Mistral’s Trillion-Parameter “Le Chonk” Goes Live at Half Price — Free Weights in Three Weeks
Mistral turned on a trillion-parameter model this week and started charging full preview price for something it plans to give away for free in three weeks. Large 4, nicknamed “Le Chonk” inside the company, went live on Mistral’s own platform on October 6 at a 50%-off launch rate; the open weights are due around October 27.
The same week, Amazon shipped a model built to do one thing — hold a real-time spoken conversation — and xAI quietly dropped its cheapest video-generation tier yet. None of this came from Anthropic, OpenAI, or Microsoft, the three names most people reach for when a week like this happens.
Put together, the three moves point at where AI money is actually moving right now: toward open weights, toward voice, and toward volume pricing — not just toward a smarter chatbot.
| Story | Lab | What it means | Date |
|---|---|---|---|
| Nova 2.5 Sonic | Amazon / AWS | Voice-agent infrastructure | Oct 5, 2026 |
| Grok Imagine Video 1.5 Lite | xAI | Cheaper AI video | Early Oct 2026 |
| Large 4 “Le Chonk” LEAD STORY | Mistral | 1T-param open model | Oct 6, 2026 |
Amazon Built an AI That Only Talks
A speech-to-speech model aimed squarely at voice agents, live now in four AWS regions.
AWS announced Nova 2.5 Sonic on October 5 — a speech-to-speech model purpose-built for voice agents, with sharper reasoning than its predecessor. It skips the usual pipeline of speech-to-text, then a separate text model, then text-to-speech, and handles the spoken exchange directly. It’s live now in Amazon Bedrock in four regions: US East (N. Virginia), US West (Oregon), Europe (Stockholm), and Asia Pacific (Tokyo).
- Live in Bedrock in 4 regions now
- Built specifically for voice, not chat
- Phone support can sound and respond more naturally without stitching together three separate AI systems
- It’s developer infrastructure, not a plug-and-play consumer app
xAI’s Cheapest Video Yet
A new budget tier for AI video — with one competitor claim worth testing yourself before you trust it.
xAI quietly added a “Lite” tier to Grok Imagine, its video-generation model, in early October. Per xAI’s own developer docs, it prices at $0.02 per second at 480p, $0.03 per second at 720p, and $0.14 per second at 1080p — the company’s lowest per-second video rate so far. One industry analysis claims it undercuts Google’s Veo 3.1 by roughly two-thirds at similar quality; that comparison comes from a third party, not xAI, so treat it as a claim to test, not a settled fact.
- Official per-second pricing at three resolutions
- xAI’s cheapest video tier to date
- A 15-second 720p clip runs about $0.45 — cheap enough to generate several versions and compare
- The Veo price comparison is unverified; don’t plan a campaign on it alone
Mistral built one of the largest open-weight models released this year — and started charging for early access to it before a single outside benchmark could confirm it’s any good.
Mistral’s Trillion-Parameter “Le Chonk”
Live now at half price. Free to download in three weeks.
On October 6, Mistral announced Large 4 — internally nicknamed “Le Chonk” — a 1-trillion-parameter model that only activates 52 billion parameters at a time, thanks to a mixture-of-experts design that keeps it fast despite its size. It’s natively multimodal, live now in preview on Mistral’s own platform at 50% off for two weeks, and was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs that Mistral owns and runs itself in Europe. Mistral is promising full open weights around October 27. Other labs have shipped comparably large open models this year too — Moonshot’s Kimi K3 and DeepSeek’s V4 Pro both claim more active or total parameters in their own configurations — so “one of the largest” open-weight models is the fair description here, not “the largest.” Mistral’s own benchmark numbers (61.7% on DeepSWE v1.1, 93% on Cybench) are self-reported and haven’t shown up yet on independent leaderboards like Artificial Analysis; Mistral has also cited different benchmark sets in different venues, which is one more reason to treat any single number here as a claim rather than a verified result.
- Live preview now on Mistral’s own platform at 50% off
- Trained on 3,800 GPUs in Mistral’s own European data centers
- Open weights this large change who can afford frontier-scale AI without a cloud subscription
- Once the free weights land Oct 27, the API discount stops being the only way in
Three Things Worth Pricing Out Before the Weekend
Test Mistral Large 4 before the discount ends — but don’t treat it as a now-or-never window. The 50%-off preview opened October 6 and runs two weeks, closing around October 20, a week before the open weights go public on roughly October 27. If you want to try frontier-scale AI on a side project without renting a GPU cluster, the discounted API is the easy way in right now; once the weights are public, you or a developer you hire could self-host it for free instead, on your own hardware or cloud bill.
If you’ve been putting off automating phone support, pilot Amazon Nova 2.5 Sonic. It’s already live in Bedrock in four regions, so a small business with a developer on retainer — or a freelance contractor who knows Bedrock — can prototype an AI phone line or voice assistant this week instead of waiting for a wider rollout.
Price out one real marketing clip on Grok Imagine Video 1.5 Lite before signing up with a bigger video vendor. A 15-second clip at 720p runs about $0.45 on xAI’s published rates — cheap enough to generate three or four versions of a product teaser or social ad and compare them side by side, rather than taking any one competitor’s price comparison on faith.