India AI DigestJune 8, 2026
India AI Digest — Monday, June 8, 2026
- Anthropic ships Claude Fable 5, its first public "Mythos-class" model, at 95% on SWE-bench Verified and $10/$50 per million tokens — a coding-and-agentic step up that resets what the Indian services layer benchmarks against.
- The Supreme Court of India released draft regulations for AI use in courts, mandating lawyer disclosure of AI in pleadings and barring outcomes decided solely by algorithm. Public comments are open to June 20.
- Google's Gemini 3.5 Pro is days from general availability — 2M-token context, Deep Think reasoning, and a roughly $15/$60 price that sharpens the unit-economics squeeze on Indian consumer AI.
Position movements: regulatory_clarity +1 (India, judicial AI); foundation_model_capability +1 (global frontier).
MODEL · FRONTIER · CODING · June 9, 2026
Anthropic releases Claude Fable 5, a Mythos-class model topping coding benchmarks
Anthropic released Claude Fable 5 on June 9, 2026 — the first publicly available model in its "Mythos" tier, positioned a step above Opus 4.8. Reported pricing is $10 per million input tokens and $50 per million output, with a 1M-token context window and up to 128K output tokens per request. The headline numbers are on code: 95% on SWE-bench Verified and 80.3% on SWE-bench Pro, the latter reported as roughly 11 points ahead of the next model. Fable 5 shares weights with the access-restricted Claude Mythos 5, with safety classifiers that reroute high-risk queries to Opus 4.8. It is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost through June 22; from June 23 it moves behind usage credits.
What this means. The interesting figure is SWE-bench Pro, not SWE-bench Verified. Verified saturates — once models cluster in the 90s, the benchmark stops discriminating. SWE-bench Pro is the harder, multi-file, longer-horizon set, and an 11-point gap there is the part that survives scrutiny. If it holds under independent runs, the claim is narrower and stronger than "best coding model": Fable 5 sustains long, autonomous engineering work that prior models dropped partway through.
The pricing tells its own story. At $10/$50, Fable 5 is a premium tier — a model you point at the task that justifies the spend, not the default for routine calls. The weight-sharing detail with the restricted Mythos 5 is the structural note: the same capability ships in two forms, one gated by safety routing, one not. That is a deployment-policy choice as much as a capability release.
India angle. This lands hardest on the services layer. TCS, Infosys, and Wipro each scaled Microsoft 365 Copilot past 100,000 seats by early June, and the economic exposure of Indian IT concentrates in exactly the engineering and delivery work a sustained coding model rebundles. A model that holds context across a multi-file change without losing the thread changes what "AI-assisted delivery" means in client conversations — it raises the baseline a proof-of-concept has to beat, and it does so at a price point that only closes on high-value enterprise work, not commodity ticket volume.
For BFSI and healthcare deployers under DPDP cross-border constraints, Fable 5 is API-only via Anthropic's regions; the residency question that gates regulated workloads from foreign frontier models is unchanged here. The cost-capability frontier moved up. The deployability wall for India's regulated sectors did not move with it.
Behind the news. Fable 5 arrives about two weeks after Opus 4.8 (released May 27, 2026) and a week after Microsoft's in-house MAI model family (June 2). The frontier is releasing on a compressed cadence, and the competition has shifted decisively toward long-horizon coding and agentic work rather than chat quality.
What to watch. Independent SWE-bench Pro reproductions over the next few weeks. Vendor-reported benchmarks set the expectation; third-party runs on held-out tasks set the reality. Watch whether the 11-point gap survives contact with evaluators who didn't build the model.
Source: Anthropic Claude Fable 5 release coverage, June 9, 2026 — corroborated across multiple outlets reporting price, context window, and SWE-bench figures. → buildfastwithai · → computingforgeeks
Confidence: Medium — benchmark and pricing figures are consistent across secondary coverage; independent reproduction of SWE-bench Pro pending.
POLICY · GOVERNANCE · JUDICIARY · June 3, 2026
Supreme Court of India publishes draft AI-in-courts regulations, opens public consultation
The Supreme Court of India issued the draft Regulations for Use of Artificial Intelligence in Courts, 2026 on June 3, 2026, and opened them for public comment through June 20. The draft applies to the use of AI in any judicial, adjudicatory, or administrative function across the Supreme Court, High Courts, and all courts, tribunals, and statutory commissions performing adjudicatory functions. Two provisions stand out: judicial outcomes — or any part of them — may not rest solely on algorithmic decision-making or AI-generated information, and lawyers must disclose AI use in their pleadings. The framework is built on stated principles of human primacy, transparency, accountability, data protection, and judicial independence, and is expressly tied to the DPDP Act, 2023 and the DPDP Rules, 2025 (notified November 13, 2025). It proposes an institutional scaffold: an apex body, committees spanning technical, judicial, infrastructure, cybersecurity, and data-management functions, plus an AI register, an incident database, a content-verification authority, and a grievance-redressal process.
What this means. This is the most concrete piece of Indian AI governance to land on a specific high-stakes domain, rather than the horizontal "do no harm" framing that has characterized MeitY's guidance. The disclosure-in-pleadings requirement is the operationally sharp part. It converts a behavioral norm — lawyers already use AI for drafting and research — into a procedural obligation with a paper trail, which is enforceable in a way that aspirational principles are not.
The human-primacy clause is doing real work and worth reading carefully. Prohibiting outcomes based solely on algorithmic decision-making leaves the assistive-versus-determinative line where it belongs, but the boundary is where the litigation will be. An AI tool that drafts a judgment a judge signs, a tool that surfaces precedent, and a tool that scores bail risk sit at very different points on that line, and "solely" will be argued over case by case. The register and incident database are the part that, if actually maintained, would give the framework teeth — a public record of which tools are deployed and where they failed.
India angle. This is a direct signal to the Indian legal-tech layer — the document-review, e-discovery, and case-research vendors building for courts and law firms. A disclosure regime plus an AI register changes procurement: tools that can produce auditable provenance and integrate with the content-verification authority have an advantage over black-box offerings. The explicit DPDP linkage compounds it — judicial data is among the most sensitive categories, and any vendor routing court data through foreign inference now sits at the intersection of two tightening frameworks. For the broader stack, the move matters as precedent: a constitutional institution writing AI governance for itself sets a template other regulated domains will reference.
Behind the news. This builds on the Supreme Court's December 2025 assurance that judges would exercise caution on AI in the judicial process, and slots above the DPDP Rules notified in November 2025. It is a draft, not law — the consultation runs to June 20, and the final form will turn on the comments received.
What to watch. The June 20 consultation close, and specifically whether the bar associations and legal-tech vendors push back on the disclosure-in-pleadings mandate and the scope of the "solely algorithmic" prohibition. Comments go to the AI committee's member secretary at office.regcc@sci.nic.in.
Source: Supreme Court of India notice dated 03.06.2026. → SC notice (PDF) · → LiveLaw
Confidence: High — primary source is the Supreme Court's own dated notice, corroborated by multiple legal-news outlets.
MODEL · FRONTIER · PRICING · June 6, 2026
Gemini 3.5 Pro nears general availability with 2M-token context — and a higher price
Google's Gemini 3.5 Pro, unveiled at Google I/O on May 19, 2026, is reported to be days from general availability as of June 6, still in limited Vertex preview. The stated specifications are a 2-million-token context window, a Deep Think reasoning mode, and frontier multimodal handling. Reported pricing is roughly $15 per million input tokens and $60 per million output — a step up from the Flash tier that shipped at I/O. This is a forthcoming release, not a shipped one: treat the specifics as Google's stated intent until GA.
From the room.
"Give us until next month to get it to you." — Sundar Pichai on Gemini 3.5 Pro, Google I/O, May 19, 2026
What this means. Two things are moving in opposite directions, and that tension is the story. Capability is climbing — a 2M-token window and a dedicated reasoning mode push into work that needs whole-codebase or whole-corpus context. Price is climbing with it. The cheaper, faster Gemini 3.5 Flash is what agent platforms are actually wiring in — Salesforce is putting Flash inside Agentforce in its June 15 release precisely because high-volume agentic calls need low per-token cost. Pro is the capability ceiling; Flash is the deployment default. That split is now the shape of the whole frontier market, not a Google quirk.
India angle. For Indian consumer and Indic-language applications running at ARPU ceilings around ₹80–200 a month, a $15/$60 model is not the inference tier the unit economics support — it is the model you route to selectively, for the query that justifies it, with a cheaper model handling the volume. The rising top-end price reinforces the architecture Indian builders already run: tiered routing, aggressive caching, and a Flash-class or open-weights model carrying the bulk of traffic. The capability is welcome; the path to using it affordably runs through everything except calling Pro on every request.
What to watch. The actual GA date and whether the $15/$60 pricing holds at launch — preview pricing and GA pricing have diverged before. Also watch whether Google ships an India-region inference option, which would change the BFSI and healthcare deployability calculus that currently keeps regulated Indian workloads off foreign frontier APIs.
Source: Google I/O (May 19, 2026) and pre-GA coverage dated June 6, 2026. → TechTimes · → Google blog
Confidence: Medium — Pichai quote and I/O unveiling are firm; GA date, final pricing, and specs are pre-release and subject to change.
