AI News Today — October 25, 2026: OpenAI Discounts Up to 90% to Win Anthropic Customers, Muse Money Warning, and Kimi K3.1 Pro Nears the Frontier

The big picture

Today's story is price — and what price reveals about trust. OpenAI is reportedly cutting enterprise prices by up to 90% to pull large customers away from Anthropic, and named customers have shifted real spending. At the same time, a Chinese open-weight model, Kimi K3.1 Pro, scored fourth on an independent intelligence index at $2 per million input tokens — pressure from below that helps explain the discounting above.

Trust is the other thread. A close review of Meta Muse's money features drew a plain warning from an identity-theft charity: the technology is too new to recommend yet. And Pennsylvania became one of the first states to require labels on misleading AI advertising. Capability is still moving — Block showed a voice model trained only on human speech — but buyers, regulators, and safety groups are now setting the terms.

What happened today

Muse AI: money features reviewed — an identity-theft charity says the technology is too new to recommend

What happened: There was no new Muse launch on October 24–25, but Muse's money features got their closest scrutiny yet. A Yahoo Finance review published October 22 and recirculating this weekend walked through what Muse offers: tracking spending and setting budgets, auditing subscriptions, flagging unexpected and duplicate charges, negotiating bills — including by calling providers — searching unclaimed-property databases, and finding deals. Meta says purchases happen only after user approval, that it uses single-use card numbers where no card is stored, and that passwords and payment methods are stored so they stay hidden even from the agent. Meta also says Muse "isn't immune to attack" and "will sometimes make mistakes." The Identity Theft Resource Center does not recommend it yet, calling the technology too new. Muse remains available only in the United States, Canada, and Japan — Japan joined on October 23 — with more than 10 million people having set it up since the September launch, according to Meta chief executive Mark Zuckerberg.

Who it matters to: Anyone tempted to hand an agent their bills, and any business building a money-handling agent. The features are real and useful on paper; the independent safety verdict, for now, is wait.

What you can do: If you try Muse for money, start read-only: let it track spending and flag subscriptions before you let it negotiate, call, or buy anything. Keep purchase approvals on, and check its action trail after every task. If you handle other people's money professionally, wait for the independent recommendations to change.

Learn more: Trade2Spike on the Yahoo Finance Muse money review and the Identity Theft Resource Center warning

Gemini Spark: no new launch — Task mode stands, and the November 17 Gems deadline is next

What happened: There was no new Gemini Spark launch announcement on October 24–25. The only Spark item in the last cycle was the one covered in yesterday's briefing: Google listing temporary sharing of saved Google Password Manager logins with Gemini Spark in ChromeOS 156, with explicit consent plus a biometric or device unlock required for every task. Everything else is unchanged. Spark lives in the main Gemini app as Task mode, with a Chat/Task switch and a Scheduled page. It still requires a Google AI Pro or AI Ultra subscription. And Gems are still set to auto-migrate to Skills on November 17, 2026 — now the next hard date on the Spark calendar.

Who it matters to: Spark users with custom Gems. The product itself is quiet this weekend; your preparation deadline is not.

What you can do: If you rely on a Gem, open it this week, copy its name, description, and instructions somewhere safe, and re-upload any knowledge files into a Skill by hand if you want to move early. Do not wait for the automatic migration to discover a Gem that did not carry over cleanly.

Learn more: FindSkill on where Gemini Spark went — it's Task mode now and gHacks on ChromeOS 156, including the Gemini Spark login-sharing listing

OpenAI reportedly offers discounts up to 90% to win Anthropic's biggest customers

What happened: OpenAI is reportedly pulling major customers away from Anthropic with aggressive discounts, according to The Information, reported on October 25. Rippling chief executive Parker Conrad said his company raised OpenAI's share of its coding workload from 16% to 52%, while Anthropic's share fell from 57% to 18%. Replit switched after OpenAI offered more than 30% off a contract worth over $100 million a year. DataCamp chief executive Jonathan Cornelissen said OpenAI's pricing cut his costs by 80 to 90%. A U.S. bank with more than $7 billion in assets reportedly got its ChatGPT seats doubled to 500 employees at no extra charge, paying under $7 per user instead of the usual $20 or more. The reports say OpenAI is partly buying share at the expense of its margins, helped by real cost cuts from its more efficient Sol and Luna model lines, and that Anthropic's growth slowed from the second to the third quarter, per an investor memo. Treat the figures as reported, not confirmed — the coverage includes no statement from Anthropic.

Who it matters to: Every business buying frontier models. When a $100 million customer can be moved by a discount after a rival's discount expired, list price is theatre — the real price is negotiated, and the models were close enough for price to decide.

What you can do: If your team spends real money on one provider, get a competing quote before your next renewal, even if you plan to stay. Today's reporting is leverage: ask what discount, seat, or credit terms match what comparable customers are reportedly getting — and get any discount's expiry date in writing, because an expiring discount is exactly what started the Replit switch.

Learn more: The Decoder on OpenAI offering discounts up to 90% to poach Anthropic's biggest customers

Moonshot AI's Kimi K3.1 Pro lands fourth on an independent index — open weights promised by November 4

What happened: Moonshot AI launched Kimi K3.1 Pro on October 23, a post-trained upgrade of its Kimi K3 model with a 1-million-token context window. Independent benchmarker Artificial Analysis scores it 55 on its Intelligence Index — fourth overall, 11 points above K3 — while Moonshot's own benchmark runs place it close to Claude Opus 5.5 and GPT-6 Astra on coding tests; those vendor numbers are Moonshot's, not independent results. On Moonshot's API the model costs $2 per million input tokens and $10 per million output tokens, a third below K3, with a Highspeed variant running two to three times faster at twice the price. Moonshot says it removed cyber-related data from training so the open release would not push offensive hacking capability forward. The catch is in the calendar: at launch K3.1 Pro is API-only — Moonshot promises the full weights by November 4, and the licence terms are not yet published.

Who it matters to: Builders choosing between closed frontier APIs and open weights. A soon-to-be-open model scoring fourth overall, at a published price a fraction of the frontier's, is the clearest explanation yet for why the closed labs are discounting so hard.

What you can do: You can test K3.1 Pro today through the Kimi app, Kimi Work, Kimi Code, or the Kimi API — run your own real tasks against your current model before believing any benchmark table. If you need to self-host, wait for November 4 and read the licence before you build on it.

Learn more: SiliconANGLE on Moonshot launching Kimi K3.1 Pro and Skimwire on K3.1 Pro's fourth-place index score and November 4 weights

Block's Dessa is a voice model trained only on human speech

What happened: Block, Jack Dorsey's company, released Dessa from its research lab — a full-duplex voice model, meaning it can listen and speak at the same time, trained entirely on human speech with no text pipeline in between. Block says that is why Dessa carries laughs, pauses, and emotion that text-first voice systems throw away, and that it runs locally on a MacBook with an M4 Max chip or better, trained on a single Nvidia GPU rack. There is no public access yet: Block has put out a call for safety researchers to test the model first, including cases where a realistic voice agent could trick people on the other end of a call.

Who it matters to: Anyone building or buying voice agents — for phone support, reception, or sales. Every major voice product today turns speech into text and back; Dessa is a credible attempt to skip that step, and local running would change the privacy and cost maths if it ships.

What you can do: Join the waitlist if voice is core to your product, but do not plan a launch around it — there is no API, pricing, or release date yet. Use the demos as a quality bar instead: if your current voice agent cannot handle interruptions and back-and-forth naturally, that gap is now the thing to fix.

Learn more: TechCrunch on Block's new voice model that sounds eerily human

Pennsylvania enacts an AI disclosure law for misleading ads, effective December 21

What happened: Pennsylvania Governor Josh Shapiro signed Senate Bill 806, the Artificial Intelligence Deceptive Advertising Act — Act 99 of 2026 — on October 22. It takes effect December 21, 60 days after approval. This is a targeted rule, not a blanket label on every AI-made ad: it bars an advertiser from knowingly, recklessly, or wilfully using synthetic content — images, video, audio, or text significantly modified or generated by AI — in a way that creates a false consumer perception about the nature, origin, or authenticity of goods or services, unless the ad includes a clear and conspicuous disclosure. The disclosure must appear when the synthetic content is first presented, in the same medium, and, where technically feasible, be permanent or extraordinarily difficult to remove. Content limited to text, and content that is evidently fictional, parody, satire, commentary, or criticism, are exempt.

Who it matters to: Anyone who advertises into Pennsylvania — which, for national campaigns, effectively means anyone. A synthetic person, product shot, or voice doing the persuading in a Pennsylvania ad now needs a label a consumer will actually notice, not a footnote.

What you can do: Before December 21, audit your live ads for AI-generated or significantly AI-modified images, video, and audio that could overstate what a product is or does. Add the disclosure at the first presentation, in the same medium — and keep a record of which assets are synthetic, because "significantly modified" questions will be answered from your files, not your intentions.

Learn more: Data Phoenix on Pennsylvania enacting the AI advertising disclosure law

Also worth knowing

• Forced to pick a candidate, Grok and GPT picked opposite parties. A Convovo study led by Cornell professor Tapan Parikh asked the models, via API, to name one candidate in nearly 500 midterm races: Grok 4.7 chose the Republican in 92% of answers and GPT-6.1 Sol chose the Democrat in 97%, agreeing less than 10% of the time. The consumer apps refused the same prompt, and OpenAI said the API test does not reflect the ChatGPT experience. TechCrunch on the Convovo study
• The next agent interface may be your text messages. TechCrunch rounded up agents you text like a person — general assistants, family chiefs of staff, and travel specialists handling calendars, reminders, email, and bookings over SMS, iMessage, WhatsApp, and Telegram, from free tiers to about $21 a month. TechCrunch on the top AI agents that can live in your text messages
• TechCrunch published a running "AI graveyard." The list tracks products and startups that shut down, were folded into larger apps, or missed badly — and cites S&P Global Market Intelligence estimating about 42% of corporate AI initiatives are ultimately abandoned. A useful reality check before your next AI purchase. TechCrunch's AI graveyard
• Moonshot's older K2 models are on a clock. Moonshot's pricing page says the K2 series will no longer be available to newly registered users from October 31, 2026, with a full retirement planned by the end of December; existing users are not affected by the October step. If a workflow still points at a K2 model, plan the move to K3.1 Pro now. GeoToolbox's Kimi explainer

One practical takeaway

Negotiate your AI bill like the big customers do — and fence any agent that touches money. In one weekend, reported discounts of up to 90% moved a $100 million account, an independent index put a $2-a-million-token open model fourth in the world, and a safety charity told consumers not to trust the biggest money agent yet. The lesson is the same on both sides: never pay list price for a model you have not benchmarked against a cheaper rival, and never give an agent spending power faster than independent reviewers — and your own approvals — can keep up with.

---

Sources: TechCrunch AI, The Decoder (reporting The Information), SiliconANGLE, Skimwire, Data Phoenix, Trade2Spike (reporting Yahoo Finance), FindSkill, gHacks. Summaries are original; follow the links for full reporting. OpenAI discount figures, customer shares, and Anthropic's growth slowdown are reported by The Information via The Decoder and are not confirmed by OpenAI or Anthropic; Kimi K3.1 Pro benchmark comparisons other than Artificial Analysis's Intelligence Index score are Moonshot's own runs; K3.1 Pro's weights and licence were promised by November 4 but not yet published at briefing time. Muse money-feature details are from Yahoo Finance's October 22 review as reported October 24; there was no new Muse or Gemini Spark launch announcement on October 24–25.