The Price of Frontier AI Just Hit $2.
The AI price war has a third combatant, a two-trillion-dollar prospectus, and — for the first time this year — a regulator who wants to join in.
Google launched Gemini 4 Argon on 30 September at US$2 / US$10 per million tokens with up to one-million-token output windows, and Artificial Analysis ranked it #1 on Text Arena with an Intelligence Index of 53. OpenAI shipped GPT-6.1 Sol the same day at the same US$2 / US$10 price, calling it "near-Astra for agentic coding" at one-fifth of Astra's cost. Anthropic filed its IPO prospectus on 30 September, confirming the US$517 billion in decade-long compute obligations reported earlier this month and targeting a valuation in the trillions. The US FTC opened a formal probe into OpenAI, Anthropic and evaluation firm METR on the same day, and on 1 October California enacted SB 947, the first US state law barring AI-only employment terminations. At home, PM Anwar told SME associations on 30 September that Budget 2027 — tabled 9 October, one week from today — will "consider cost pressures" on small enterprises.
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. The frontier has a third address — and the price just hit $2
On 30 September, Google released Gemini 4 Argon with paid API access at US$2 per million input tokens and US$10 output, with a one-million-token context window and output windows that can run to the same length. Artificial Analysis ranked it #1 on Text Arena with an Intelligence Index of 53. Google also launched the Fairwind Program, giving cyber defenders Argon access first and citing 68% autonomous vulnerability-detection on CWE-bench v1 — the kind of number that used to belong to specialist security vendors.
Hours later, OpenAI shipped GPT-6.1 Sol at the same US$2 / US$10 price. OpenAI called it "near-Astra performance for agentic coding" at one-fifth of Astra's cost, and cancelled the previously planned Astra 6.1 release.
Anthropic's Opus 5.5 — released 22 September at US$4 / US$20 — now sits as the premium option in a three-horse race whose cheapest tier is US$2 per million input tokens. That is roughly RM8.50, before any wrapper markup, for a million tokens of frontier-class work.
So what for a business owner: The baseline price for frontier AI dropped from "whatever the vendor charges" to a published commodity number this week. If your software provider's "AI premium" line item was quoted before 30 September, you have a renegotiation to make before Budget 2027 lands. The money that drops into your margin is the money the vendor was keeping.
2. Anthropic went public — and showed the compute bill
On 30 September, Anthropic filed its IPO prospectus with the SEC, moving from last year's confidential submission to a public document. The filing confirmed what The Information first reported on 13 September: Anthropic has committed to up to US$517 billion in compute deals over the next decade, locked in across cloud partners during the eleven months to September. The target valuation runs into the trillions.
The tension in the filing is open now: the compute bill is larger than many national economies, roughly 80% of it is non-cancellable, and the company's growth plan assumes enterprises keep buying more agent-hours every quarter at prices that fell this week. That is one bet made on your behalf.
So what for a business owner: A vendor that has pre-committed US$517 billion to its own supply will not undercut itself for your business — it will push volume. Expect enterprise sales pressure to intensify through Q4: unused credits, "pilot programs," expanded seats, extra workflows. Say yes only to what moves one of your real problems. The vendor's cost base is not your problem.
3. The regulators finally walked through the door
Three regulatory moves landed on the same two days.
30 September — FTC probe. The US Federal Trade Commission opened a formal investigation into OpenAI, Anthropic, and evaluation firm METR, examining whether safety claims and agent incidents violated US consumer-protection law. Information demands and executive testimony are expected.
30 September — White House "Super Intelligence" rename. OpenAI, Anthropic, Nvidia, Google, Meta, and Microsoft signed a "morally binding" voluntary safety accord with the administration. In the same guidance, Washington renamed AI to "Super Intelligence" in official usage.
1 October — California SB 947. The first US state law barring AI-only employment decisions. From 1 July 2027, no termination may be executed on AI output alone; the employer must document human corroboration and give written notice.
Last week, the frontier labs were asking the UN to regulate them. This week, the FTC and California started doing it without asking. The gap: three working days.
So what for a business owner: The idea that AI rules will "take years" survived about a week. Any process in your business where AI output can trigger a decision affecting a customer, supplier or employee — a credit refusal, a return rejection, a shortlisted CV — needs a documented human step above it. Not because Sacramento will prosecute a Melaka trading firm, but because that is the rule your biggest clients will start demanding by contract within six months.
Closer to Home: Malaysia
One dated story moved locally this week, and the clock on everything else kept running.
On 30 September, PM Anwar Ibrahim held a Budget 2027 engagement session at Perdana Putra, Putrajaya, with representatives of SME associations and Bumiputera contractor firms. The line carried by Malay Mail on 1 October: Budget 2027 will "consider the cost pressures in procurement and operations faced by small enterprises following the global supply crisis," and the MADANI Government will "open up wider avenues and opportunities" to help businesses build capability. No numbers were announced. Budget 2027 will be tabled on 9 October — one week from today. If your industry association's submission has not reached the Finance Ministry by Monday, it will not shape this budget.
The e-invoice clock has not moved. The RM10,000-per-transaction rule still applies to every business — including the 1.1 million now exempt below RM3 million turnover. The RM1–5 million band grace period runs to 31 December 2027, with penalties of RM200 to RM20,000 per invoice from 1 January 2028. LHDN's February 2026 scorecard — 500,000+ non-compliant cases, RM1.4 billion in unreported income — has not been publicly updated. Assume the audit engine is still running quietly.
Standing support remains live: the MSME Digital Grant MADANI (50% matching, up to RM5,000 per first tool), the GGSM4 government guarantee (up to 80%, up to RM20 million per company, applications open to 31 December 2026 or until the pot is drained), and HRD Corp levies redirectable toward AI-literacy training. First in the queue is first processed.
Singapore note: no material regional AI or SME story broke this week that changes the picture for Malaysian owners.
What This Means for Your Business
- Renegotiate before 9 October. The three biggest AI labs now publish a US$2 / US$10 per million token price for frontier-class work. If any software vendor in your stack charges an "AI premium" above those numbers, ask in writing when the 30 September price cuts flow into your bill. A vendor who cannot answer within a week has told you where their margin lives.
- Put a human step between AI and any customer-facing decision — this quarter. Credit terms, return rejections, hiring, supplier onboarding: if AI output influences the decision, document who signs off, within what deadline, and with what override authority. California's SB 947 and the FTC probe set the direction of travel. Your enterprise clients will contractualise this before any regulator reaches you.
- File your Budget 2027 ask by Monday 6 October. Three working days before tabling. SME association submissions still land. Your line: one specific grant, one specific exemption threshold, one specific e-invoice portal fix. Vague asks get vague allocations.
The Practical Question
"If I can buy frontier-class AI at RM8–10 per million output tokens next month, which of my repeat processes is still worth paying a person to do by hand?"
That is today's maths, not next year's. One million output tokens is roughly 750 pages of English prose. At RM10 that works out to about RM0.013 per page at the vendor's cost — before your wrapper, before training, before cleanup. Three processes in almost every Malaysian SME now fail that test on volume alone: statement-to-invoice matching, routine supplier correspondence, and first-pass CV screening. Pick one. Spend an hour this weekend writing down what each manual run currently costs you — in headcount hours and in errors — and bring that number into Monday's management meeting. The question is not whether AI is cheaper this quarter. It is whether you have stopped doing the maths.
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.