Frontier AI Moved Down a Shelf
The frontier tier stopped costing frontier money this week.
“On 30 June, Anthropic released Claude Sonnet 5 at US$2 per million input tokens and US$10 per million output tokens — near-Opus 4.8 performance on agentic workloads at roughly a third of Opus prices — and made it the default model on Claude Free and Pro from 1 July. On 1 July, Bloomberg reported that Meta is standing up a cloud services business, internally called Meta Compute, to sell excess AI infrastructure and hosted access to its own models. Google shipped Nano Banana 2 Lite the same week at US$0.034 per 1K-resolution image in about four seconds. And in Putrajaya on 29 June, Prime Minister Anwar Ibrahim launched Malaysia Digital 2030, a seven-thrust national plan targeting 30% of GDP from the digital economy by 2030.”
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. Claude Sonnet 5 Put Near-Frontier Performance on a Mid-Tier Bill
On 30 June, Anthropic released Claude Sonnet 5, the newest version of its mid-tier model. The headline isn't the benchmark score — it's the price.
Launch pricing sits at US$2 per million input tokens and US$10 per million output tokens through 31 August (then US$3 and US$15). On SWE-bench Pro coding, Sonnet 5 scored 63.2% against Opus 4.8's 69.2%. On OSWorld-Verified agent tasks, 81.2% against 83.4%. On Terminal-Bench 2.1 — closest to just do the thing for a business owner — it beat Opus 4.8, 80.4% to 74.6%. The context window is 1 million tokens. From 1 July, Sonnet 5 is the default model for Claude Free and Pro users.
The workloads that were "call Opus, pay Opus prices, wait" three months ago — chasing overdue invoices across three systems, cleaning a supplier's messy CSV, drafting a proposal from four scattered documents — now run on the middle tier. If your business piloted an agentic workflow this year and shelved it because the token bill looked ugly, the arithmetic changed on Tuesday.
2. Meta Wants to Sell You Cloud
On 1 July, Bloomberg reported that Meta is standing up a cloud services business — internally called Meta Compute — to sell access to its excess AI infrastructure and, potentially, hosted access to its own models. Meta shares jumped more than 10% on the news. The initiative is led by Meta's head of infrastructure and its Superintelligence Labs leader.
Look past the stock move. Three hyperscalers — AWS, Azure and Google Cloud — set the underlying price of every AI feature your SaaS vendors resell to you. Adding a fourth well-capitalised competitor with its own data centres and its own models does one thing reliably over the next eighteen months: it pushes prices down. Meta already runs WhatsApp, Instagram and Facebook, which is where a large share of Malaysian SME commerce actually happens. The company that carries your customer conversations is now planning to also carry your compute. The practical implication isn't "switch clouds" — it's assume AI features inside your existing SaaS keep getting cheaper, and renegotiate accordingly on any annual renewal due later this year.
3. Google Made a Decent Product Photo Cost Fifteen Sen
On 30 June, Google released Nano Banana 2 Lite, the cheapest tier of its Gemini image model family, at US$0.034 per 1K-resolution image — about RM0.15 at current rates — produced in around four seconds. Alongside it, Gemini Omni Flash arrived for fast multimodal work. The family now stacks three tiers: Lite for speed and volume, 2 for generalist work, Pro for the top end.
For a Malaysian retailer or F&B operator, this is the news of the week even if it never makes a local headline. A cost of roughly fifteen sen per image, at four seconds each, prices out a full week of Shopee thumbnails, Instagram carousels, TikTok product shots and WhatsApp broadcast graphics for less than a plate of chicken rice. Whether an owner touches the tool directly or through Canva, the marginal cost of a passable product image just fell close enough to zero that "we don't have creative for this SKU" stops being a defensible reason not to list it.
Closer to Home: Malaysia
Putrajaya had a busy week. On 29 June, Prime Minister Anwar Ibrahim launched Malaysia Digital 2030 (MD2030) — a national digital action plan built on seven strategic thrusts (government, economy, infrastructure, talent, society, trust and security, innovation). The headline targets by 2030: 30% of GDP from the digital economy, 500,000 high-value digital jobs, RM4.5 billion in public sector savings through digitalisation, and 95% of government services delivered end-to-end online. Free Malaysia Today's opinion column the next day read fair — the diagnosis is right, the execution is the harder half.
On 1 July, MTDC launched SemiconStart Malaysia, the country's first dedicated semiconductor venture-building programme, backed initially by RM10 million with ten startups selected for a year-long cohort and up to RM1 million in incubation support per company. And on 4 July, The Star's business page carried a Xerox survey of 1,033 Malaysian SMEs reporting that 81% had adopted AI and 48% plan to expand its use within a year. Read those numbers carefully — the competitive question in your industry may already have moved from "should we?" to "how well?"
The e-invoice clock did not stop for any of this. Phase 4 businesses (RM1–5 million turnover) still have grace to 31 December 2027, with penalties of RM200–RM20,000 per invoice from 1 January 2028. The RM10,000 individual e-invoice rule is live now. LHDN's February disclosure of 500,000+ non-compliant cases and RM1.4 billion in unreported income has not been retracted; the grace period suspends penalties, not audits.
Singapore, briefly: regional enterprise pricing tends to flow through Singapore-hosted contracts first. Malaysian SMEs on cross-border SaaS will feel the Sonnet 5 price drop through their existing vendors within the quarter.
What This Means for Your Business
1. Rerun the automation you shelved earlier this year
If you looked at agentic AI in January or February and decided the token spend didn't justify the time saved, redo the sum. A workflow that would have called Opus 4.8 twenty times a day — reconciling your bank feed against your ledger, tagging incoming customer emails, drafting supplier chase letters — costs roughly a third of what it did on Monday. The businesses that quietly ship one of these workflows this quarter won't announce it. They'll just have a monthly close that closes a day earlier.
2. Rewrite your image inventory as a workflow, not a project
List the product SKUs, service pages and menu items on your channels where the image is dated, missing or off-brand. That was a two-quarter creative brief. It is now a Saturday afternoon. Set a house style, feed it into whichever image tool your team already uses, and update the top 100 assets in one sitting. Google's price change makes this a maintenance task, not a campaign.
3. Read MD2030 for the incentives, not the vision
The plan is real, but the useful part for an SME today is the funding stack that continues to sit underneath it — the 50% MADANI digitalisation grant up to RM5,000 through MDEC-listed providers, the 50% additional tax deduction on certified AI and cybersecurity training, and the SJPP RM5 billion financing guarantee at up to 80% coverage. MD2030 did not add fresh cash this week. The cash that already exists is still unspent and available.
The Practical Question
Which task did you delay on cost grounds three months ago that is now inside your budget — and which competitor already noticed?
The week's pattern was unusually uniform. Anthropic dropped the price of frontier reasoning. Google dropped the price of image generation. Meta signalled that cloud compute pricing has more room to fall. And Malaysia's most-read business page reported that four in five SMEs say they already use AI. If your working assumption is still that competitors are "figuring it out" alongside you, this week suggests the majority now claim to be doing something with it. Whether they're doing it well is a separate question — but the price of finding out just dropped.
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.