Mervin Praison

At WWDC26 on 8 June 2026, Apple previewed Siri AI and the next generation of Apple Intelligence on iOS 27, iPadOS 27, and macOS 27—powered by Apple Foundation Models built with Google Gemini and split across on-device Apple silicon and Private Cloud Compute.

What changed

• Siri AI replaces the legacy assistant with personal-context search across Messages, Mail, and Photos; on-screen and Camera-mode awareness; expanded systemwide app actions; web-grounded answers; and a dedicated Siri app with iCloud-private conversation sync across iPhone, iPad, Mac, Watch, and Vision Pro.
• Invocation surfaces expand beyond “Hey Siri” to Dynamic Island swipe (iPhone), Spotlight (iPad/Mac), control-click context menus, and Vision Pro look-to-speak with 3D visualisation.
• On-device plumbing includes a system orchestrator, Spotlight index, and App Toolbox that keep personal-context processing local before escalating frontier workloads.
• Apple Foundation Models are custom-built in collaboration with Google Gemini for deeply integrated experiences—not exposed as a raw Gemini API to consumers per Apple’s Intelligence announcement.
• Hybrid execution runs models on device and on Private Cloud Compute; PCC retains Apple’s no-storage privacy promise with ongoing external verification.
• Image Playground adds photorealistic generation on PCC with hidden SynthID watermarks; Photos gains Spatial Reframing and other on-device intelligence features.
• Developer betas for Siri AI ship 8 June 2026 on iOS, iPadOS, macOS, and visionOS; watchOS follows in a future beta.

Developer integration surface

Foundation Models framework (Swift) is the primary LLM integration path: on-device sessions, Private Cloud Compute for frontier tasks, tool calling, Dynamic Profiles for multi-model routing, and third-party models via the Language Model protocol (Gemini, Claude, and others). Apple plans to open-source the framework core later in summer 2026. Use it when you want Apple-hosted intelligence inside your app without managing API keys or PCC authentication.

Core AI is a separate stack for deploying custom PyTorch models on Apple silicon—Python conversion tools, ahead-of-time compilation in Xcode, Swift inference APIs, and Core AI debugging instruments. Use Core AI when you bring your own weights; use Foundation Models when you consume Apple’s Foundation Models or attach approved third-party LLM providers.

App Intents and Spotlight integrations extend Siri AI personal context to third-party apps. View Annotations and on-screen-awareness APIs let apps participate in Siri’s screen-context flows without exposing raw screenshots to external model vendors.

Why it matters for engineers

Apple’s WWDC26 stack is a platform inference architecture, not a single model API. Builders should plan for dual execution paths: on-device Foundation Models for latency- and privacy-sensitive personal context, and PCC for frontier workloads (photorealistic image generation, broad world knowledge) with quota limits. This article covers the consumer Siri AI and developer framework launch; it is distinct from Apple’s PCC infrastructure expansion on Google Cloud NVIDIA hardware, which focused on attestation, fleet ledgers, and confidential-GPU hosting rather than Siri UX and App Intents.

Feature-detect against two hardware tiers before shipping voice or dictation features: the base Apple Intelligence list (iPhone 16+, M1+ Mac/iPad) differs from the advanced on-device model tier (M4+/M3+ with ≥12GB unified memory, iPhone 17 Pro family) required for expressive voices and advanced dictation.

Server-model daily caps and iCloud+ entitlements mean client apps must degrade gracefully when users exhaust allotments—Apple has not published numeric quotas, but photorealistic Image Playground and similar PCC-backed features are explicitly rate-limited. Enterprise Mac teams should plan fall GA as a coordinated OS 27 rollout with regional gates: EU iOS/iPadOS Siri AI is deferred whilst Mac and Vision Pro proceed.

For teams comparing hyperscaler assistants: Apple exposes no raw Gemini or Claude endpoint. Capabilities arrive through Foundation Models framework sessions and Siri AI system channels—simplifying privacy review but limiting custom prompt engineering relative to direct API integrations.

Personal-context Siri workloads stay on Apple silicon; frontier models run in Private Cloud Compute without storing user prompts.

Intelligence routing at WWDC26

flowchart TB
  USER["User or app request"]
  LOCAL["On-device Foundation Models"]
  PCC["Private Cloud Compute"]
  ANS["Response to user"]
  USER --> LOCAL
  LOCAL -->|"personal context"| ANS
  LOCAL -->|"frontier workload"| PCC
  PCC --> ANS

Anthropic shipped Claude Fable 5 on 9 June 2026—a Mythos-class frontier model for general use with classifier fallbacks to Claude Opus 4.8 on sensitive cyber, biology, and distillation queries—alongside restricted Claude Mythos 5 access for Project Glasswing defenders and separate biology trusted-access programmes.

Short video walkthrough

Engineering walkthrough — ElevenLabs narration, HeyGen bookends, API vs claude.ai defaults, and official Anthropic B-roll (~6 min).

What changed

• Claude Fable 5 is Anthropic’s first Mythos-class model generally available, with state-of-the-art scores on software engineering, knowledge work, vision, and long-horizon agent benchmarks—lead grows as tasks become longer and more complex per the launch post.
• New safety classifiers extend constitutional-classifier work: cyber (exploitation plus offensive agentic hacking), biology/chemistry (broad fallback during launch), and distillation (large-scale capability extraction) all route flagged prompts to Claude Opus 4.8 instead of refusals.
• Claude Mythos 5 shares Fable 5 weights with cyber safeguards lifted for existing Project Glasswing partners upgrading from Mythos Preview; comparable or stronger performance at substantially lower cost.
• Biology trusted access (separate from Mythos 5) will offer Fable 5 with bio/chem classifiers removed but cyber classifiers still active to a small life-sciences cohort—broader enrolment planned as safeguards narrow.
• Pricing halved versus Mythos Preview on API and consumption-based Enterprise plans.
• 30-day retention is required for Mythos-class business traffic to detect novel jailbreaks; data deleted after 30 days with logged human access (Anthropic support article).
• Red-team validation: external bug bounty reported no universal jailbreak in 1,000+ hours; zero compliance on harmful single-turn cyber requests across 30 public jailbreak techniques in partner testing.
• Subscription rollout is demand-sensitive: included at no extra cost on paid Claude plans through 22 June 2026, then usage credits until capacity stabilises.

Capability evidence for builders

• Software engineering: Stripe reported a 50-million-line Ruby migration in one day (versus an estimated two-plus months manually); Cognition’s FrontierCode ranks Fable 5 highest among frontier models at medium effort with improved token efficiency.
• Knowledge work: highest score on Hebbia’s Finance Benchmark; IMC reported near-perfect trading-analysis results across factual lookup, root-cause analysis, and expected-value reasoning.
• Vision: state-of-the-art on vision tasks; completed Pokémon FireRed vision-only without navigation harnesses that prior Claude models required.
• Memory: on Slay the Spire agent runs, file-based memory produced threefold improvement versus Opus 4.8 and threefold higher final-act completion rates.
• Alignment: automated assessments place Mythos 5 misaligned behaviour similar to Opus 4.8 per the system card.

Why it matters for engineers

Teams wiring production agents must treat Fable 5 as a two-model endpoint: more than 95% of sessions never trigger fallback, but cyber-hardening, bioinformatics, or suspicious bulk-extraction patterns transparently downgrade to Opus 4.8 with user notification. Log response metadata and surface fallback events to operators—latency and capability profiles differ, and conservative classifier tuning means benign security research queries can still trip safeguards during the launch window.

The API and consumption-based Enterprise path is the reliable integration surface from day one. Subscription inclusion is time-boxed and demand-sensitive; capacity planning for long autonomous coding runs should prefer metered API tiers. Mythos 5 remains outside general API access—cyber defenders need Glasswing or a future trusted-access application; biology researchers follow the separate Fable-without-bio-classifiers programme.

Long-context and file-backed memory improvements matter for multi-hour agent loops: Fable 5 sustains focus across millions of tokens and benefits disproportionately from persistent notes versus Opus 4.8. Vision-only harnesses now complete screenshot-to-code and scientific-figure extraction tasks that previously required scaffolding.

Regulated workloads must account for 30-day Mythos-class retention on business traffic, logged human access to stored prompts, and explicit prohibition on training use. Benchmark harnesses that resemble distillation attacks may trigger classifiers—design eval pipelines to tolerate Opus 4.8 fallbacks or isolate test traffic from production API keys.

Most Fable 5 sessions run at full frontier capability; cyber, biology, and distillation classifiers route sensitive prompts to Opus 4.8 instead of blocking.

Classifier fallback in production

flowchart LR
  REQ["Agent or app request"]
  CLS["Safety classifiers"]
  FABLE["Fable 5 response"]
  OPUS["Opus 4.8 fallback"]
  OUT["Answer delivered"]
  REQ --> CLS
  CLS -->|"typical workload"| FABLE
  CLS -->|"cyber bio distillation"| OPUS
  FABLE --> OUT
  OPUS --> OUT

Anthropic doubled Claude Cowork’s five-hour session rate limits for Pro, Max, and Team subscribers from 5 June through 5 July 2026, leaving weekly caps and the shared quota across Claude products unchanged.

What changed

• Boris Cherny, who leads Claude Code at Anthropic, announced the promotion on 5 June 2026 via social post—no dedicated article appeared on the Anthropic newsroom index by 9 June 2026.
• Claude Cowork five-hour rolling session limits are doubled for approximately one month, ending 5 July 2026.
• Eligible plans: Claude Pro, Claude Max, and Claude Team.
• The change applies to five-hour rate-limit windows only—Anthropic’s weekly usage cap is unchanged.
• Claude Code and Claude.ai retain standard session limits; the promotion is Cowork-specific.
• Subscription quota remains a shared pool across Claude surfaces—heavier Cowork bursts can still exhaust the weekly budget faster.

Why it matters for engineers

Anthropic meters paid plans with two leaky buckets: a five-hour rolling session window for burst fairness and a weekly cap for cost control. Doubling only the first bucket optimises long desktop agent runs—folder reorganisation, batch report generation, scheduled digests—without raising Anthropic’s weekly compute exposure. Teams scheduling Cowork jobs should treat the promotion as session headroom, not unlimited capacity.

Cowork is not the Claude API. It runs in the desktop app with filesystem and Office integration, autonomous loops, and user approval gates—ideal for knowledge-worker delegation, unsuitable for production services. Engineers should keep CI and production agents on API metering while pilots use Cowork inside the promo window for deferred “messy folder” projects Cherny highlighted.

Unified quota across Cowork, Claude Code, and web chat means platform leads need allocation policy. A seat running heavy Code sessions the same week as a doubled Cowork migration may hit the unchanged weekly ceiling before the session window resets. Monitor Settings → Usage for both progress bars before kicking off multi-hour agent tasks.

Enterprise admins already manage Cowork feature access and org spend caps separately from consumer tiers. Communicate the 5 July revert date so programme managers do not assume permanent 2× session limits in capacity plans.

Anthropic doubled the five-hour Cowork usage bucket for eligible paid plans from 5 June through 5 July 2026 whilst leaving weekly caps unchanged.

Limit windows over the promotion

flowchart TB
  START["5 Jun 2026 promo starts"]
  SESSION["Five-hour rolling window resets continuously"]
  DOUBLE["Cowork session allowance 2x"]
  WEEKLY["Weekly cap unchanged"]
  SHARED["Shared pool: Cowork chat and Code"]
  END["5 Jul 2026 promo ends"]
  START --> DOUBLE
  DOUBLE --> SESSION
  SESSION --> SHARED
  SHARED --> WEEKLY
  WEEKLY --> END
  classDef agent fill:#8B0000,color:#fff
  classDef tool fill:#189AB4,color:#fff
  class DOUBLE agent
  class WEEKLY tool

Timeline view: session windows roll continuously and temporarily widen for Cowork; the weekly ceiling and cross-product pool stay fixed.

Microsoft’s 2026 Work Trend Index gives engineering leaders a vocabulary for human–agent collaboration and ships Copilot Cowork mobile, plugins, and Agent 365 so Frontier Firms can orchestrate work across Microsoft and third-party systems.

What changed

• Microsoft named four collaboration patterns—Author, Editor, Director, and Orchestrator—and argued leaders must match workstreams to the right pattern rather than defaulting every process to multi-agent orchestration.
• The 2026 Work Trend Index analysed trillions of anonymised Microsoft 365 signals and surveyed 20,000 AI-using knowledge workers across ten countries (February–April 2026).
• 49% of sampled Copilot chats support cognitive work; 58% of AI users produce work they could not a year ago, rising to 80% among Frontier Professionals.
• Microsoft described a Transformation Paradox: 65% fear falling behind without AI, yet 45% prefer current goals over redesigning work, and only 13% feel rewarded for AI-driven reinvention.
• Organisational factors—culture, manager support, talent practices—account for more than twice the reported AI impact of individual mindset (67% vs 32%).
• Respondents map to five readiness zones: Frontier (19%), Blocked Agency (10%), Unclaimed Capacity (5%), Stalled (16%), and Emergent (50%).
• Copilot Cowork Mobile launched on iOS and Android; native plugins for Dynamics 365 and Fabric are GA, with partner plugins (LSEG, Miro, monday.com, S&P Global Energy) rolling out.
• Custom plugins let organisations codify internal workflows; federated Copilot connectors are GA in Researcher and Microsoft 365 Copilot Chat.
• Microsoft Agent 365 is the control plane for governing, observing, and securing agents at scale, including visibility into local agents.

Why it matters for engineers

Platform teams often ship agents without changing incentives. The WTI data suggests most adoption friction is organisational, not model quality—skilled builders frequently land in Blocked Agency zones where legacy metrics punish workflow redesign. Pair agent rollouts with evaluation criteria that reward reinvention, not only throughput.

The four-pattern ladder is a practical safety taxonomy. Author and Editor modes suit low blast-radius tasks with human review on every artefact. Director mode needs job isolation, rollback, and audit trails. Orchestrator mode demands a control plane—Agent 365 in Microsoft’s stack—for connector scopes, identity, and exception routing. The same framing applies whether you build on Copilot or run Claude Code beside it.

Cowork’s plugin and connector model is the integration surface to design for: native first-party data (Fabric, Dynamics), packaged partner actions, and custom plugins for proprietary expertise. Federated connectors let agents read external knowledge without migrating data. That graph-of-connectors pattern is portable beyond M365.

Frontier Professionals—multi-step agent users who redesign workflows and publish team standards—are a benchmark for internal playbooks. They pause to allocate human versus AI work, deliberately practise skills without AI, and treat model output as draft material. Telemetry showing 49% of Copilot use in cognitive tasks suggests backlog priority belongs in analysis and synthesis features, not generic chat wrappers.

Frontier Firms redesign work around human–agent teams: people set goals and own accountability whilst agents execute repeatable analysis and orchestration.

Readiness zones at a glance

flowchart LR
  subgraph lowOrg["Low organisational readiness"]
    ST["Stalled 16%"]
    EM["Emergent 50%"]
  end
  subgraph highOrg["High organisational readiness"]
    UC["Unclaimed capacity 5%"]
    FR["Frontier 19%"]
  end
  subgraph indiv["Individual capability"]
    LO["Low"]
    HI["High"]
  end
  BA["Blocked agency 10%"]
  HI --> BA
  BA --> lowOrg
  FR --> highOrg
  HI --> FR
  LO --> ST
  classDef agent fill:#8B0000,color:#fff
  classDef tool fill:#189AB4,color:#fff
  class FR agent
  class BA tool

Matrix view: Frontier sits where individual skill and organisational support reinforce each other; Blocked Agency is the engineering-heavy zone where talent outruns incentives.

Apple is extending Private Cloud Compute to Google Cloud NVIDIA GPU clusters so the heaviest Apple Intelligence workloads can run on third-party infrastructure without abandoning stateless, attestable privacy guarantees.

What changed

• PCC leaves Apple-only data centres. For the first time, Apple Intelligence cloud inference runs on Google Cloud systems, whilst Apple retains cryptographic control over which PCC software builds devices will trust.
• New hardware trust stack. The implementation combines NVIDIA Confidential Computing GPUs, Intel CPUs with Trust Domain Extensions (TDX), and Google’s Titan security chip — replacing the Apple-silicon-only hosts used since PCC launched in 2024.
• Foundation model collaboration. Apple worked with Google to apply Gemini-family techniques when building next-generation Apple Foundation Models; on-device tiers still handle lighter tasks, but agentic tool-use and complex reasoning target the cloud tier on NVIDIA hardware.
• Supply-chain and attestation hardening. Apple maintains a cryptographically verifiable, append-only ledger of every Google Cloud machine in the PCC fleet. Components that could exfiltrate data if compromised are attested with at least two independent vendor roots of trust.
• Architectural patterns carry over. Initial request parsing runs in a dedicated namespaced process; shared inference processes recycle on a short time-to-live; attested keys live in a separate confidential VM isolated from external inputs.
• Transparency programme unchanged. PCC binaries remain published for public inspection, with research tooling and live research-mode nodes offered through the Apple Security Bounty Programme.

Why it matters for engineers

Confidential VMs and GPU encryption are now commodity cloud options. Apple’s claim is different: those primitives have not, until now, been composed into an end-to-end confidential inference pipeline that also ships public binaries and bounty-grade verification at global scale. PCC on Google Cloud is a reference for treating the entire stack — firmware through application code — as the trusted computing base, rather than trusting the guest VM boundary alone.

Platform teams building multi-tenant AI should study the operational patterns, not only the silicon. Stateless computation is enforced through short-lived inference workers and isolated parsers, reducing the blast radius if a host is misconfigured. Hardware inventory ledgers matter when you neither manufacture servers nor operate the facility: they convert supply-chain risk into auditable state. Dual roots of trust make it harder for a single vendor compromise to forge the entire attestation story.

For Apple Intelligence client engineers, the device-side contract is stable: only Apple-cryptographically-approved PCC releases execute, regardless of whether inference lands on Apple metal or a Google Cloud A3-class confidential GPU node. Preview ramp during summer 2026 means protection depth may converge over weeks — plan feature flags and telemetry accordingly until Apple declares parity with Apple-data-centre PCC.

Security researchers should watch the Confidential Computing Summit session and the forthcoming PCC Security Guide update for attestation quote formats, research-node access mechanics, and fleet geography. Until then, treat this announcement as architectural intent with preview availability, not a finished open inference API.

Apple Private Cloud Compute extends its privacy envelope to Google Cloud nodes using NVIDIA confidential GPUs, Intel TDX, and Titan-backed attestation.

flowchart LR
    DEV["Apple device"]
    TRUST["Apple-approved PCC client"]
    NODE["Confidential cloud node"]
    GPU["Stateless GPU inference"]
    RESP["Encrypted response"]

    DEV --> TRUST
    TRUST --> NODE
    NODE --> GPU
    GPU --> RESP
    RESP --> DEV

    classDef agent fill:#8B0000,color:#fff
    classDef tool fill:#189AB4,color:#fff
    class NODE,GPU tool
    class DEV,RESP agent

Amazon Bedrock now documents EU geographic cross-region inference profiles so teams in Europe can pool model capacity across Union Regions whilst keeping prompts and outputs inside a fixed EU routing boundary.

What changed

• Inference profile IDs replace plain model IDs. Applications opt into CRIS by passing system-defined profile strings such as eu.amazon.nova-2-lite-v1:0 (EU geographic) or global.amazon.nova-2-lite-v1:0 (global commercial) to Converse, InvokeModel, streaming APIs, batch jobs, Agents, and knowledge-base generation.
• EU geographic profiles constrain destination Regions. All destinations in EU CRIS lie within the European Union. Requests from EU sources cannot be routed to non-EU commercial Regions whilst using an eu.* profile.
• London and Zurich are special-cased. Sources in eu-west-2 may route among EU Regions plus London; eu-central-2 sources among EU Regions plus Zurich. Non-EU sources using EU profiles are optimised across the source Region and EU destinations only.
• Geographic profile Region lists are static. AWS will publish a new inference profile ID rather than silently expanding an existing EU geography definition.
• Audit fields ship in CloudTrail. Invocation metadata is logged in the customer source Region; additionalEventData.inferenceRegion records where Bedrock actually processed the request. Optional Model Invocation Logging keeps full payloads in the source Region only.
• Compliance framing is explicit. The post ties CRIS to GDPR records-of-processing expectations, IAM least privilege, and Amazon Bedrock’s inclusion in the CISPE Data Protection Code of Conduct.

Why it matters for engineers

EU SaaS teams no longer choose between single-Region throttling and unaudited multi-Region sprawl. EU CRIS is a deliberate contract: your SDK client stays in a familiar source Region, but Bedrock may execute inference in another EU Region selected for capacity. Inter-Region traffic remains on the AWS private backbone with encryption in transit — a detail that matters when security reviewers ask whether prompts leave controlled networks.

The integration surface is small; the governance surface is not. IAM policies for geographic CRIS must grant bedrock:InvokeModel on the inference profile and on foundation-model ARNs in every destination Region listed for that profile, often conditioned on bedrock:InferenceProfileArn. Service Control Policies that block any destination Region in the profile will fail requests even when the source Region is allowed. Cross-Region inference can also target Regions you have not manually enabled — SCP design must allow the full destination set.

Operational teams should dashboard inferenceRegion alongside application metrics. That field supports data-protection impact assessments without enabling payload logging. When maximum throughput or ~10% cost savings outweigh residency constraints, global.* profiles remain available — but that is an explicit product decision, not a framework default.

Discover profiles via the Bedrock console cross-Region inference page, per-model Regional availability tables in the user guide, or list_inference_profiles(typeEquals='SYSTEM_DEFINED') from your source Region. Treat profile choice as architecture documentation: EU geographic for GDPR-aligned processing, global for performance-first workloads with accepted cross-border inference risk.

EU geographic Bedrock inference profiles keep prompts and outputs inside Union Regions whilst pooling capacity across EU destination Regions.

flowchart LR
    APP["App in source Region"]
    API["Bedrock runtime API"]
    ROUTER{"CRIS profile router"}
    DEST["Destination Region inference"]
    RET["Response to source Region"]

    APP --> API
    API --> ROUTER
    ROUTER --> DEST
    DEST --> RET
    RET --> APP

    classDef agent fill:#8B0000,color:#fff
    classDef tool fill:#189AB4,color:#fff
    class ROUTER tool
    class APP,RET agent

Xiaomi MiMo and TileRT shipped MiMo-V2.5-Pro-UltraSpeed, a trillion-parameter API tier that sustains roughly 1000 tokens per second decode on a single eight-GPU commodity node—aimed at agent builders who need frontier-scale models inside realtime loops.

What changed

• 1000+ tps on 1T MoE. Xiaomi claims the first public trillion-parameter decode above 1000 tokens per second using one standard eight-GPU server, via model–system co-design rather than custom wafer or SRAM-only silicon.
• Selective FP4 on experts. MoE expert matrices quantise to FP4 (MXFP4) with quantisation-aware training; routers and attention stay higher precision to protect reasoning and code quality versus naive full-model FP4.
• DFlash speculative decoding. Block-level masked parallel drafting replaces serial draft-token generation; reported acceptance lengths reach ~6.3 (coding), ~5.6 (maths/reasoning), and ~4.3 (agent) tokens per verification round with block size eight.
• TileRT ultra-low-latency stack. Persistent engine kernels and warp-specialised pipelines cut microsecond execution gaps that dominate at kilohertz decode rates.
• Open weights. Hugging Face release XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash ships FP4 weights plus DFlash draft parameters for offline study.
• Gated trial. Approved users get free chat at ultraspeed.xiaomimimo.com during the promotion; enterprise partnerships via business-mimo@xiaomi.com.

Why it matters for engineers

Latency redefines what a trillion-parameter model can do. Below roughly ten tokens per second, 1T MoE models sit behind batch jobs and human-tolerated waits. Near 1000 tps, the same weights can participate in parallel Best-of-N search, sub-minute codegen sessions, or millisecond think–act loops in trading, fraud, and clinical triage—without downsizing to a 70B shortcut model.

The architectural lesson is co-design: bandwidth-bound expert matmuls shrink with FP4, serial decode expands via DFlash acceptance, and TileRT removes per-operator launch tax. Teams self-hosting open weights can benchmark the HuggingFace checkpoint on vLLM or SGLang; teams buying API capacity should measure cost per successful agent task during the June trial, not headline tokens per dollar alone.

Treat UltraSpeed as a latency SKU on MiMo-V2.5-Pro, not a new foundation family. Trial pricing and slots end 23 June 2026 unless extended; plan production fallbacks if FP4 quality drifts on your longest agent traces.

MiMo UltraSpeed stacks FP4 expert quantisation, DFlash speculative decoding, and TileRT persistent GPU pipelines to deliver roughly 1000 tokens per second from a one-trillion-parameter MoE on commodity hardware.