Anthropic Releases Claude Opus 4.8 With Dynamic Workflows at an Unchanged Price
Anthropic's new frontier model introduces native orchestration of hundreds of parallel subagents, effort controls, and a Fast mode three times cheaper, while keeping its predecessor's pricing.
On May 28, 2026, Anthropic unveiled Claude Opus 4.8, its new frontier model, 41 days after the release of Opus 4.7 and at an unchanged price of $5 per million input tokens and $25 per million output tokens. The most significant addition is Dynamic Workflows, in research preview: they let the model handle complex tasks by orchestrating hundreds of subagents in parallel within a single Claude Code session, available to Enterprise, Team, and Max plans. According to Anthropic, Claude Code with Opus 4.8 can now complete codebase-wide migrations — hundreds of thousands of lines of code, from green light through to merge — using the existing test suite as the validation benchmark. Also arriving are effort controls on claude.ai and Cowork (high, extra, and max levels), which let you choose the balance between response quality and rate-limit consumption, along with a Fast mode three times cheaper than previous models. On the reliability front, Anthropic reports that Opus 4.8 is four times less likely than Opus 4.7 to let code defects slip through and scores 84% on Online-Mind2Web. As TechCrunch reports, the model also shows a greater tendency to flag uncertainty and to avoid unsupported claims, a behavior confirmed as well by a comment from Bridgewater Associates cited by the company.
Why it matters
- LLM BUILDER/DEV · INGEGNERI ICT / IT MANAGER: Native orchestration of hundreds of subagents and effort controls change the design of agentic pipelines: codebase-wide migrations become a concrete use case, and model upgrades should be treated as part of the development cycle, not as isolated events.
- IMPRENDITORI: A more capable frontier model at an unchanged price, with a Fast mode three times cheaper, improves the cost/performance ratio of automation: the competitive bar rises without an increase in spending for early adopters.
«Do Language Models Need Sleep?»: Consolidating Context During Sleep to Streamline Online Inference
A paper on arXiv proposes an offline «sleep» phase that converts recent context into persistent fast weights within the SSM blocks, clears the KV cache, and moves computation off the online inference path.
A new arXiv paper (submitted on May 25, 2026, revision v2 on the 27th) by Sangyun Lee, Sean McLeish, Tom Goldstein, and Giulia Fanti proposes giving language models a «sleep» phase: Do Language Models Need Sleep? Offline Recurrence for Improved Online Inference. Instead of discarding the KV cache when the context window fills up (every L tokens), the model enters an offline phase and runs N recurrent passes over the accumulated context, updating the «fast weights» in the state-space (SSM) blocks via a local learning rule. The update is Hebbian with gating (S_t = α_t·S_{t-1} + β_t·v_t·k_tᵀ); once the N passes are complete, the KV cache is cleared. The key choice: at inference time, each response token is predicted with a single standard forward pass, so the «reasoning» happens during sleep rather than at test time.
The results show that sleep duration (N) helps the more complex tasks the most. On the Rule 110 cellular automaton with horizon t=32, the baseline tops out around 10% accuracy, two cycles climb to 20%, and three-to-four cycles exceed 30%, for the same training budget. On multi-hop retrieval (the Depo task) at 16 hops, only the 4-cycle model begins to improve. On the GSM-Infinite mathematical reasoning benchmark, Jet-Nemotron 2B goes from 0.351 to 0.388 on 8-operation problems with six cycles, while Ouro 1.4B rises from 0.419 to 0.615 on six operations with four cycles.
Why it matters
- RICERCA DI FRONTIERA: It is a signal toward hybrid attention-SSM architectures and toward a clean separation between offline computation (consolidation) and online computation (prediction), with explicit biological inspiration. The fact that the gains grow with N precisely on reasoning tasks opens a direction that is alternative — or complementary — to test-time compute.
- LLM BUILDER/DEV: This is latency engineering in its purest form: the cost of long context shifts to a batch/offline phase, leaving inference at a single forward pass per token. Relevant for anyone designing persistent memory and context management who wants to contain latency on the online path.
EU AI Act: the "Digital Omnibus" Delays High-Risk Obligations and Introduces New Prohibitions
The provisional agreement between the Council and Parliament pushes back the deadlines for high-risk systems and, from 2 December 2026, bans the generation of non-consensual intimate content and child sexual abuse material.
On 7 May 2026, the Council and the European Parliament reached a provisional agreement to simplify the AI Act, as part of the "Digital Omnibus" (Omnibus VII) package of the EU simplification agenda (Council of the EU).
At the heart of the agreement is a longer timeline for the rules on high-risk systems. The obligations for "stand-alone" Annex III systems — biometrics, critical infrastructure, education, employment, migration, asylum and border control — slip from 2 August 2026 to 2 December 2027 (a 16-month deferral). For high-risk systems embedded in products (such as elevators or toys, Annex I), application moves to 2 August 2028. The Member States' obligation to establish regulatory sandboxes is also postponed to 2 August 2027 (Inside Global Tech).
On the machine-readable labelling of synthetic content (art. 50(2)), systems already on the market before 2 August 2026 have until 2 December 2026; those placed on the market after that date must comply immediately.
Bucking the trend toward relaxation, two new prohibitions take effect from 2 December 2026: the generation of non-consensual intimate images ("nudifier" apps) and of child sexual abuse material (CSAM). The agreement also reintroduces two requirements: the obligation to register in the EU database for systems that providers deem exempt, and the strict-necessity standard for processing special categories of data for the purpose of bias detection. Violations of Article 25 remain punishable by up to 3% of worldwide turnover or 15 million euros.
Why it matters
- Entrepreneurs: Pushing back the Annex III obligations to December 2027 gives roadmaps and legal budgets some breathing room, but it is not a pause: the new prohibitions from 2 December 2026 and penalties of up to 3% of worldwide turnover make it imperative to review the governance and risk perimeter of systems in use right away.
- ICT engineers / IT managers: The technical deadlines diverge: machine-readable labelling of synthetic content must be implemented by 2 December 2026 for systems already on the market (immediately for new ones), while the compliance requirements for Annex III high-risk systems now have a December 2027 horizon — a timeline against which to re-plan development, documentation and registration in the EU database.
KPMG Brings Claude to 276,000 Professionals with Digital Gateway
Global KPMG-Anthropic alliance announced on May 19, 2026: Claude is being integrated into the Digital Gateway delivery platform on Microsoft Azure, with an initial focus on tax and private equity and full roll-out expected by September 2026.
KPMG has signed a global alliance with Anthropic to integrate Claude into its Digital Gateway, the technology delivery platform built on Microsoft Azure where tax expertise, proprietary tools, and client data come together. The deal was announced on May 19, 2026 and puts Claude's capabilities — in particular Cowork and Managed Agents — in the hands of the network's 276,000 professionals across 138 countries, with full implementation on Azure expected by September 2026 (Anthropic, May 19, 2026; picked up by the aggregator Build Fast with AI, May 28, 2026).
This is not a chatbot bolted onto workflows, but an agentic integration into core processes: KPMG professionals can build agentic workflows in real time directly within the platform they already use for clients. The cited example is the creation of an agent to keep up with changes in tax regulation: a task that previously took weeks and required switching between multiple tools and chat windows, and that now, with Cowork and Managed Agents embedded in the Digital Gateway, goes from weeks to minutes.
The initial focus is on tax clients and private equity funds, for whom KPMG positions itself as the go-to advisor. Among the offerings is KPMG Blaze, which embeds Claude Code to modernize legacy IT systems. The initiative builds on two years of internal Claude adoption across KPMG's US operations, including its AI and Data Labs and the firm's internal teams (Anthropic).
Why it matters
- Entrepreneurs: This is one of the largest enterprise deployments of Claude and marks the shift from chatbot to deep agentic integration into core processes — tax analysis, IT modernization, real-time workflows — on a global scale. For anyone running a business, it's a concrete signal of how agentic AI is entering professional services delivery and of what to expect, in terms of time and cost, from your own advisor.
At Google I/O 2026 the search box reinvents itself with information agents, and Gemini 3.5 Flash reaches GA
Google redesigns the search box for the first time in over 25 years, introduces information agents that work 24/7, and ships Gemini 3.5 Flash, a fast and affordable model for agentic tasks and coding.
At Google I/O 2026 (the May 19 keynote) Google announced the biggest update to the search box in over 25 years: a search box rethought with AI. It expands dynamically for more complex queries, offers suggestions that go beyond autocomplete, and accepts multimodal inputs — text, images, files, video, and Chrome tabs. The rollout is global, in every country and language where AI Mode is available (Google — Search's I/O 2026 updates).
The most substantial new feature is information agents: agents that run in the background 24/7 to monitor a topic — stock prices, flights, sporting events, news, the real estate market — and send synthesized updates via push notifications. Unlike the old Google Alerts, they compare sources and perspectives instead of merely notifying you. They are activated through AI Mode with a natural-language prompt and debut in summer 2026 for AI Pro and Ultra subscribers in the US (TechCrunch). Google reports that AI Mode has surpassed one billion monthly users, with queries more than doubling every quarter since launch (Google — 100 things at I/O 2026).
On the model front, Gemini 3.5 Flash is now generally available and is the default for AI Mode. Google places it above Gemini 3.1 Pro on coding and agentic benchmarks: Terminal-Bench 2.1 at 76.2%, GDPval-AA at 1656 Elo, MCP Atlas at 83.6%. Pricing is $1.50/M input tokens and $9/M output tokens, with output up to 4 times faster and a context window of roughly 1 million tokens (MarkTechPost).
Why it matters
- End users: Search stops being reactive: with information agents, Google promises proactive, synthesized updates on the topics you follow over time, while the new multimodal box changes the everyday act of searching. The initial limitation to Pro/Ultra subscriptions and to the US market alone is worth weighing.
- LLM builders / devs: Gemini 3.5 Flash in GA delivers frontier performance on agentic and coding tasks at reduced cost and latency ($1.50/9 per million tokens, output up to 4x faster), shifting the price/performance bar for anyone building agents. The stated benchmarks (Terminal-Bench, GDPval-AA, MCP Atlas) remain a partisan reference, though, to be validated against your own workloads.