Three Big Cuts and a Warning
The AI industry cut its own prices twice on Tuesday and walked into the UN on Wednesday to ask someone else to slow it down.
Anthropic released Claude Opus 5.5 on 22 September with input and output prices 20% lower than Opus 5 and cache reads 60% cheaper. The same Tuesday, OpenAI shipped GPT-6 Sol and GPT-6 Luna at cuts of roughly 50% versus their predecessors. On 23 September at the UN Security Council in New York, Sam Altman and Dario Amodei asked for global oversight; the US and China said no. Twenty-four hours later, Anthropic said Claude — running as 950 agents for 21 hours — had discovered a novel bacteriophage enzyme system the lab does not yet fully understand.
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. Two labs cut model prices 20–50% on the same Tuesday
On 22 September, Anthropic released Claude Opus 5.5 at US$4 per million input tokens and US$20 output — down 20% from US$5/US$25 for Opus 5. Cache reads dropped from US$0.50 to US$0.20 per million tokens, a 60% cut in the price of showing the model the same document twice. On CursorBench for coding agents it scored 52.5% against GPT-5.6 Sol's 41.7% at roughly one-third the cost.
Hours later, OpenAI shipped two mid-tier models: GPT-6 Sol at US$2/US$10 per million tokens and GPT-6 Luna at US$0.10/US$0.50. Both cut predecessors' prices by roughly 50–58% and carry a 1.05 million-token context window. Sol handles repeated complex work — feature-building, code review, debugging. Luna handles high-volume clerical jobs like summarising documents and extracting fields. GPT-6 Astra remains the flagship for the hardest problems.
So what for a business owner: The AI capability your vendor was pricing at a monthly premium in January is materially cheaper to run this week. Ask in writing: "Given the 22 September price cuts, when will they show in our bill?" A vendor who cannot answer is keeping the margin.
2. AI CEOs asked the UN to write the rules — the US and China said no
On 23 September, at a France-chaired UN Security Council briefing during the General Assembly, Anthropic CEO Dario Amodei told the room: "If managed poorly, I even believe AI could be a risk to humanity as a whole." Sam Altman warned humanity could "lose control of the future of AI" and said big AI decisions "shouldn't be made by labs in San Francisco alone." Hugging Face CEO Clement Delangue and AI expert Yoshua Bengio joined them.
The US representative Michael Kratsios dismissed the ask bluntly: the Trump administration would "totally reject all efforts by international bodies to assert centralised control." China took the same line. So the strongest voices asking for global rules on frontier AI this week were the men who own the frontier AI — and the strongest voices saying no were the two governments whose signatures would matter most.
So what for a business owner: International rules on how AI vendors handle your data are not arriving soon. The rules governing your use of these tools for the next three years are the ones you write yourself — who touches what customer data, who signs off on what output, where the audit log lives. "We'll wait until it's regulated" is a two-year delay, not a plan.
3. Claude found a new enzyme in 21 hours — and Anthropic can't yet say what it does
Also on 23 September, Anthropic published research showing Claude — deployed as roughly 950 agents in parallel for 21 hours over 210 million tokens — identified a previously uncharacterised enzyme system in bacteriophages: array-associated reverse transcriptases (ART). It pairs a reverse transcriptase with a repeating DNA array that looks structurally similar to CRISPR arrays used in gene editing. MIT CRISPR pioneer Feng Zhang called it "an exciting example of how AI agents can contribute to biological discovery." Anthropic added carefully that what ART actually does biologically is still under investigation.
So what for a business owner: AI agents can now do weeks of specialised work in a day when pointed at a clean problem — the kind of bottleneck you have in reconciling statements, chasing outstanding invoices, or matching POs to deliveries. But even the lab that built the agents did not fully understand what came out. If your AI produces something surprising — an odd customer segment, an unusual invoice pattern, a rule the bot invented — slow down and check.
Closer to Home: Malaysia
Two things moved locally this week, and one has a date on it.
On 24 September, MDEC chief executive Anuar Fariz Fadzil told the Data Centre Nexus 2026 conference that 83 firms with data-centre activities now hold Malaysia Digital status, representing RM308.3 billion in approved investments. Data centres alone were RM95.8 billion — 44% of Malaysia's total approved investment in the first half of 2026. Anuar's point was blunt: most of that concrete and silicon sits in Johor and Selangor, and most Malaysian SMEs cannot afford to sit next to it. His remedy is to make the compute layer more affordable so SMEs can move "from experimenting with AI to developing applications that solve real problems." The shape it takes will decide whether a Melaka trading firm can rent Malaysian GPU time by the hour in 2027 or keeps paying in US dollars.
The Pre-Budget Statement 2027 landed on 22 September, ahead of Budget 2027 on 9 October. It names AI, digital services and semiconductors as core priorities, promises "AI literacy, digital capability and lifelong learning" tied to SME training, and points to a RM10 billion BNM-CGC guarantee for MSME financing. The Ministry of Finance's own figure: business AI adoption in Malaysia climbed from 27% in 2025 to 38% in 2026. If your business is in the 62% still on the sidelines, you have just over two weeks before your competitor's grant clears the queue.
The e-invoice clock has not moved. The RM10,000-per-transaction rule still applies to every business, including the 1.1 million exempt below RM3 million turnover. The RM1–5 million band's grace period runs to 31 December 2027, with penalties of RM200 to RM20,000 per invoice from 1 January 2028. LHDN's February 2026 audit numbers — 500,000+ non-compliant cases, RM1.4 billion in unreported income — have not been updated publicly. Assume the audit engine is still running.
Singapore note: no material Southeast Asian AI or SME story broke this week that changes the picture for Malaysian owners.
What This Means for Your Business
- Renegotiate this month. If any vendor charges you an "AI premium" as a line item, ask in writing when the 22 September price cuts from Anthropic and OpenAI will flow into your bill. A vendor who does not know or will not say is telling you their margin is not your margin.
- Write your own rules before someone else does. International frontier-AI regulation is not landing soon. Put on paper this week: who in your business may use AI on customer data, what they may and may not do with the output, and where the log lives. Two pages beats zero pages.
- Line up your Budget 2027 asks before 9 October. The MSME Digital Grant MADANI still covers 50% of a first tool up to RM5,000. HRD Corp levies you already pay (1% of monthly wages for firms with 10+ staff) can be redirected toward AI-literacy training. If fresh allocations land on 9 October, first in the queue is first processed. Have a shortlist of one to three tools ready by Wednesday 8 October.
The Practical Question
"If AI costs half as much next quarter, what do I do differently — hire, invest, or take the margin?"
The answer changes what kind of owner you are for the next two years. Take the margin, and competitors who reinvest will outbuild you. Hire, and you may be adding headcount to a workflow about to change. Invest — in a first tool, in staff training, in one clean process — and you spend it once and keep the compounding. Spend an hour this weekend deciding. The prices announced on Tuesday made it a real question.
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.