India AI Digest — Thursday, July 2, 2026

• The Supreme Court set aside insolvency-tribunal orders built on fabricated, AI-generated precedents, declared zero tolerance for unverified AI citations in court filings, and directed the Bar Council of India to frame disciplinary norms.
• Lightstorm, Microsoft, Singtel, and Tata Communications signed contracts to build I-2SEA, a ~3,600 km subsea cable linking Singapore and Kuala Lumpur to dual east-coast landings at Machilipatnam and South Chennai, targeting service in Q4 2029.
• Anthropic restored Claude Fable 5 globally on July 1 after the US government lifted the June 12 export-control order — ending a 19-day shutoff that hit India, Anthropic's second-largest market, hardest. Mythos 5 stays US-only.
• A day earlier, Anthropic shipped Claude Sonnet 5 at an introductory $2/$10 per million tokens, putting near-Opus-4.8 agentic capability at the mid-tier price point.
• Position movements: regulatory_clarity +1 (Supreme Court), compute_infrastructure +1 (I-2SEA), enterprise_adoption_depth +1 (Fable 5 restoration; Sonnet 5 pricing).

---

LEGAL · GOVERNANCE · POLICY · July 2, 2026

Supreme Court sets aside tribunal orders built on AI-hallucinated precedents, directs BCI to frame norms

The Supreme Court of India, in a judgment dated July 2, 2026, set aside insolvency-tribunal orders in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd after finding they relied on fabricated, AI-generated case citations. Singh, suspended director of Essel Infraprojects Ltd, was challenging Section 7 IBC proceedings initiated by J&K Bank; her counsel, senior advocate Madhavi Divan, showed that judgments the tribunal relied on were either fictional or carried AI-generated content falsely attributed to real Supreme Court rulings. Per Bar & Bench's reporting, two of the cited precedents — ICICI Bank Ltd v. Urban Infrastructure Real Estate Ltd (2019) 16 SCC 528 and Sarbjit Singh v. Union Bank of India (2022) 7 SCC 464 — do not exist. The Court called for a zero-tolerance approach to citing AI-generated precedents without verification, held that doing so is misconduct on an advocate's part, and directed the Bar Council of India to constitute a committee to prescribe guiding principles and the disciplinary consequences for violations.

What this means. The Indian judiciary has moved from anecdote to doctrine on AI hallucination. Sanctions for fake citations have been accumulating case by case — lawyers pulled up in individual matters, benches issuing warnings. This is different in kind: a tribunal's own orders stood on citations that do not exist, and the Supreme Court's remedy was structural rather than punitive. Setting aside the orders establishes that AI-fabricated authority contaminates the decision, not just the filing. Directing the BCI to write profession-wide norms converts a judge-by-judge irritant into a regulatory mandate with disciplinary teeth.

The unresolved layer is detection. The fake citations here were caught because appellate counsel checked them. Nothing in the judgment builds verification into the system — no citation-checking requirement at filing, no tribunal-side tooling. Until the BCI norms land, the operative rule is that the advocate carries the risk: cite it unverified, and it is misconduct.

India angle. For Indian legal-tech, the ruling cuts both ways. Products that draft or research with LLMs now sell into a market where an unverified citation is professional misconduct — which raises the bar for retrieval-grounded, citation-verified tools and effectively ends the market for raw chatbot output in litigation workflows. Vendors that can show verifiable, database-anchored citations (the SCC Online / Manupatra integration layer) get a compliance argument handed to them. For courts and tribunals themselves, the judgment is an uncomfortable precedent: the failure here was judicial, not just professional — an NCLT bench incorporated fabricated authority into its reasoning. Whatever norms the BCI writes for advocates, the tribunal-side verification gap remains open.

Behind the news. Indian institutions are writing AI-failure rules sector by sector rather than waiting for an omnibus AI law. The RBI's June 24 draft model-risk guidance put kill-switch and lifecycle-governance obligations on AI models in regulated finance (covered in the June 27 digest); a week later the Supreme Court is doing the judicial equivalent — naming the failure mode, assigning accountability, and delegating rulemaking to the profession's statutory body.

What to watch. The Bar Council of India's committee: its constitution, composition, and whether the norms it produces require affirmative citation verification at filing or stop at post-hoc discipline. A named, dated BCI notification is the concrete next step the judgment orders.

Source: Bar & Bench, July 2, 2026 → link; The Federal, July 2, 2026 → link.

Confidence: High on the judgment, its date, the zero-tolerance holding, and the BCI direction — corroborated across Bar & Bench, The Federal, and Free Press Journal. Medium on the precise procedural scope: coverage differs on whether the NCLT order, the NCLAT judgment, or both were formally set aside.

---

COMPUTE · INFRASTRUCTURE · July 2, 2026

Lightstorm, Microsoft, Singtel, and Tata Communications sign contracts to build the I-2SEA subsea cable

Lightstorm announced on July 2, 2026 the signing of contracts to build the India–Southeast Asia (I-2SEA) submarine cable system, in a consortium with Microsoft, Singtel, and Tata Communications, with Lightstorm as majority owner. The cable runs approximately 3,600 km from Singapore, connecting Kuala Lumpur, to dual landings on India's east coast at Machilipatnam (Andhra Pradesh) and South Chennai (Tamil Nadu), where Lightstorm operates carrier-neutral landing stations. NEC Corporation is the system supplier and ASEAN Cableship the marine installation partner. Target ready-for-service is Q4 2029. The consortium positions the system as purpose-built for AI, GPU-cluster, and hyperscale cloud traffic on the India–Southeast Asia corridor.

What this means. The signal is in who signed and what stage this is. Contract signing with a named supplier and a named cableship operator is a committed build, not a memorandum — the step at which subsea projects acquire real capital and real delivery dates. Microsoft's presence reads as capacity procurement for its India and Southeast Asia AI build-out; Tata Communications and Singtel bring the carrier layer on both ends. The purpose-built-for-AI framing is positioning language — a fibre pair does not care what workload it carries — but the underlying claim is about route diversity and latency engineering for east-coast Indian data-centre clusters, and the landing choices support it: Machilipatnam is the short path toward Hyderabad's data-centre belt, and a South Chennai landing adds diversity to a congested corridor.

The date that matters is 2029. Nothing about India's connectivity position changes this quarter. What changes now is the committed pipeline — and the fact that a US hyperscaler is underwriting part of an Indian east-coast landing rather than routing everything through Mumbai.

India angle. India's subsea geography has been concentrated at Mumbai, with Chennai and Tuticorin the secondary landings — a known single-region concentration risk for a country adding AI data-centre capacity at multi-gigawatt scale. A new high-capacity system landing at both Machilipatnam and South Chennai rebalances that map eastward and shortens the path between Indian GPU capacity and Singapore's interconnection hub. For Andhra Pradesh, a Machilipatnam landing station stitches the state's data-centre ambitions into international routing. For Indian cloud and AI-infrastructure operators, the practical question is landing-station access terms: Lightstorm's carrier-neutral positioning matters more than the cable's headline capacity if third parties can actually buy in on open terms.

Behind the news. The east-coast rebalancing has been building all year. Avendus's May data-centre report projected three new subsea cables landing in Chennai across 2026–27 (covered in the May 28 digest), and Google's Visakhapatnam campus was announced with its own multi-region subsea landings (covered in the April 27 digest). I-2SEA adds a committed, consortium-financed system to what had been a projection line.

What to watch. The DoT cable-landing licensing process for the Machilipatnam and South Chennai stations, and whether Lightstorm publishes third-party access terms for the landing stations before the marine build starts. Q4 2029 ready-for-service is the delivery date to hold the consortium to.

Source: Lightstorm press release, July 2, 2026 → link; TNGlobal, July 2, 2026 → link.

Confidence: High on the consortium, route, landings, supplier, and RFS target — consistent across the Lightstorm release and trade coverage. Fibre-pair count and design capacity were not disclosed in the coverage reviewed.

---

POLICY · MODEL ACCESS · ENTERPRISE · July 1, 2026

Anthropic restores Fable 5 worldwide after the US lifts its export-control order; Mythos 5 stays US-only

Anthropic restored global access to Claude Fable 5 on July 1, 2026, after the US government lifted the export-control directive that had forced a worldwide suspension of Fable 5 and Mythos 5 on June 12 — a 19-day shutoff of the company's two newest frontier models. CNBC reported the lifting on June 30; Fable 5 returned to the Claude platform, claude.ai, and Claude Code for global users on July 1. Mythos 5 remains restricted to approved US organizations under a separate government clearance granted June 26. Anthropic's own account says testing showed the jailbreak finding that triggered the order — a report that Fable 5 could be prompted to identify software vulnerabilities — was reproducible on many less capable models, and that access returns with additional cybersecurity classifiers layered on.

What this means. The restoration resolves the outage and leaves the precedent fully intact. For 19 days, a deployed frontier model was switched off worldwide by a US government directive — not a weights-export rule but an inference-access order — and it came back not because the mechanism was withdrawn but because the government changed its assessment. Both readings from June still stand. The system worked: a misuse claim got investigated, found overstated, and access returned in under three weeks. And the system now exists: model access has been demonstrated as an instrument of US policy, applied and released on Washington's timetable. The Mythos 5 asymmetry sharpens the second reading — the most capable tier returns as a US-only clearance regime, which is a new shape for frontier-model geography.

India angle. India, Anthropic's self-described second-largest market, gets its access back — and keeps the lesson. The SI layer and the Indian product startups that spent June migrating agentic workloads off Claude now face the reverse question: migrate back, or bank the diversification they were forced into. The rational answer for most is both — restore Claude where it was strongest, keep the fallback path warm — which means the shutoff's durable effect on India is an ecosystem that architects for model substitution as a matter of course. For BFSI and public-sector deployers, the episode hardened the self-hosting argument in a way no vendor assurance can now undo: an open-weights model on Indian GPUs cannot be switched off by a foreign directive. That argument was made loudly during the June debate, and the restoration does not answer it.

Behind the news. This closes the arc opened on June 12, when the Commerce Department's directive forced Anthropic to disable both models three days after launch (covered in the June 14 digest). That digest flagged restoration terms as the signal to watch — a quiet reinstatement versus a carve-out regime. The answer is now visible and split: Fable 5 came back globally without ID-verification conditions; Mythos 5 came back as a US-organizations-only clearance tier.

What to watch. Whether Mythos 5 access extends beyond approved US organizations — and on what verification terms — or whether the two-tier geography holds. That is now the cleanest single indicator of whether frontier-model access is renormalizing or restratifying.

Source: Anthropic, "Redeploying Claude Fable 5," July 1, 2026 → link; CNBC, June 30, 2026 → link; Forbes, July 1, 2026 → link.

Confidence: High on the lifting of controls, the July 1 global restoration of Fable 5, and the US-only status of Mythos 5 — corroborated across Anthropic's announcement, CNBC, Forbes, and Al Jazeera. Medium on the interim usage terms (plan-level inclusion windows and post-restoration credit mechanics), which rest on secondary summaries of Anthropic's announcement.

---

MODEL RELEASE · PRICING · AGENTS · June 30, 2026

Anthropic ships Claude Sonnet 5 at an introductory $2/$10, pushing near-Opus agentic capability down the price curve

Anthropic released Claude Sonnet 5 on June 30, 2026, positioning it as its most agentic Sonnet-tier model — able to plan and run autonomously against tools like browsers and terminals at a level Anthropic says approaches Opus 4.8. Introductory API pricing is $2 per million input tokens and $10 per million output through August 31, 2026, after which it moves to the standard Sonnet price of $3/$15. The model is the new default for Free and Pro plans. Anthropic reports a lower rate of undesirable behaviors than Sonnet 4.6 in agentic contexts.

What this means. The Sonnet tier has held its $3/$15 sticker since the original Claude 3 launch in March 2024, while the capability shipped at that price has climbed tier over tier — the June 2024 Claude 3.5 Sonnet release delivered then-Opus-class capability at Sonnet pricing, and Sonnet 5 repeats the move against Opus 4.8. The introductory $2/$10 is the new element: a time-boxed discount on a frontier-lab mid-tier reads as a volume play against the open-weight price floor, where models like Moonshot's Kimi K2.7-Code have been shipping agentic coding capability at $4 per million output tokens against Fable 5's $50 (the comparison drawn in coverage of that release, covered in the June 12 digest). Anthropic shipping its cheapest-ever near-frontier tier the day before Fable 5's restoration also says something about the company's June: the export-control crisis did not pause the release cadence.

India angle. Agentic workloads are where Indian deployment economics have been tightest — long tool-use loops multiply output tokens, and output is where frontier pricing bites. Near-Opus capability at $10 output (then $15) changes the per-task math for the SI layer pricing agentic delivery into client SOWs, and for Indian product startups whose agent features were marginal at Opus-tier prices. The two-month introductory window is effectively a paid trial at scale: Indian teams that benchmark Sonnet 5 against their open-weights fallbacks before August 31 will know exactly what the $3/$15 reversion is worth to them.

What to watch. August 31, when introductory pricing lapses — whether Anthropic converts the discount into a permanent price move under open-weight pressure, or Sonnet 5 usage holds at $3/$15.

Source: Anthropic announcement, June 30, 2026 → link; TechCrunch, June 30, 2026 → link.

Confidence: High on the release date, pricing, and plan positioning — corroborated across Anthropic's announcement, TechCrunch, and VentureBeat. The near-Opus-4.8 capability claim is Anthropic's own, pending independent evaluation.

---

Position movements