Astra, Anwar, and the Audit Reprieve
“This week: OpenAI released GPT-6 Astra and used the word "AGI" out loud. Anthropic answered with Claude Fable 5.1 at the same headline price. Prime Minister Anwar raised the e-invoice exemption threshold from RM1 million to RM3 million during his National Day address, and LHDN made it official on 1 September — more than 1.1 million MSMEs are now off the mandatory queue.”
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. OpenAI shipped GPT-6 Astra and said the AGI word
On 3 September, OpenAI announced GPT-6 Astra as a staged rollout — trusted partners first, then Plus, Pro, Business and Enterprise users, then the API and AWS. President Greg Brockman called it a "generational leap" that could in retrospect be seen as the arrival of artificial general intelligence. Astra is priced at US$10 per million input tokens and US$50 per million output tokens on the standard API — 2.5 times the current promotional rate of GPT-5.6 Sol.
Two things are worth separating. First, the benchmarks: Astra reportedly posted perfect or near-perfect scores on the reasoning tests OpenAI cares about, beating GPT-5.6 Sol and Claude Fable 5. Second, the framing: labs have called every release since 2023 "a step towards AGI." Brockman's version is louder because the model is measurably better at software engineering, computer use and cybersecurity, not because a threshold has been formally crossed.
So what: the ceiling of what an AI system can do for you rose again this week, and the price of the ceiling rose with it. If your business runs on the cheap tier — GPT-5.6 Sol, Gemini 3.7 Flash, DeepSeek V4 — nothing about your bill changed. If you were planning to buy the newest thing on principle, the newest thing now costs 2.5x what you were paying six weeks ago.
2. Anthropic quietly matched, and cut the cache price 75%
Two days earlier, on 1 September, Anthropic launched Claude Fable 5.1 and Mythos 5.1. Same headline pricing as Astra — US$10 per million input, US$50 per million output — with a 1 million-token context window and cached input tokens dropping from US$1 to US$0.25 per million, a 75% cut. Fable 5 moved to legacy the same day. Anthropic also inked a US$35 billion cloud deal with Lambda in the same week.
The Fable/Mythos split matters more than the version bump. Fable 5.1 is the model most customers get; it blocks sensitive cybersecurity and biology tasks. Mythos 5.1 is the same model without those blocks, and it is being handed to "a limited number of trusted organisations" only. Two labs, in one week, chose to gate their most capable models behind explicit trust lists. That is a new pattern, not a footnote.
So what: if you have a workflow that hammers the same reference documents — a chatbot over your SOPs, a review tool over the same contract templates, a compliance checker over the same MFRS clauses — the 75% cache discount is where the real ringgit savings live. Cached tokens are how you take advantage of a 1M-token context window without your bill following the context around.
3. McKinsey called it a two-speed race
On 2 September, McKinsey's latest enterprise AI report landed with a blunt frame: the market is splitting into companies that have re-architected workflows around AI and companies that have bolted chatbots onto existing processes. The first group is compounding gains. The second group is not.
The report is a consultancy report and should be read as one — the numbers are self-reported and the methodology favours the customers McKinsey already sells to. But the underlying observation is being echoed by every serious operator this year: value comes from redesigning a task end-to-end, not from adding an assistant to the task as it currently exists. Docusign's announcement the same week — opening its contract tools to ChatGPT and "every AI agent worldwide" — is the same idea from the vendor side. Docusign wants to be the layer other agents call, not the app you open.
So what: if you have deployed AI in your business and cannot point to one workflow that runs differently than it did a year ago, you are in the slow lane of the race McKinsey is describing. The fix is not more AI. The fix is picking one process — quotations, reconciliation, incoming enquiries — and redesigning it, then measuring.
Closer to Home: Malaysia
The e-invoice deadline just moved for a lot of you. On 1 September, LHDN confirmed the exemption threshold for mandatory e-invoicing has been raised from RM1 million to RM3 million in annual turnover, aligning with the announcement Prime Minister Anwar made during his 2026 National Day address. More than 1.1 million MSMEs are now exempt from the mandatory rollout.
Read the fine print carefully, because two things did not change:
- The RM10,000 rule is still live. Any single business transaction of RM10,000 or more still requires its own individual, LHDN-validated e-invoice and cannot be rolled into a monthly consolidated e-invoice. This has been in force since 1 January 2026 and the September announcement did not touch it. If a customer above the threshold — or a corporate buyer of any size — asks you for an e-invoice on a large purchase, you still have to produce one.
- The audit exposure did not disappear. LHDN reported in February that it had already logged more than 500,000 non-compliant cases and RM1.4 billion in unreported income. Exemption from the mandatory rollout is not exemption from proper record-keeping, and it is not amnesty.
Separately, TikTok Shop Malaysia announced it will run more than 1,000 All-Star LIVE sessions across 2026, up from around 100 in 2025 — a signal that the platform is doubling down on live commerce as the acquisition channel for MSMEs. Shopee still holds roughly 53% of Southeast Asian marketplace share, but TikTok Shop's growth rate keeps compressing that gap.
Singapore note: no material SME-relevant policy movements this week.
What This Means for Your Business
- If your turnover is under RM3 million, do not switch off compliance thinking — switch off the panic. You are no longer racing the Phase 4 deadline. But keep your bookkeeping tight, keep your books ready to produce an e-invoice on any RM10,000+ transaction, and keep an eye on the threshold, because it can move again. What was RM500,000 in Phase 1 is now RM3 million; policy is directional, not permanent.
- Stop buying the newest model on reflex. The cheap tier — GPT-5.6 Sol, Gemini 3.7 Flash, DeepSeek V4, Fable 5 (now legacy but still available), open-weights Kimi K3 — is where 90% of SME workflows live comfortably. The US$10/US$50 tier is for coding assistants, contract review, and long-context research over your own documents. Match the model to the job. If you can't tell whether a task needs the frontier tier, it doesn't.
- Pick one workflow and redesign it this month. Not "add AI to." Redesign. The McKinsey slow lane is any business where AI is a browser tab. Candidates that consistently work for Malaysian SMEs: quotation generation from an emailed enquiry, reconciliation of DuitNow QR settlements against POS receipts, first-response triage on WhatsApp Business, and monthly management reporting off your accounting export. Pick one. Measure the before. Measure the after.
The Practical Question
Which task in your business will run differently at the end of this month than it did at the start?
If the answer is "none" — either because you have not deployed anything, or because everything you have deployed sits alongside the old process instead of replacing it — you have just described the slow lane. The e-invoice reprieve gave a lot of Malaysian SMEs back the RM10,000 to RM50,000 they were about to spend on compliance software. The question is whether that money buys you a filing cabinet you didn't need, or one redesigned workflow that pays for itself every month from here on.
At The Empyrean, we help Malaysian SMEs find the practical, repeatable tasks where AI delivers value without disruption. If you're not sure where to start, we're happy to take a look at your operations and tell you honestly what would make sense.