AI June 2026: What Shipped in Week 3 and What Slipped

The AI frontier moved fast in the third week of June 2026, but not always in the directions the calendar promised. Gemini 3.5 Pro missed its own deadline — again. Claude Sonnet 4.8 never existed outside speculation. And Grok 5 is still training. What actually shipped is worth your attention precisely because the signal-to-hype ratio is finally manageable: three concrete developments, real numbers, and a platform shift at Apple that changes who controls AI distribution on a billion devices.

What Actually Shipped — and What Didn't

The table below sets the record straight before anything else.

Lab	Model	Release date	Spec headline	Dev access	Pricing (input / output per 1M tokens)
Anthropic	Claude Fable 5	June 9, 2026	80.3% SWE-Bench Pro, 1M ctx / 128K output	API GA (claude-fable-5)	$10 / $50
Anthropic	Claude Mythos 5	June 9, 2026	Same weights as Fable 5, safety classifiers lifted in specific domains	Restricted API	$10 / $50
OpenAI	GPT-5.5 Instant	May 5 → June default	1M+ ctx, 128K output, hallucination reduction in law/med/finance	API GA (chat-latest)	$5 / $30 (standard); $10 / $45 long-ctx
Google	Gemini 3.5 Flash	May 19, 2026	1M ctx, 65K output, 4× faster than 3.1 Pro	API GA	$1.50 / $9.00
Google	Gemini 3.5 Pro	Still expected	2M ctx, Deep Think reasoning	Limited Vertex enterprise preview only	Not yet published
xAI	Grok 5	Still expected	~6T-param MoE, training on Colossus 2	Not available	Not yet published

Claude Sonnet 4.8 — despite showing up in leaked roadmaps — has not shipped. Anthropic's Sonnet line remains at 4.6. The company moved its frontier tier naming entirely, going from Opus-as-ceiling to the Mythos class that debuted with Fable 5 and Mythos 5. You can stop waiting for a Sonnet 4.8 announcement.

Gemini 3.5 Pro: At Google I/O on May 19, Sundar Pichai asked developers to "give us until next month" — prompting an audible groan in the room. That next month is now more than half over and Pro remains in limited Vertex enterprise preview with no public model card, no published benchmarks, and no API pricing. It is still expected, but the window is closing.

Grok 5: xAI missed its Q1 target, then its Q2 target. Polymarket had it at roughly one-in-three odds of shipping by June 30. The model is reportedly a Mixture-of-Experts architecture with around 6 trillion parameters, training on Colossus 2. Meanwhile xAI has been iterating on the Grok 4 series — the 4.3 beta shipped in April — and shipped Grok Voice (June 4) and Grok Imagine Video 1.5. Grok 5 is real work in progress, not vaporware, but it is not here yet.

Claude Fable 5 and Mythos 5: The Code Benchmark to Take Seriously

Anthropic's June 9 launch introduced a tier they are calling Mythos-class — positioned above Opus. Two models share the same weights: Fable 5 (public API, with safety classifiers that reroute high-risk sessions to Opus 4.8) and Mythos 5 (restricted access, classifiers partially lifted for specific research domains).

The benchmark that matters for working developers is SWE-Bench Pro: Fable 5 posts 80.3%. By comparison, Opus 4.8 sits at 69.2%, GPT-5.5 at 58.6%, and Gemini 3.1 Pro at 54.2% on the same evaluation. Fable 5 also takes the top position on FrontierCode and posts a 1932 Elo on GDPval-AA. Anthropic's framing — "the longer and more complex the task, the larger Fable 5's lead" — is consistent with what the numbers show, but independent third-party evaluations are still sparse at this writing.

Pricing is $10 / $50 per million tokens, exactly double Opus 4.8's rates. That is not cheap for production use at scale. The June 22 deadline matters: Pro, Max, Team, and Enterprise seat plans got Fable 5 included at no extra cost through that date. On June 23 it moves to usage credits. If you have active workflows you want to test, the next few days are the window.

Context: 1M input tokens, 128K output — matching Opus 4.8's ceiling. For the truly long-context use cases Anthropic keeps teasing (multi-repo refactors, legal discovery over document sets), the model is ready in principle; the practical question is cost per session.

API access is straightforward via the claude-fable-5 model ID. No waitlist.

GPT-5.5: Calibration Over Capability

OpenAI's headline move this week was less a launch than a consolidation. GPT-5.5 Instant became the default model for ChatGPT across free and paid tiers, replacing GPT-5.2 entirely (that model retired June 12). The API identifier is chat-latest.

The pitch from OpenAI is calibration: fewer bullet-heavy responses, reduced hallucination in law, medicine, and finance, stronger use of shared context across sessions. The 1M+ context window was already present in GPT-5.5's April API release; the June work is about polish rather than new capability.

Pricing: $5 / $30 per million tokens (standard), dropping to $2.50 / $15 in Batch/Flex mode. Long-context pricing above 200K tokens steps up to $10 input, $45 output. The gap versus Fable 5 on input tokens is real ($5 vs $10), but GPT-5.5 trails on the SWE-Bench Pro score by about 22 points.

Memory improvements — pulling context from past chats, files, and connected Gmail — are rolling out to Plus and Pro users this week, with Free and Go tiers to follow.

Gemini 3.5 Flash: The Quiet Workhorse

Gemini 3.5 Flash shipped GA at Google I/O (May 19) and has been quietly gaining adoption without much additional press noise this week. For developers who missed the initial noise: it runs at $1.50 / $9.00 per million tokens, handles 1M input tokens, and outperforms Gemini 3.1 Pro on coding and agentic benchmarks (Terminal-Bench 2.1: 76.2% vs 70.3%; MCP Atlas: 83.6% vs 78.2%) while running 4× faster.

The practical positioning: if your workload does not require the raw coding ceiling of Fable 5 and you are price-sensitive, Flash is the most capable model at this price point from any of the three major labs. The catch is that Pro — which Pichai positioned as Google's frontier answer to Opus and GPT-5.5 — is still not available. Developers who need a Gemini equivalent to what Fable 5 offers are still waiting.

Apple WWDC and the Distribution Shift

At WWDC on June 8, Apple announced an Extensions framework for iOS 27, iPadOS 27, and macOS 27 that lets users set Claude, Gemini, or ChatGPT as the AI backbone for Siri — not just as external handoff targets, but as replacements for Apple Intelligence features including Writing Tools and Image Playground. Users select their preferred provider in Settings → Apple Intelligence & Siri.

This is strategically significant in a way that benchmark point-scoring is not. Until now, OpenAI had privileged access to the iOS surface via the existing ChatGPT integration. The Extensions model puts all three major providers on equal footing at the OS level. It also turns every Siri interaction into a potential switching moment — meaning API usage data from mobile will start diverging from web/desktop patterns in ways worth watching.

For Indian developers: all three providers will need apps on the Indian App Store to be selectable in Siri Extensions. Claude is available; Gemini is available; ChatGPT is available. The Extensions system ships with iOS 27 in September, not today, but the API-level implications for routing mobile queries are upstream of that.

India Access Snapshot

This is the section most India-based developers actually need.

Google (Gemini): The only Tier-1 AI service that accepts UPI, bills natively in INR, and requires no international card. The Gemini Developer API (AI Studio) is accessible from India with INR conversion at approximately ₹95 per dollar as of mid-June. Vertex AI India regional endpoint pricing comes into effect July 1, 2026 — before that date, non-global endpoints default to global pricing. Gemini 3.5 Flash's ₹142 / ₹855 per MTok (at that rate) is the most accessible entry point for bootstrapped Indian teams.

Anthropic: The Bengaluru office opened in early 2026 with Irina Ghose (former Microsoft India head) as MD. India is now Anthropic's second-largest market after the US. The API is accessible from India, but all billing is in USD — no INR, no UPI support directly from Anthropic. The workaround is AWS Bedrock or Google Cloud (which does support INR billing) for model access that avoids direct dollar payment. Anthropic is working with Karya and the Collective Intelligence Project on Indic language evaluations, which suggests future model improvements will include better Hindi, Tamil, Telugu, and regional language performance — no timeline yet.

OpenAI: API billed in USD. Third-party gateways like AICredits accept INR via Razorpay/UPI for API access. No Indian data residency endpoint as of this writing, though regional processing is available with a 10% pricing uplift for eligible models.

What to Watch

• Gemini 3.5 Pro GA: The remaining days of June are the make-or-break window. If it slips to July, Google will have spent six weeks building expectations it could not meet. A model card with real benchmarks is the minimum needed to evaluate whether it competes with Fable 5 on coding tasks.
• Grok 5: Every week past the Q2 target narrows xAI's window before Google and Anthropic lock in the developer tooling ecosystem for the second half of 2026. The Colossus 2 training run is at scale; the question is whether the eval numbers justify the wait.
• Fable 5 independent evals: Anthropic's own benchmarks show an 11-point gap over Opus 4.8 on SWE-Bench Pro. Third-party replication on real production repos — not curated eval sets — will determine whether the $10/$50 price point is justified for teams currently on Opus or Sonnet 4.6.
• iOS 27 Extensions monetization model: Apple has not disclosed how revenue sharing or attribution will work when a user sets Claude as their default AI in Siri. That commercial structure will matter for how aggressively Anthropic and Google invest in the iOS surface post-September.
• INR billing parity: Anthropic has India as its second-largest market but still bills in USD. Either they extend INR billing before the year end or Indian developers consolidate on Gemini Flash for cost-sensitive production workloads. Watch for an announcement tied to the Bengaluru office's first major milestone.