Two of Claude Code’s best-known builders leave in the same week.
AI NEWS
Past AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
This day
October 25, 2026
11 items for 2026-10-25, ranked by combined engagement. Unverified items are marked.
Top 3
AI video moves from clips to episodic series.
Switching reasoning mid-session without losing cache saves cost on long runs.
All
Claude Code’s Thariq Shihipar leaves Anthropic; Claude Code desktop creator Anthony Morris has also left
Two of Claude Code’s best-known builders leave in the same week.
CollapseClaude Code’s Thariq Shihipar leaves Anthropic; Claude Code desktop creator Anthony Morris has also left
Thariq Shihipar (@trq212) announced he has left Anthropic after more than a year working on Claude Code, which he said grew larger than he ever imagined. He said he still believes deeply in Anthropic’s mission and its idea of a “race to the top,” and called his colleagues the most talented and earnest people he has met. He plans to take a break before starting something new, and pointed people to @addyosmani, @lydiahallie, @_catwu, @ClaudeDevs and Claude Code originator @bcherny for more Claude Code content. The same day Anthony Morris (@amorriscode), who created Claude Code on desktop, posted that October 23 was his last day at Anthropic and that he will share what’s next soon.
Grok Imagine releases episode 1 of its AI series “Odyssey”, a ~7-minute film made with director Dev Walker and Wonder Studios
AI video moves from clips to episodic series.
CollapseGrok Imagine releases episode 1 of its AI series “Odyssey”, a ~7-minute film made with director Dev Walker and Wonder Studios
xAI’s Grok Imagine account released episode 1 of its “Odyssey” series, made with Grok Imagine; the video runs about 7 minutes 12 seconds and is available in 4K. It is a creative collaboration with Wonder Studios (@WonderStudiosX) and director Dev Walker (@glumlot_). The account asked viewers whether they want a full breakdown of how it was made. Elon Musk shared it with the caption “Odyssey with Grok Imagine,” and his post also passed 1.8 million views.
OpenAI Codex lead: faster call connections in dots, and changing reasoning mid-conversation in Codex no longer breaks the cache
Switching reasoning mid-session without losing cache saves cost on long runs.
CollapseOpenAI Codex lead: faster call connections in dots, and changing reasoning mid-conversation in Codex no longer breaks the cache
OpenAI Codex lead Tibo Sottiaux (@thsottiaux) posted “Day 19” of his running updates, calling them small ones. Calls in dots now connect much faster; changing the reasoning setting in Codex and work mid-conversation no longer invalidates the cache; and the Codex CLI has been updated. The post gives no version numbers or speed figures.
Block unveils Dessa, a full-duplex speech model trained from scratch only on human conversation that runs locally on an M4 Max laptop
A new path that learns language straight from sound, skipping text.
CollapseBlock unveils Dessa, a full-duplex speech model trained from scratch only on human conversation that runs locally on an M4 Max laptop
Jack Dorsey’s Block released Dessa, which it says is the first AI trained only on human speech that can hold a realistic live conversation: no text and no pretrained models, learning to talk from scratch. Dessa is a true speech-to-speech, full-duplex model that listens and talks at once and knows when to cut in or backchannel; inference runs faster than real time and it runs locally on a MacBook Pro M4 Max or better. Block says it learned language from hundreds of times fewer words than LLMs and was trained on a single NVIDIA GB200 NVL72. The small lab, founded by former members of Toronto deep-learning startup Dessa, reports directly to Jack Dorsey, and Block is inviting a small group of AI safety researchers to get early access and probe its risks.
a16z: GPU rental rates are up across the board this year — H200 ~+160%, B200 ~+120%, even the 6-year-old A100 costs more
Data showing compute is still scarce.
Collapsea16z: GPU rental rates are up across the board this year — H200 ~+160%, B200 ~+120%, even the 6-year-old A100 costs more
In its Charts of the Week, a16z said GPUs are “aging like fine wine”: rental rates are up across the board since January 1, roughly +160% for H200, +120% for B200, +65% for RTX PRO 6000 and +35% for H100, and even the six-year-old A100 rents for more than at the start of the year. In follow-ups it said rental prices for H200 and B200 sit about 143% and 111% above a standard six-year depreciation schedule, with most of the gap opening after April, when token demand took off. One explanation it offers: token demand on OpenRouter is up about 27x since last November, from ~6T to ~165T tokens a week, and open-weight models now handle about 72% of those tokens, up from about 28%.
Artificial Analysis: Kimi K3.1 Pro scores 64%, the top openly available model for agentic cyber defense, behind only trusted-access GPT-6 Sol
Open models’ cyber capability is nearing restricted closed models.
CollapseArtificial Analysis: Kimi K3.1 Pro scores 64%, the top openly available model for agentic cyber defense, behind only trusted-access GPT-6 Sol
Artificial Analysis published Cyber Index results: Kimi K3.1 Pro scores 64%, 7 points ahead of the previous best openly available model, DeepSeek V4.1 Flash 1017 (57%), and 5 points behind GPT-6 Sol (Daybreak Blue, max, 69%), which is only offered through OpenAI’s trusted-access program; it is up 23 points from Kimi K3 (41%). K3.1 Pro posts the highest CyberGym-E2E-AA score of any model at 95%, with 37% on DeepsecBench-AA and 61% on CWE-Bench-AA. AA notes that proprietary models from OpenAI, Anthropic and Google may be stronger but lose points to safety refusals (the trusted-access GPT-6 Sol scores 32 points above its public version), while Kimi K3.1 Pro hit no refusals. It is 33% cheaper per token than K3 but uses about twice the input tokens, so its cost per task is $2.81, 15% higher.
LiveBench: Kimi K3.1 scores below DeepSeek Flash, and Bindu Reddy says it looks “bench-maxxed”
Benchmarks and early hands-on verdicts diverge on a new model.
CollapseLiveBench: Kimi K3.1 scores below DeepSeek Flash, and Bindu Reddy says it looks “bench-maxxed”
Bindu Reddy, CEO of Abacus.AI and a backer of LiveBench, posted that Kimi K3.1 scores below DeepSeek Flash on LiveBench. In her view K3.1 seems “bench-maxxed” on public benchmarks, is competitive on price per task, but is not in the Fable or Astra class, and DeepSeek is still the king of open source. AI2 researcher Nathan Lambert separately said that at the top end, real-world usability matters far more than ±2 points on AA, and K3.1 may be closer to a miss. Both are individual evaluations and opinions, and they contrast with the AA index result above.
Apollo’s chief economist: AI has delivered no financial payoff to corporate America outside tech, with flat revenue per worker and margins for the S&P 493
A sober Wall Street view on AI’s payoff.
CollapseApollo’s chief economist: AI has delivered no financial payoff to corporate America outside tech, with flat revenue per worker and margins for the S&P 493
Gary Marcus quoted Apollo chief economist Torsten Slok, who says that despite massive investment and hype, AI has delivered zero financial payoff to corporate America outside tech. Slok notes that revenue per worker and profit margins are flat for the S&P 493 (the index minus the biggest tech names), with no sign of AI boosting measured productivity growth. He says mainstream companies are pulling back and starting to treat token spending as a cost center, and don’t need or want to pay for the latest model; as long as agentic AI and token spending stay confined to a small part of the economy, AI’s macro effects will be limited. Marcus argues this strengthens the bear case on AI.
The Information: OpenAI wins Anthropic coding customers with steep discounts, as Rippling and Replit shift most spend to OpenAIUnconfirmed
The coding-model market has turned into a price war.
CollapseThe Information: OpenAI wins Anthropic coding customers with steep discounts, as Rippling and Replit shift most spend to OpenAI
According to The Information (as relayed by AI commentator kimmonismus and others), OpenAI is using steep discounts to win Anthropic’s biggest coding customers. Replit moved most of its AI spending to OpenAI after getting more than 30% off a contract worth over $100 million a year, after Anthropic declined to extend its earlier discount without a bigger commitment. At Rippling, Anthropic’s share of coding work fell from 57% to 18% since May while OpenAI’s rose from 16% to 52%, and OpenAI offered DataCamp 40% off on condition it use Codex. The report notes this may deepen OpenAI’s losses, while Anthropic’s renewed lead on OpenRouter suggests better models can win users back quickly.
Perplexity Computer adds the open-source /impeccable skill for interfaces that don’t look AI-generated, with 5,000 free credits per user
Agent products start competing on design taste.
CollapsePerplexity Computer adds the open-source /impeccable skill for interfaces that don’t look AI-generated, with 5,000 free credits per user
Perplexity CEO Aravind Srinivas announced that /impeccable is now a skill inside Perplexity Computer. It is an open-source skill built by Paul Bakaus (@pbakaus) to help people build interfaces that don’t look like AI-generated slop. Perplexity is giving each user 5,000 credits to get started with it.
The Information: Altman told staff the industry needs ways to slow dangerous AI, from international regulation to merging labs or one company dominating computeUnconfirmed
A leading lab CEO is discussing how to put on the brakes.
CollapseThe Information: Altman told staff the industry needs ways to slow dangerous AI, from international regulation to merging labs or one company dominating compute
According to The Information, Sam Altman told OpenAI employees that the industry needs better ways to slow the development of potentially dangerous AI. His proposed paths include international regulation, merging the leading AI labs, or allowing one company to dominate computing resources. The post gives only the key points; the full story is paywalled and we could not check it.