Mistral AI releases Large 4; DeepMind announces EmbeddingGemma 2

#1 𝕏

Mistral AI released Mistral Large 4 (Le Chonk), a natively multimodal model with 1T parameters and 49B active, via API. The company claims leading open-weight benchmark performance and plans an open-weights release for the end of October.

Also covered by: @Mistral AI, @Julien Chaumond, @clem 🤗

#2 𝕏

Google DeepMind announced EmbeddingGemma 2, its first natively multimodal open model for on-device embeddings. Expanding beyond text, it unifies code, images, audio, and video in a shared space.

Also covered by: @Google DeepMind, @Philipp Schmid, @Sundar Pichai

#3 𝕏

Claude now works inside Google Docs, Sheets, and Slides, and files from those apps can also be opened in Claude. In Google Workspace, Claude appears in a sidebar, reads the open file, and can edit it in place; users can approve each edit before it is applied.

#4 𝕏

Cursor announced that its iOS app lets users control agents on their computer from their phone, including checking in, replying, and starting new tasks.

Also covered by: @Cursor

#5 𝕏

Philipp Schmid says Nano Banana 2.1 (`gemini-nano-banana-2.1`) is now available at $0.034 per image versus $0.134, reportedly outperforming the previous Pro model at 4x cheaper pricing. It improves instruction following and in-image text rendering, uses Google Image Search grounding, supports consistency for up to five characters and 14 reference images, and is available through Google AI Studio, the Gemini API, Google Cloud, the Gemini app, and Flow by Google.

Also covered by: @Logan Kilpatrick

#6 📝 Anthropic News

Expanding the Cyber Verification Program - Anthropic News: On October 6, 2026, Anthropic launched an expanded Cyber Verification Program with three access tiers, giving qualifying security professionals access to advanced cyber capabilities and reduced blocking classifiers.

Also covered by: @Anthropic

#7 in

Dan Shipper says Every released Every Agent, an agentic coworker in Slack for long-running tasks, with pre-installed Every workflows and personalized Frontier Alerts. The release includes $15 in free credits per person, up to $500 per workspace; afterward, access requires a $30/month Every subscription, with tokens billed at cost and zero markup.

Also covered by: @Claude

#8 📝 OpenAI News

Sharing AI progress in mathematics - OpenAI News: On October 6, 2026, OpenAI released a broad range of new mathematical results produced by an unnamed internal frontier model, with the work available on GitHub.

Also covered by: @OpenAI

#9 𝕏

claire vo 🖤 demonstrated a YouTube thumbnail workflow using the Decisions API with vision to select her best face from 40 minutes of footage, associating the process with a $0.13 cost. Her post quoted an OpenAI Decisions API announcement.

#10 𝕏

Santiago commented that Ampersand can connect an agent to a customer’s Salesforce, applying configured object and field mappings while handling API calls, authentication, token refreshes, and retries. He highlighted the pain of working with two different Salesforce accounts.

#11 𝕏

Madhu Guru said teams often mistakenly treat evaluations as an extra QA step after building an agent; for AI products, evaluations are the product specification.

#12 𝕏

bolt.new shared that users can ask Bolt to configure Sentry in a GDPR-safe way, storing no network addresses, replacing emails with internal IDs, and recording sessions only when something breaks. The post says this keeps bugs visible while users remain anonymous.

#13 𝕏

Garry Tan described his gbrain workflow as chaotic, spinning up a new thread for each bug or issue. He called Capy’s cross-thread coordination and “virtual standup” conflict avoidance state of the art, adding that GStack’s autoplan and eli5 help him work quickly.

#14 𝕏

Aravind Srinivas announced that Perplexity cut Perplexity Decision API pricing by half, to 2 cents per million input tokens.

#15 📝 OpenAI News

Atlassian and OpenAI expand their partnership - In an October 6, 2026 post, OpenAI said Atlassian and OpenAI are expanding their partnership to bring frontier models in the GPT‑6 family across Atlassian’s platform, powering AI experiences intended to help teams plan, build, and deliver work.

#16 𝕏

Santiago commented on Gumloop’s infrastructure-hosted browser, highlighting vault-based account management, persistent profiles, 1Password integration, live activity viewing, session replays, and reusable workflows.

#17 𝕏

Guillermo Rauch commented on a Vercel post about its AI Gateway, describing slower thinking under a confidence threshold as a simple feature that could be extremely impactful for AI decision-making at scale.

#18 𝕏

Tal Raviv recapped his 40-day “can Fable grow my food” experiment, in which he gave Claude Code three tools—a water pump, lamp, and camera—and used approximately 50 million tokens. He said he did not eat any radishes. He shared a deep dive into what worked, what disappointed, and open questions, along with the open-sourced transcripts and Raspberry Pi filesystem.

Also covered by: @Tal Raviv

#19 𝕏

DeepLearning.AI recapped how Xiaomi addressed test-only training that led a model to add unrequested code and let errors pass silently by multiplying each test result by quality checklist scores. Xiaomi’s MiMo-V2.6-Pro RL leads open-weights models on Artificial Analysis’ Intelligence Index with a score of 46.

#20 𝕏

NVIDIA shared how AI agents can investigate telecom network problems, compare fixes, and test them in simulation. NVIDIA Nemotron supports agent reasoning, while NVIDIA OpenShell controls what agents can access and do.