. The Replit 30% discount and $100M contract, and the DataCamp 40% discount, currently exist only in The Information's version as relayed secondhand and are unverified.
  • Take: Competition among model vendors is starting to look like cloud contracts: discounts, share terms, renewal windows. For buyers, "model portability" is becoming real negotiating leverage, and the relayed detail that Replit switched its code in a few hours says more than any benchmark.

Other

  • The Information reported that customers automating work in Salesforce with external AI tools such as Claude may face new charges, and some enterprises are already reconsidering usage, with some even considering moving data out: The Information's post (10-24 21:00, 12K views, 20 likes). The original is paywalled; this is a sourced claim.

3. Chinese AI Ecosystem

No Chinese-language AI news worth a separate entry was found in this cycle. There was some Chinese-language relaying of Thariq's departure, for example @charles12388 (10-24 16:59, 703 views), consistent with the English side and adding nothing new.

Events including the first episode of Grok Imagine's "Odyssey" were not included because no official original post was obtained.