Two frontier models in three days. The useful news is pricing and defaults. Anthropic released Claude Fable 5.1 on 1 September and cut cache reads by 75% to $0.25 per million tokens, lowering typical workloads by roughly 25% and highly agentic work by up to roughly 45%. OpenAI followed with GPT 6 Astra on 3 September at $10 per million input and $50 per million output tokens. - Repetition is where savings sit. The cut lands on reused context, so long agentic runs benefit most. - Check the default first. Astra is off by default in enterprise workspaces at launch and an administrator has to enable it. - Data custody moved as well. Anthropic's Enterprise Frontier Safeguards keep customer data on the customer's own cloud, phasing in from this autumn. - Speed is a paid option. Astra offers a Fast mode at up to twice the speed for twice the standard price. Read the pricing page before the benchmark chart.

The expensive mistake is not the tool. It is the first project. DBS surveyed 730 Singapore companies between December 2025 and January 2026. Close to two in five, 39%, are seeking expert advice on how to integrate AI meaningfully. DBS, EnterpriseSG and IMDA have since restructured the Spark GenAI programme around readiness, opening access to IMDA's Open Innovation Platform and its network of more than 16,000 solution providers. - Support is tiered now. Start, Accelerate and Scale match the depth of help to how far along a business actually is. - Diagnose before you buy. A new playbook sorts firms into AI Observer, AI Explorer, AI Practitioner and AI Trailblazer. - Some of it costs nothing. IMDA runs complimentary workshops, and pre approved AI solutions carry EnterpriseSG grant support. - The evidence is local. These are Singapore firms, not a global average borrowed from a different market. Spend the first month diagnosing. It is the cheapest month in the project.

Agents are in production. Trust is not, and the gap now has a price. Forrester Consulting surveyed 409 senior technology decision makers for Boomi in July 2026. 86% of organisations have moved past piloting agents, yet only 34% trust what those agents do. Among the least ready quartile, 77% shipped anyway, carrying an average of $2.1 million in added costs from fines, lost customers, downtime and rework. - Integration platforms split the field. 46% of the most ready run agentic workflows on an integration platform, against 25% of the least ready. - APIs decide what starts. The most ready were three times likelier to say well managed APIs determine whether a use case gets piloted at all. - Focus differs too. 47% of leaders in control prioritise integration with tools and apps, against 31% still tuning the model. - Sprawl is the symptom. Some organisations run as many as 200 agents that were never connected to anything. Connect one workflow properly before you approve the next agent.

Your customer's compliance deadline has a way of becoming your procurement questionnaire. CSA will publish an updated Cybersecurity Code of Practice for Critical Information Infrastructure later this year, requiring owners to attain Cyber Trust Mark Level 5 certification. A separate Code of Practice for Cloud Services follows in the second half of 2026, with companion guides written alongside Amazon Web Services, Google Cloud and Microsoft Azure. - Boards are named. Owners must keep a documented cyber resilience framework covering risk tolerance, mitigation, transfer and recovery, reviewed at least annually. - Interconnected systems count. Oversight extends to everything that connects to and communicates with the CII, which reaches suppliers. - Exercises become standing practice. A comprehensive cybersecurity exercise plan is required rather than an occasional tabletop. - Certification is already moving. CSA has required all CII owners to attain Cyber Trust mark certification by end 2027. If you sell into a critical sector, expect these questions at your next renewal.

Four checks from the week, none of them optional for very long. The week ran from a regulator deadline to the gap between feeling productive and being profitable, then closed on who controls your tools. - A date you can diarise. PDPC steps up enforcement from the start of 2027 against NRIC numbers used to authenticate. The audit belongs this quarter. - A ratio that explains your integration bills. Most applications in the average estate are not connected, and a third of IT time goes to wiring them by hand. - A gap between productive and profitable. Smaller organisations did not move on agent scaling while larger ones did, and financial impact stayed flat. - A reminder that tools rest on contracts. One coding tool loses a model family in November, and two further vendor posts covered security findings and agents running physical devices. Take the deadline first. The other three are decisions, not dates.