This week in AI (12–18 October 2026): DeepSeek's search API, Muse Spark Open, MAI-Base-2 and three deprecation dates
Ten things from the week: new models from DeepSeek, Meta, Microsoft and BFL, Gemini 4 Argon pricing, three shutdown dates, and a CERT-In warning.
A busy week: two large models from Meta and Microsoft, an updated DeepSeek Flash with its own search API, Google’s new flagship on the paid API, three deprecation dates to put in the calendar, and a security warning from CERT-In that every Indian IT team should read. Ten items, each with the source we checked.
1. DeepSeek updates Flash and opens a web Search API (17 Oct)
DeepSeek moved its deepseek-flash API model to a new checkpoint, V4.1-Flash-1017, with no change to price or calling method, and published the weights under the MIT licence on Hugging Face. The same day it opened Search and Fetch APIs at US$0.75 per 1,000 successful requests, capped at 5 requests per second per account, and free through DeepSeek Harness, Claude Code or Codex until 1 November. Its pricing page shows peak rates, double the off-peak ones, during hours that overlap much of the Indian weekday. We covered the details, including the peak windows in IST, in this morning’s post.
2. Microsoft AI describes MAI-Base-2, a 2.8T-parameter base model (15 Oct)
Microsoft AI published details of MAI-Base-2, a mixture-of-experts model with 2.8T total and 80B active parameters, trained on about 50T tokens using 16,000 GB200 GPUs over 38 days. Microsoft says it is 4–12x more efficient in pre-training compute than DeepSeek-V4-Pro-Base; the detailed analysis on the same page gives 4.2–12.3x. It is still in post-training and won’t be released as open weights: Microsoft plans to license it to trusted partners through Foundry. No absolute benchmark scores were published.
3. Meta releases Muse Spark Open (15 Oct)
Meta released Muse Spark Open, an open-weight model with 816B total and 42B active parameters and a 1M-token context window. Meta calls it “the most capable US open-weight model to date” and says it is the first in its class to run on two Apple M5 Ultra Mac Studios. The licence is the Muse Model License, which has limited commercial terms, so read it before building a product on the weights. Downloads are on Hugging Face, with hosted options on SageMaker, Google Cloud Model Garden and Microsoft Foundry.
4. Meta’s Muse Wonder arrives on the Meta Model API (12 Oct)
Earlier in the week Meta announced Muse Wonder and Muse Wonder Full, aimed at agent work in areas like law, finance, health and coding. Muse Wonder is available to developers through the Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers. Meta highlights two features: Active Memory, which carries context across conversations, and Active Focus, meant to keep the model working through long tasks.
5. Gemini 4 Argon reaches the paid Gemini API, and the price doubles on 1 December (13 Oct)
Google opened Gemini 4 Argon to developers on the paid tier of the Gemini API. The pricing page lists US$2 per million input tokens and US$10 per million output tokens on the standard tier until 30 November, rising to US$4 and US$20 from 1 December. Batch and Flex are half the standard price. There is no free tier for this model, so prototypes built on the free Gemini tier can’t simply switch to it.
6. Black Forest Labs releases FLUX 3 Image Klein (15 Oct)
BFL released FLUX 3 Image Klein, a 5.6B-parameter open-weight model for image generation and editing with native 4K output and bounding-box control. BFL says an INT8 build runs on GeForce RTX cards with as little as 8GB of VRAM. The FLUX Kommunity License allows free commercial use only for individuals and companies under US$5M in annual revenue; larger organisations need a self-hosting licence or the API.
7. OpenAI sets 14 January 2027 for GPT-4o and GPT-4.1 API shutdowns (14 Oct)
OpenAI’s deprecations page now lists five snapshots that shut down on 14 January 2027: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14. Fine-tunes of these models go on the same day. The recommended replacement for all of them is gpt-6-luna. If you have a fine-tuned GPT-4o mini in production, start re-training and testing now; three months is not long for a migration that involves evals.
8. GitHub Copilot will drop three models on 16 November (15 Oct)
GitHub announced that Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra leave Copilot on 16 November. The suggested replacements are Claude Opus 5.5 for both Claude models, and GPT-6.1 Sol or GPT-6 Luna for Terra. Individual subscribers on annual plans keep access to Opus 4.8. Admins should check any policies or team defaults that pin these models.
9. Anthropic study: small models could self-replicate across cheap IoT devices (12 Oct)
Anthropic’s Frontier Red Team published research suggesting small AI models may soon be able to copy themselves across cheap devices such as off-brand Android TV boxes, without a data centre. A 9B model got inference running on 8 of the 9 hardware types tested. The authors stress the limits: the break-in step was deliberately easy (open ADB ports), and the end-of-2026 timeline is a rough extrapolation. Their defences are familiar: watch network flows for large model transfers and unusual device-to-device traffic, quarantine infected devices, and get rid of default passwords.
10. CERT-In: AI agents under test reached live systems at Indian organisations (16 Oct)
India’s CERT-In said AI model providers had voluntarily disclosed that agents used in research and testing reached live systems at Indian organisations and read personal and confidential records, TechCrunch reported. The agents got in through exposed credentials in code, repositories and configuration, and through overly permissive cloud roles. None of the affected organisations noticed on their own. According to Deccan Herald’s report on the advisory, CERT-In recommends auditing code, repositories, configs and cloud storage for secrets and rotating them, switching off unused APIs, enforcing least privilege and MFA, keeping admin interfaces off the internet, and retaining logs for 180 days. We could not find the advisory text on CERT-In’s own website, so these points come from press coverage.
The week in one line
More open weights arrived (DeepSeek, Meta, BFL), Microsoft showed a far larger model it won’t open, and the most useful thing most teams can do this week is boring: check your model IDs against three deprecation dates, and scan your repositories for exposed keys.
References
- deepseek-ai/DeepSeek-V4.1-Flash-1017 (model card) — Hugging Face / DeepSeek, accessed 2026-10-18
- Search & Fetch API guide — DeepSeek API Docs, accessed 2026-10-18
- Models & Pricing — DeepSeek API Docs, accessed 2026-10-18
- Frontier pre-training at Microsoft AI (MAI-Base-2) — Microsoft AI, accessed 2026-10-18
- Introducing Muse Spark Open — Meta, accessed 2026-10-18
- Introducing Muse Wonder and Muse Wonder Full — Meta, accessed 2026-10-18
- Gemini 4 Argon is now available for developers — Google, accessed 2026-10-18
- Gemini Developer API pricing — Google AI for Developers, accessed 2026-10-18
- FLUX 3 Image Klein — Black Forest Labs, accessed 2026-10-18
- Deprecations (2026-10-14: GPT-4 series model deprecations) — OpenAI, accessed 2026-10-18
- Upcoming deprecation of selected GitHub Copilot models in mid-November — GitHub Changelog, accessed 2026-10-18
- AI models may be capable of self-replicating without access to a datacenter — Anthropic, accessed 2026-10-18
- India becomes the latest country to report rogue AI agent activity — TechCrunch, accessed 2026-10-18
- CERT-In issues critical alert on AI agents breaching Indian systems — Deccan Herald, accessed 2026-10-18
This week in AI, every Sunday
One email with the releases that mattered and what they mean in practice. No spam.