Bigger, Smaller, Faster31 posts
← Back to InsightsAI Trends

Bigger, Smaller, Faster

November 14, 2026·7 min read

This week, new AI models came with new sticker prices — and most of them went down.

On 9 November, Anthropic released Claude Fable 5.2 — its new flagship — at $8 per million input tokens and $40 output, a 20% cut from Fable 5.1, taking the #1 spot on Artificial Analysis's Intelligence Index v5 at 65. The same day, Meta shipped Muse Ember at $0.10 input and $0.50 output per million tokens, matching Claude Haiku 5.5's top score at roughly a third the cost on Meta's own benchmarks. On 10 November, ElevenLabs released Scribe v3 and Scribe v3 Live, cutting word errors 27% on recorded audio and 31% on realtime across 102 languages, with a Singapore data-residency option for enterprise. One flagship cheaper. One small model cheaper still. Speech sharper. All in two days.

Here's what happened, why it matters, and what your business should do about it.


The Big Three

1. Anthropic releases Claude Fable 5.2 — and trims the price

On 9 November, Anthropic shipped Claude Fable 5.2 at $8 per million input tokens and $40 per million output — 20% below Fable 5.1, and the first Fable release to cut its list price. It took the #1 spot on Artificial Analysis's Intelligence Index v5 at 65, generates output up to 30% faster than 5.1, and keeps the same 1M-token context window. Anthropic also released Claude Mythos 5.2, a restricted sibling for approved cybersecurity and biomedical research on the same underlying model.

For the last two years, every new frontier model cost at least as much as the one before it. This is the first visible crack in that pattern at the top of the market.

So what: When the flagship tier starts cutting prices, the pressure rolls downhill. The mid-tier AI tool your sales team pays RM500 a month for does not run on Fable; it runs on something cheaper that competes with other cheaper things. Ask whoever bills you whether they've repriced since November. Several will not have, and the answer tells you how competitive your vendor's own pricing really is.

2. Meta launches Muse Ember — a small model that undercuts Haiku

The same day, Meta released Muse Ember at $0.10 per million input tokens and $0.50 per million output — the same headline rate as Claude Haiku 5.5 and GPT-6 Luna. Meta's own numbers show Ember matching Haiku 5.5's top score for roughly a third of the cost per task, and scoring 37.8% on AutomationBench (a business-workflow benchmark) versus Haiku 5.5's 13.9%. Ember has a 1M-token context, up to 512K output, and is pitched at exactly the work SMEs most want to automate: tool calls, routing, extraction from documents, and small coding fixes.

The caveats matter. These are Meta's own numbers, measured on its own setup, using a benchmark composite rather than independent tests. Expect them to shift when third parties run their own measurements.

So what: The cheapest capable model keeps getting cheaper and more capable. For any SME considering AI for a repetitive job — chasing invoices, drafting WhatsApp replies, pulling numbers out of a supplier PDF — the question "which model do I need?" is no longer "the best one I can afford". It is closer to "the cheapest one that actually works for this task". If your vendor can't tell you which model sits behind their product, that in itself is useful information.

3. ElevenLabs ships Scribe v3 — speech-to-text sharp enough for mixed-language audio

On 10 November, ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for realtime. Scribe v3 cuts word errors 27% versus Scribe v2 across 102 languages. Scribe v3 Live cuts semantic errors 43% versus the previous realtime version, and adds speaker diarization — knowing who said what as the conversation is happening — for the first time in realtime. ElevenLabs also added a keyterm library for product names and jargon, a multilingual mode for conversations that switch languages mid-sentence, and isolated EU, India and Singapore environments for enterprise data-residency.

Malaysian meeting audio — Bahasa Melayu, English, Mandarin and Tamil often in the same conversation, often from the same speaker — is exactly the kind of messy input transcription has historically failed on. The numbers are vendor-reported, not independent. The accuracy floor has still shifted.

So what: If you have been quietly waiting for voice AI (meeting transcripts, call summaries, a receptionist bot) to be accurate enough for Malaysian audio, the gap narrowed this week. Before committing, run your own recordings through it for a day — not the demo file — and check the speakers are being tagged correctly. Price is the other variable; weigh the accuracy gain against the per-minute delta on your actual audio.


Closer to Home: Malaysia

Two stories close to home.

First, on 3 November, LHDN disclosed at its Innovation and Integrity Day that 2.14 billion e-invoices have been submitted by more than 269,900 taxpayers as at 2 November 2026. The system now has enough data to do what LHDN has already confirmed it is doing: run analytics on every submission. The compliance clock has not moved. The RM1 million+ turnover mandate is live. Every transaction at or above RM10,000 still requires an individual e-invoice today, not in 2027. Phase 4 businesses (RM1–5 million turnover) remain inside the grace period that ends 31 December 2027, with penalties of RM200–20,000 per invoice from 1 January 2028. LHDN's February figures — 500,000+ non-compliant cases, RM1.4 billion unreported income — still describe the audit environment.

Second, TikTok Shop's standalone app arrived in Malaysia. Apple's store data gives its first release as 27 October 2026; TechCrunch reported it on 5 November, and TikTok's Malaysian newsroom had nothing on it as of 6 November. The listings showed US dollar prices in the screenshots and no ratings around launch. The functional change is small — same orders, same catalogue, a new entry point for shoppers who want to buy without scrolling the feed — but the direction is not: TikTok Shop is moving from "a feature inside TikTok" to "its own app you keep on your phone." Shopee still led Malaysia at 83% penetration to TikTok Shop's 64% in Ipsos's May 2026 survey. The gap keeps closing.

Singapore note: ElevenLabs' Singapore data-residency environment matters for Malaysian businesses only if a compliance team has specifically asked where audio data sits.


What This Means for Your Business

  1. Audit what you pay for AI against this week's price curve. Not whether your current tool "works" — that is the easy part. Whether what you pay matches the market. If Muse Ember, Haiku 5.5 and GPT-6 Luna all headline at $0.10 / $0.50 per million tokens, any vendor quoting a figure that assumes a much higher per-call cost is either running a bigger model than they need to, or marking up heavily. Both are fair grounds for a renegotiation.
  2. Run one cheap model against your most boring task. Not the customer-facing one — the one that quietly consumes the most staff hours. Quotation drafts, invoice chases, WhatsApp routing, line-item extraction from a supplier PDF. If a $0.10 / $0.50 small model handles it well enough, that is where to shift spend. If it does not, you now know exactly what the premium models buy you, which is a stronger position than guessing.
  3. Use November for e-invoice hygiene, not AI pilots. The grace period is a window to clean up past failures, not a reason to postpone compliance work. Confirm every ≥RM10,000 transaction is e-invoiced and that your accounting system is talking cleanly to MyInvois. A backlog LHDN surfaces through analytics in 2028 will cost more than any productivity gain a new AI tool delivers this quarter.

The Practical Question

Which of your current software tools would you still choose, at full price, if a direct competitor cut its price by 20% tomorrow?

That is the pressure the week's news just put on every vendor pitching you. The ones who respond to market price cuts by holding their own prices are telling you how locked in they think you are. The ones who move with the market are telling you something else. Either answer is useful. The one most SMEs never get is silence from a vendor they never asked.


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.

Talk to us →