AI Weekly Digest26 June – 3 July 2026

Fable 5 is back globally from July 1 on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Access on AWS, Google Cloud, and Microsoft Foundry is being restored as quickly as possible. Enhanced safeguards shipped with restoration: the specific jailbreak technique cited by the government is now blocked in more than 99% of cases by a new classifier layer. Trade-off: slightly more sensitive safety routing, with a higher rate of false positives on legitimate security development and debugging tasks. Anthropic is proposing an industry-wide jailbreak severity framework (with Amazon, Microsoft, Google, and Glasswing partners) built on four criteria: capability gain provided to the attacker, scope of that gain, ease of weaponisation, and discoverability. HackerOne bug-bounty programme launched for reporting cybersecurity jailbreaks. Subscription terms: Fable 5 is available for up to 50% of weekly usage limits on Pro, Max, Team, and select Enterprise plans through July 7, after which usage credits are required.

Claude Sonnet 5 launches June 30 as the new default for Free and Pro. Key benchmarks (Anthropic's own figures): SWE-bench Verified 72.7% (vs Sonnet 4.6: 62.3%, Opus 4.8: 79.4%); Terminal-Bench 76.1% (vs Sonnet 4.6: 55.4%, +20.7 points — the largest single benchmark jump of any Sonnet launch); BrowseComp (agentic search) and OSWorld-Verified (computer use) both approach Opus 4.8 at medium effort. Safety: lower hallucination, sycophancy, and undesirable-agent-behaviour rates than Sonnet 4.6; refuses malicious requests more consistently; resists prompt injection more reliably. Cyber safeguards enabled by default. In a Mozilla Firefox test, Sonnet 5 never produced a working exploit despite multiple attempts. Sonnet 5 is included in the Cyber Verification Program on the native Claude Platform, Claude Platform on AWS, and Claude in Microsoft Foundry (Vertex AI coming soon). Tokenizer note: Sonnet 5 uses the same newer tokenizer as Opus 4.7/4.8 and Fable 5 — approximately 30% more tokens for the same text. Audit your token budgets before switching production pipelines.

Rate limits raised across all paid plans on Chat, Cowork, Claude Code, and the Claude Platform. API tier structure simplified to Start, Build, and Scale (from the previous multi-tier structure). Sonnet and Haiku API rate limits raised at every tier. View your tier and current limits in the Claude Console.

Claude in Microsoft Foundry is now generally available. Hosted on Azure: inference in Azure, Azure authentication, billing, governance, US and global data zones, Anthropic as data processor. Hosted on Anthropic: full Claude API feature set, previously called Foundry Preview. Claude Sonnet 5 available in Foundry at launch on July 1. NVIDIA GB300 GPU support confirmed for Foundry. Foundry Agent Service can use Sonnet 5 as the reasoning core for multi-step planning, tool use, and task execution across enterprise systems. Feature and model parity between the two hosting options is the roadmap goal.

Self-hosted gateway ships inside the claude binary as a single stateless container. Identity: acts as an OIDC relying party against Google Workspace, Microsoft Entra ID, Okta, or any standards-compliant OIDC provider — issues short-lived sessions, no long-lived secrets on developer machines. Policy: managed settings defined once on the server; clients receive policy at sign-in and the gateway enforces it on every request (allowed models, default settings). Telemetry: the client stamps a usage metric on every request; the gateway relays it over OTLP to a collector you configure in your own network and on your own retention schedule.

Richer admin analytics: usage and cost breakdowns by model, team, and user in the admin console. Model-level entitlements: set model defaults and access permissions by group, not just organisation-wide. Spend alerts: configure spend-threshold notifications to catch overages before they land on the invoice. Analytics API: finance and IT teams can pull the same metrics into existing reporting systems programmatically. These controls complement existing spend caps, access and model routing, and the usage analytics dashboard already in Claude Enterprise.

Claude in Chrome is now generally available — Claude can browse, click, fill forms, and act on web pages directly inside Chrome. Update to v2.1.197: Sonnet 5 is now the default model in Claude Code with introductory pricing of $2/$10 per MTok through August 31. Background PR auto-open: background agents launched from claude agents now commit, push, and open a draft PR automatically when they finish code work in a worktree, instead of stopping to ask. New /dataviz skill: chart and dashboard design guidance with a runnable colour-palette validator. Background agent notifications: sessions that need input or finish now fire the Notification hook (agent_needs_input / agent_completed). Claude Platform on AWS added as a gateway upstream provider (anthropicAws); model-not-found responses now advance the failover chain rather than erroring out. Built-in Explore agent now inherits the main session's model (capped at Opus) instead of always running on Haiku. Subagents and context compaction now inherit the session's extended thinking configuration. Org default models: admins set the default model in the org console; it shows as 'Org default' in /model. Fixed: brief network drops mid-response now retry with backoff instead of aborting the turn. Fixed: Opus 4.7 fast mode deprecated, removal July 24.

Opus 4.7 fast mode deprecated: removal July 24, 2026. After removal, requests to claude-opus-4-7 with speed: 'fast' will return an error. Migrate to Opus 4.8 fast mode. Opus 4.6 fast mode already removed as of June 29. MCP tunnels API migrated: management API moved from /v1/organizations/tunnels on the Admin API to /v1/tunnels on the Claude API; new header: anthropic-beta: mcp-tunnels-2026-06-22; new WIF scope: workspace:manage_tunnels. Previous surface remains available during migration window. SDK update: Python, TypeScript, Go, Java, Ruby, PHP, and C# SDKs now include support for code_execution_20260120 (REPL state persistence, minimum version for programmatic tool calling). No beta header required — set type to code_execution_20260120.

OpenAI frames this as evidence that AI is moving beyond question answering towards autonomous task execution. Active users increased more than fivefold during 2026. More than 10% of users now manage three or more concurrent Codex agents. The paper argues that agentic AI is beginning to reshape how organisations organise work rather than simply making existing work faster.

New reasoning effort options: Instant, Medium, High, Extra High, Pro Standard, Pro Extended. Thinking Light removed. This reflects OpenAI's broader strategy of hiding model complexity from end users and allowing routing systems to determine the appropriate reasoning level automatically.

Sub-millisecond processing speeds and drastically lowered compute requirements relative to Gemini 3 Pro Image. Supports a 1M token input context window. Output: 4K-token image matrices paired with up to 64K tokens of explanatory text. Exceptional performance in internationalized text rendering across non-Latin scripts and character consistency across sequential edits. Pricing: approximately 80% cost reduction per thousand generated matrices relative to gemini-3-pro-image, at an identical input/output token cost ratio to the standard text-based gemini-3.1-flash-lite model card.

macOS native client with local file system access (user-defined directory gating). Upcoming: remote execution pipeline — issue instructions from mobile to remote Mac hardware (retrieve sales reports, extract metrics, email summaries). Background task orchestration with automated real-time polling of web hooks, RSS feeds, financial tickers, and sports data streams. MCP support: developers can build and expose custom local or remote MCP servers directly to the Gemini Spark client, enabling custom internal enterprise applications to be called natively within the Gemini chat interface. New third-party extensions: Canva, Dropbox, Instacart, OpenTable, Zillow Rentals. Native integrations: Google Tasks, Google Keep.

Vertex AI brand retired; all enterprise ML workloads now under the Gemini Enterprise Agent Platform. Agent Designer: visual no-code/low-code flowchart environment for designing trigger-based autonomous agent paths that run in isolated cloud sandboxes. Agent Inbox: centralised operations room for prioritising and auditing agent activity. Projects: strictly walls off an agent's memory to assigned datasets, Drive repositories, and Group chats — enforcing clean data boundaries across departments. Skills: custom macros via @mentions to execute deterministic tasks across the enterprise workspace. Agent Gallery: out-of-the-box connectors for Asana, Workday, Mailchimp, Adobe, Atlassian, Lovable, and ServiceNow.

External coding agents can natively call foundation models from Google's Model Garden, pull approved internal prompt templates, or manage notebooks within an active project. Implementation requires three steps: enable the Agent Platform API in Google Cloud Console, configure the local client JSON payload to point to the remote Google Cloud MCP endpoint, and copy the provided Toolset Endpoints into the external IDE workspace. Native Cloud IAM integration handles enterprise token exchange and session validation automatically at the boundary layer.

Gemini Formula Fixer (Sheets): diagnoses syntax/logic errors and updates formulas inline. Iterative Prompt Refinement (Gmail): persistent instruction bar for surgical text edits without rewriting the full prompt. 28-language sheet localisation (Spanish, Japanese, French, German, Korean, and more). AI Inbox (Alpha Track only): high-volume executive triage.

Worldwide GA: all enterprise tenants. Model auto-selection: Cowork automatically routes between GPT-5.5 Thinking, Claude, and GPT-5.5 Instant based on task type. Expanded plugin ecosystem: Enosix, Harvey, LSEG, Miro, monday.com, Moodys, Morningstar, S&P Global Energy, TeamsMaestro, plus Databricks via sideloading. Microsoft Fabric and Dynamics 365 apps now integrated. Cost Management Dashboard in admin center: admins can monitor Cowork credit usage, budgets, and spend. Customize tab for custom skills.

Word: model choice now available; Copilot Catchup content card; image creation; iOS agentic capabilities; preserves chat history across apps; applies edits based on comments. Excel: Skills, personalisation, and rules; Planner Agent availability; deeper reasoning over Power BI enterprise data. Outlook: Compose canvas refinement; Copilot settings in Outlook Classic rolling out next month. PowerPoint: Brand Kit Picker for approved templates; reusable Copilot skills; references to SharePoint libraries and OneDrive folders.

Cost Management Dashboard: admins can monitor Cowork credit usage, budgets, and spend in the admin center. Purview DLP controls: email restriction and Cowork governance added. Federated Connectors GA: secure external app integration with staged rollout.

Outlook emails can now be added as references in Copilot Notebooks. Notebooks are now available to all Copilot Chat users, not just licensed Copilot 365 users.

Fable 5 je ponovo globalno dostupan od 1. jula na Claude Platformi, Claude.ai, Claude Code i Claude Cowork-u. Pristup na AWS-u, Google Cloud-u i Microsoft Foundry-u se obnavlja što je brže moguće. Poboljšane zaštitne mere isporučene uz obnavljanje: konkretna tehnika jailbreak-a na koju je vlada ukazala sada je blokirana u više od 99% slučajeva novim slojem klasifikatora. Kompromis: nešto osetljivije bezbednosno rutiranje, s višom stopom lažnih pozitivnih na legitimnom bezbednosnom razvoju i zadacima debagovanja. Anthropic predlaže industrijski okvir za ozbiljnost jailbreak-a (s Amazon-om, Microsoft-om, Google-om i Glasswing partnerima) zasnovan na četiri kriterijuma: dobijeni kapaciteti napadača, obim tih kapaciteta, lakoća naoružavanja i otkrivljivost. HackerOne bug-bounty program lansiran za prijavljanje sajber bezbednosnih jailbreak-ova. Uslovi pretplate: Fable 5 je dostupan za do 50% sedmičnih ograničenja upotrebe na planovima Pro, Max, Team i pojedinim Enterprise planovima do 7. jula, nakon čega su potrebni krediti za upotrebu.

Claude Sonnet 5 lansiran 30. juna kao novi podrazumevani za Free i Pro. Ključni benchmarkovi (Anthropicovi sopstveni podaci): SWE-bench Verified 72.7% (nasuprot Sonnet 4.6: 62.3%, Opus 4.8: 79.4%); Terminal-Bench 76.1% (nasuprot Sonnet 4.6: 55.4%, +20.7 poena — najveći skok na jednom benchmarku u istoriji lansiranja Sonnet modela); BrowseComp (agentna pretraga) i OSWorld-Verified (upotreba računara) oba se pri srednom naporu približavaju Opus 4.8. Bezbednost: niže stope halucinacija, sycophancy-a i neželjenog agentnog ponašanja nego Sonnet 4.6. Sajber zaštitne mere su podrazumevano omogućene. Napomena o tokenizeru: Sonnet 5 koristi isti noviji tokenizer kao Opus 4.7/4.8 i Fable 5 — otprilike 30% više tokena za isti tekst. Proverite budžete tokena pre prelaska produkcijskog pipline-a.

Rate limiti povećani na svim plaćenim planovima na Chat-u, Cowork-u, Claude Code-u i Claude Platformi. API struktura tierova pojednostavljena na Start, Build i Scale. Sonnet i Haiku API rate limiti povećani na svakom tieru. Pogledajte vaš tier i trenutne limite u Claude Konzoli.

Claude u Microsoft Foundry je sada opšte dostupan. Hosted on Azure: zaključivanje unutar Azure-a, Azure autentifikacija, naplata, upravljanje, US i globalne data zone, Anthropic kao procesor podataka. Hosted on Anthropic: pun skup Claude API funkcija (ranije zvano Foundry Preview). Claude Sonnet 5 dostupan u Foundry-u od lansiranja 1. jula. Podrška za NVIDIA GB300 GPU potvrđena za Foundry. Foundry Agent Service može koristiti Sonnet 5 kao jezgro rezonovanja za višestepeno planiranje, upotrebu alata i izvršavanje zadataka.

Sopstveni gateway se isporučuje unutar claude binarne datoteke kao jedan stateless kontejner. Identity: funkcioniše kao OIDC relying party prema Google Workspace-u, Microsoft Entra ID-u, Okta-i ili bilo kom standardnom OIDC provideru — izdaje kratkotrajne sesije, bez dugoživelih tajni na razvojnim mašinama. Politika: upravljana podešavanja definisana jednom na serveru; klijenti primaju politiku pri prijavi i gateway je primenjuje na svaki zahtev (dozvoljeni modeli, podrazumevana podešavanja). Telemetrija: klijent pečatira metriku upotrebe na svaki zahtev; gateway je prosleđuje putem OTLP-a na kolektor koji sami konfigurirate u vlastitoj mreži i po sopstvenom rasporedu zadržavanja.

Bogatija admin analitika: raščlamba upotrebe i troškova po modelu, timu i korisniku u admin konzoli. Ovlašćenja na nivou modela: postavljanje podrazumevanih modela i dozvola pristupa po grupi, ne samo na nivou cele organizacije. Upozorenja potrošnje: konfigurišite notifikacije praga potrošnje za hvatanje prekoračenja pre nego što stignu na fakturu. Analytics API: finansijski i IT timovi mogu programatski povlačiti iste metrike u postojeće sisteme izveštavanja.

Claude u Chrome-u je sada opšte dostupan — Claude može da pregleda, klikće, popunjava formulare i deluje na web stranicama direktno unutar Chrome-a. Ažurirajte na v2.1.197: Sonnet 5 je sada podrazumevani model u Claude Code-u po uvodnoj ceni od $2/$10 po MTok do 31. avgusta. Auto-otvaranje PR-a u pozadini: pozadinski agenti pokrenuti iz claude agents sada automatski commituju, pushuju i otvaraju draft PR kada završe rad s kodom u worktree-u, umesto da staju da pitaju. Novi skill /dataviz: smernice za dizajn grafikona i dashboardova s validatorom palete boja koji se može pokrenuti. Notifikacije pozadinskog agenta: sesije kojima je potreban unos ili koje završe sada aktiviraju Notification hook (agent_needs_input / agent_completed). Claude Platforma na AWS-u dodata kao gateway upstream provider (anthropicAws). Ugrađeni Explore agent sada nasleđuje model glavne sesije (ograničen na Opus) umesto da uvek radi na Haiku-u. Podagenti i kompakcija konteksta sada nasleđuju konfiguraciju proširenog razmišljanja sesije. Podrazumevani modeli org-a: admini postavljaju podrazumevani model u org konzoli; prikazuje se kao 'Org default' u /model. Popravljeno: kratki mrežni prekidi usred odgovora sada se ponavljaju s backoff-om umesto da prekidaju izmenu.

Deprecacija Opus 4.7 fast mode: uklanjanje 24. jula 2026. Nakon uklanjanja, zahtevi za claude-opus-4-7 s speed: 'fast' vratiće grešku. Migrirajte na Opus 4.8 fast mode. Opus 4.6 fast mode je već uklonjen 29. juna. Migracija MCP tunnels API-ja: management API premešten iz /v1/organizations/tunnels na Admin API-ju na /v1/tunnels na Claude API-ju; novi header: anthropic-beta: mcp-tunnels-2026-06-22; novi WIF opseg: workspace:manage_tunnels. Prethodna površina ostaje dostupna tokom prozora migracije. Ažuriranje SDK-a: Python, TypeScript, Go, Java, Ruby, PHP i C# SDK-ovi sada uključuju podršku za code_execution_20260120 (persistentnost REPL stanja, minimalna verzija za programatsko pozivanje alata). Nije potreban beta header — postavite type na code_execution_20260120.

OpenAI ovo okviruje kao dokaz da AI prelazi iz odgovaranja na pitanja ka autonomnom izvršavanju zadataka. Aktivni korisnici su se povećali više od peterostruko tokom 2026. Više od 10% korisnika sada upravlja s tri ili više istovremenih Codex agenata. Studija tvrdi da agentni AI počinje da menja način na koji organizacije organizuju rad, a ne samo da ubrzava postojeći rad.

Novi nivoi napora rezonovanja: Instant, Medium, High, Extra High, Pro Standard, Pro Extended. Thinking Light je uklonjen. Ovo odražava OpenAI-jevu širu strategiju skrivanja složenosti modela od krajnjih korisnika i prepuštanja sistemima rutiranja da automatski odrede odgovarajući nivo rezonovanja.

Sub-milisekundne brzine obrade i drastično smanjeni zahtevi za računarima u poređenju s Gemini 3 Pro Image. Podržava ulazni kontekstni prozor od 1M tokena. Izlaz: matrice slika od 4K tokena parene s do 64K tokena objašnjivačkog teksta. Izvanredne performanse u međunarodnom renderovanju teksta kroz ne-latinična pisma i konzistentnosti karaktera kroz uzastopne izmene. Cena: ~80% smanjenje troškova po hiljadu generisanih matrica u poređenju s gemini-3-pro-image.

macOS nativni klijent s pristupom lokalnom fajl sistemu (korisnikovim direktorijum gatovanjem). Predstojeće: pipeline za udaljeno izvršavanje — izdajte instrukcije s mobilnog uređaja svom udaljenom Mac hardveru. Orkestracija pozadinskih zadataka s automatskim praćenjem web hookova, RSS feedova, finansijskih tikera i sportskih podataka u realnom vremenu. MCP podrška: programeri mogu graditi i izlagati prilagođene lokalne ili udaljene MCP servere direktno Gemini Spark klijentu. Nove ekstenzije trećih lica: Canva, Dropbox, Instacart, OpenTable, Zillow Rentals. Nativne integracije: Google Tasks, Google Keep.

Brend Vertex AI povučen; sva enterprise ML radna opterećenja sada pod Gemini Enterprise Agent Platformom. Agent Designer: vizuelno no-code/low-code okruženje flowchart-a za dizajn autonomnih agentnih putova u izolovanim cloud sandbox-ovima. Agent Inbox: centralizovana operativna soba za prioritizaciju i reviziju agentne aktivnosti. Projekti: striktno ograđuju memoriju agenta na dodeljene datasete, Drive repozitorijume i Group chatove. Skills: prilagođeni makroi putem @mentiona za izvršavanje determinističkih zadataka. Agent Gallery: gotovi konektori za Asana, Workday, Mailchimp, Adobe, Atlassian, Lovable i ServiceNow.

Eksterni coding agenti mogu nativno pozivati modele iz Google Model Garden-a, povlačiti odobrene interne šablone prompta ili upravljati notebook-ovima unutar aktivnog projekta. Implementacija zahteva tri koraka: omogućite Agent Platform API u Google Cloud Console-u, konfigurišite lokalni client JSON payload da pokazuje na udaljenu Google Cloud MCP krajnju tačku, i kopirajte pružene Toolset Endpoints u eksterni IDE radni prostor. Nativna Cloud IAM integracija automatski rukuje enterprise token exchange-om i validacijom sesije na nivou granice.

Gemini Formula Fixer (Sheets): dijagnostikuje sintaksne/logičke greške i ažurira formule inline. Iterativno dorada prompta (Gmail): trajna instrukcijska traka za hirurške izmene teksta bez prepisivanja celog prompta. 28-jezička lokalizacija tabele (špansko, japansko, francusko, nemačko, korejsko i više). AI Inbox (samo Alpha Track): visokoobjemno izvršno razvrstavanja.

Globalna GA: svi enterprise tenanti. Auto-selekcija modela: Cowork automatski rutira između GPT-5.5 Thinking, Claude-a i GPT-5.5 Instant-a na osnovu vrste zadatka. Prošireni ekosistem pluginova: Enosix, Harvey, LSEG, Miro, monday.com, Moodys, Morningstar, S&P Global Energy, TeamsMaestro, plus Databricks putem sideloading-a. Microsoft Fabric i Dynamics 365 aplikacije su sada integrisane. Cost Management Dashboard u admin centru: admini mogu pratiti Cowork kreditnu upotrebu, budžete i potrošnju. Tab Prilagodavanje za prilagođene skills.

Word: izbor modela sada dostupan; content card Copilot Catchup; kreiranje slika; iOS agentne mogućnosti; čuva istoriju čata između aplikacija; primenjuje izmene na osnovu komentara. Excel: Skills, personalizacija i pravila; dostupnost Planner Agent-a; dublje rezonovanje nad Power BI enterprise podacima. Outlook: dorada Compose canvas-a; Copilot podešavanja u Outlook Classic-u koja se uvode sledećeg meseca. PowerPoint: Brand Kit Picker za odobrene šablone; višekratni Copilot skills; reference na SharePoint biblioteke i OneDrive fascikle.

Cost Management Dashboard: admini mogu pratiti Cowork kreditnu upotrebu, budžete i potrošnju u admin centru. Purview DLP kontrole: dodato ograničenje email-a i upravljanje Cowork-om. Federated Connectors GA: bezbedna integracija eksternih aplikacija s faznim uvođenjem.

Outlook email-ovi se sada mogu dodavati kao reference u Copilot Notebooks-u. Notebooks su sada dostupni svim Copilot Chat korisnicima, ne samo licenciranim Copilot 365 korisnicima.