AI Weekly #13/2026: OpenAI Buries Sora — AI Industry Pivots to Profitability

TL;DR

This week in 30 seconds:

• OpenAI Sora: Video app discontinued after less than a year, $1B Disney deal dissolved — AI profitability beats product experiments.
• Anthropic vs. Pentagon: Federal court lifts DoD blacklisting — first successful First Amendment ruling in the AI context in the US.
• Apple opens Siri: iOS 27 brings Extensions API — Claude and Gemini as Siri backend on one billion devices, planned for fall 2026.
• LiteLLM hack: Supply chain attack activates on installation — over 47,000 users of affected versions potentially exposed.

---

Audio Version

11:34 | Download MP3

Chapters

- 0:00 - TL;DR - 0:42 - Story of the Week - 3:07 - More Top Stories - 6:57 - Tool of the Week - 7:59 - Fail of the Week - 9:00 - Number of the Week - 9:48 - Reading List - 10:52 - Next Week

Read aloud with edge-tts (en-US-AndrewNeural)

---

Story of the Week

Sora Is Dead — And That Says Everything About OpenAI in 2026

On a single Tuesday in the last week of March, OpenAI axed the Sora video app, dissolved the billion-dollar Disney deal, reassigned a manager — and simultaneously announced a fresh $10B funding tranche [1]. The message was clearer than any press release: when you have $120B in investor capital behind you [10], you need to deliver, not experiment.

Sora had a structural problem after its spectacular launch. Runway, Pika, and Chinese competitor Kling caught up significantly within months and were considered competitive with OpenAI’s solution in many respects [1]. Internally, the app was deemed not competitive enough for a standalone market presence. Rather than continuing to invest, the leadership team decided to redirect resources toward enterprise products and B2B revenue — the only viable course ahead of a potential IPO.

The numbers are stark: A Disney deal worth $1B was dissolved [1]. SoftBank is financing its $30B share in OpenAI’s round through a 12-month, unsecured loan of $40B from JPMorgan and Goldman Sachs [10]. Private equity investors receive a guaranteed minimum return of 17.5% [10] — unusual for a startup and a strong signal that analysts interpret as preparation for an imminent public offering.

For developers and users, this means: The OpenAI portfolio is shrinking strategically. Anyone who bet on Sora integration needs to replan — Runway Gen-3, Pika 2.0, and Kling are mature alternatives without OpenAI’s valuation pressure. For enterprise customers, the pivot signals that OpenAI is focusing more heavily on API products and B2B offerings.

Open questions: The decision raises questions about the “how,” not the “what.” External developers who had built on the Sora API learned about its end from the press [1]. That’s not a good sign for platform trust in a pre-IPO year when OpenAI is counting on enterprise credibility.

Bottom line: OpenAI is cleaning house — not because Sora was bad, but because $120B in investor capital demands a different focus than a consumer video experiment that China can copy faster than OpenAI can scale.

---

More Top Stories

Anthropic vs. Pentagon: Federal Court Sets Precedent for AI and Free Speech

A federal judge has preliminarily lifted the Pentagon’s blacklisting of Anthropic [2]. The DoD had classified Anthropic as a “supply chain risk” — citing security concerns around Anthropic’s major investors (including Google, Amazon, and Saudi PIF). The court, however, also saw potential retaliation against the company’s critical press coverage of the DoD procurement process and deemed the classification a potential violation of the First Amendment [2]. The case was confirmed by TechCrunch, CNBC, and WSJ [2].

The ruling is historic: it is the first time an AI organization has successfully won a First Amendment lawsuit against a US federal agency. The blacklisting would have effectively excluded Anthropic from the government procurement market — one of the highest-volume AI markets globally [2]. Bloomberg also reports that Anthropic is weighing an IPO as early as October 2026 [7] — the company is visibly fighting from a position of strength.

So what? For the entire AI industry, this is a relevant signal: regulatory arbitrariness through government classifications now has a proven legal counterweight. Future administrations will think twice before restricting AI companies for political reasons.

---

iOS 27 Opens Siri to Claude, Gemini & Co. — Apple’s Quiet Revolution

According to Bloomberg reporter Mark Gurman [3], iOS 27 will bring a new extensions system allowing third-party chatbots as a Siri backend. Users will be able to set Claude, Gemini, or other AI assistants as the default intelligence for Siri requests — analogous to choosing a default browser [3]. Apple plans to announce at WWDC; launch is scheduled for fall 2026.

The strategic dimension is significant: Apple is opening access to approximately one billion active devices [3] to Anthropic, Google, and potential additional providers. The backdrop includes both competitive pressure and regulatory requirements under the EU Digital Markets Act. For Anthropic, this represents a direct distribution channel that would otherwise require years of independent app development.

So what? Apple loses little — Siri was never a differentiator against Android. The AI assistant market gains a distribution surface that could fundamentally break up the duopoly between ChatGPT and Google Assistant. Whoever builds the better API experience wins.

---

Quick Hits

In brief:

• David Sacks: Trump’s AI and crypto czar leaves the White House after 130 days — the statutory maximum for Special Government Employees (SGEs). A successor for the US AI agenda has not yet been named [5].
• ARC-AGI-3: New benchmark tests “agentic intelligence” in interactive environments — humans solve 100% of tasks on first contact, frontier models collectively under 1% at time of publication [6]. 25 tasks are publicly playable at arcprize.org.
• IPO race 2026: Bloomberg reports Anthropic is weighing an IPO as early as October 2026 [7] — in parallel with OpenAI’s SoftBank loan signal. Analysts consider it possible that both leading AI labs could go public before the end of 2026.
• LiteLLM patch: Users of versions 1.82.7 and 1.82.8 should update immediately and treat all credentials as compromised — SSH keys, AWS credentials, Kubernetes configs, and ~30 other file types were within the attack scope [4].

---

Tool of the Week

Claude Auto Mode — Autonomous Action with Built-in Classifier Protection

Disclosure: This newsletter is generated using Claude (Anthropic). The following tool is from the same provider — we recommend it because it is technically relevant, but disclose this connection.

Claude Code has a new “middle path” between full permission dialogs and the risky --dangerously-skip-permissions flag [8]. Auto Mode allows longer, autonomous tasks — an integrated classifier checks every action before execution and automatically blocks: mass file deletion, credential exfiltration, and suspicious code.

“It’s a middle path that lets you run longer tasks with fewer interruptions while introducing less risk.” [8]

Available immediately for Team Plan users (Research Preview), Enterprise and API to follow in the coming days. Compatible with Claude Sonnet 4.6 and Opus 4.6 [8]. Particularly useful for teams that want to automate complex multi-step tasks without giving up security controls.

Learn more about Claude Auto Mode

---

Fail of the Week

“We’re Building Our Own Model” — Cursor’s Kimi Surprise

Cursor presented “Composer 2” as an in-house AI coding model — and forgot to mention that it is built on Kimi 2.5 from Moonshot AI, an Alibaba-backed Chinese company [9]. This wasn’t discovered through an audit, but by an X user named Fynn who found code hints and made them public. Cursor subsequently confirmed.

Root Cause: Not a technical bug — a deliberate communication failure. Cursor chose not to proactively disclose the model’s origin. In the geopolitical climate of 2026, where Chinese AI models are a sensitive topic, this was a calculated non-disclosure that has come back to bite them [9].

What we learn: Anyone building an AI product must disclose the model’s origin — especially when the base model raises geopolitical questions. “Nobody asked” is not an acceptable communication strategy in 2026 for a highly valued tool with a broad developer user base.

---

Number of the Week

$120,000,000,000

OpenAI’s total funding after the latest $10B tranche, confirmed by CFO Sarah Friar [10]. For comparison: the entire European AI investment landscape was under $50B in 2025 — OpenAI alone exceeds the European annual volume by more than double.

Particularly revealing is the financing structure: SoftBank is covering its $30B share through an unsecured 12-month loan from JPMorgan and Goldman Sachs [10]. Private equity investors receive a guaranteed minimum return of 17.5% [10]. Such structures don’t emerge for companies going public in three years — but for those that analysts say could be there in 12 months. The clock is ticking.

---

Reading List

For the weekend:

1. The Malicious LiteLLM Package — Simon Willison walks through step by step how a supply chain attack abuses CI/CD credentials and activates on installation — without ever executing import litellm. Required reading for anyone using PyPI packages in production environments. (8 min)

2. ARC-AGI-3 Thread — The researchers behind ARC-AGI explain why the new benchmark is intentionally unsolvable through memorization — and why <1% frontier performance against 100% human performance remains the most honest AGI distance measurement tool. 25 tasks are directly playable. (5 min)

3. Why SoftBank’s $40B Loan Points to a 2026 OpenAI IPO — TechCrunch explains the financing structure behind OpenAI’s $120B round and why a 12-month unsecured loan is a more precise IPO timing statement than any official announcement. (6 min)

---

Next Week

What’s coming:

• Apple WWDC lead-up (CW 14): Following the iOS 27 Siri Extensions leak, analysts expect first official comments or denials from Apple — speculation cycle begins
• Anthropic vs. DoD: The preliminary injunction has been granted, the main hearing is still pending — further movement in this precedent-setting case is expected
• OpenAI IPO signals: Following the SoftBank loan structure, coverage of a potential 2026 IPO is likely to intensify — Bloomberg and WSJ are watching closely

---

🤖 Behind This Newsletter

Generated in: ~45 minutes (Auto Mode) Sources scanned: 20 articles from 8+ feeds Stories found: 20 → 7 selected Validation: 4 Agents (Fact-Check, Devil’s Advocate, Quality Editor, Legal Compliance) Model: Claude Sonnet 4.6 + Haiku (Validation) Images: Pollinations.ai (Phase 3.5)

Full Metrics

Phase	Metric	Value
Source collection	RSS feeds	8+
Source collection	WebSearch queries	12
Selection	Stories presented	20
Selection	Stories selected	7
Draft	Words	~1,400
Draft	Sources cited	10
Validation	Fact-Check issues	3
Validation	Balance issues	5
Validation	Quality issues	3
Validation	Legal issues	1

---

This newsletter was researched and written AI-assisted. Images generated with Pollinations.ai.