AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.

OCR had a week. Mistral dropped OCR 4 with bounding boxes. Baidu open-sourced a model that beats DeepSeek-OCR. Claude got a permanent home inside Slack. And the Fable 5 ban fallout keeps getting uglier: Alibaba was apparently stealing Claude’s capabilities, and even the NSA lost access to Mythos. Meanwhile, GPT-5.6 is delayed to mid-July. Let’s go.

1. Mistral OCR 4: document AI gets serious

Mistral launched OCR 4 this week. It’s not just another OCR model. It’s a full document understanding system with paragraph-level bounding boxes, confidence scores, and support for 170 languages.

The specs:

• $4 per 1,000 pages (standard), $2 per 1,000 pages (batch)
• Paragraph-level bounding boxes with coordinates
• 72% win rate in blind tests against competitors
• Available on la Plateforme, Microsoft Foundry, and self-hosted for enterprise
• Top score on OlmOCRBench

Why this matters for developers: Bounding boxes change everything. Previous OCR models gave you text. Mistral gives you text + where it is on the page. That unlocks document search, compliance systems, and any workflow where page structure matters.

My take: At $4/1000 pages, this is competitive with Google Document AI ($5) and significantly cheaper than building your own pipeline. For enterprise document processing, this is probably the best option right now. For budget-conscious developers, Baidu’s free alternative (see below) is worth considering. Full comparison in our Mistral vs DeepSeek vs Baidu breakdown.

2. Baidu open-sources Unlimited-OCR

While Mistral went commercial, Baidu went open. Unlimited-OCR is a 3B-parameter MIT-licensed model that processes multi-page PDFs in a single inference pass.

Key features:

• Built on DeepSeek-OCR architecture (SAM+CLIP + DeepSeek-V2 MoE decoder)
• Reference Sliding Window Attention for memory efficiency on long documents
• Tables to HTML, equations to LaTeX, layout to bounding boxes
• Private by design: nothing leaves your device
• GGUF, MLX, NVFP4 quantizations already available

My take: For a 3B model you can run on a laptop, this is remarkably capable. It won’t match Mistral OCR 4 on complex enterprise documents, but for invoices, receipts, forms, and standard PDFs, it’s more than good enough and it’s free. The fact that Baidu explicitly positions it as “pushing DeepSeek-OCR one step further” tells you where the open-source OCR race is heading. See our local setup guide and open-source OCR comparison.

3. Claude Tag: always-on AI teammate in Slack

Anthropic launched Claude Tag, a persistent Claude identity that lives inside Slack channels. Think of it as an always-on AI coworker rather than a chatbot you have to DM.

How it works:

• Admin grants Claude access to selected channels
• Anyone in the channel can @Claude to delegate tasks
• Claude accumulates context across days (persistent memory per channel)
• Connects to tools, data, and codebases configured by admin
• Available for Enterprise and Team customers (beta)

Why it’s interesting: This is Anthropic’s play for enterprise sticky revenue. Once Claude becomes embedded in your team’s daily Slack workflow with accumulated context about your projects, switching costs become enormous. It’s the same playbook Notion and Slack used: make the tool part of daily muscle memory.

My take: This is less about technology and more about business model. Claude Tag turns Claude from “a tool employees open sometimes” into “a teammate that’s always there.” For the comparison with Microsoft Copilot and ChatGPT’s Slack integration, see our full comparison.

4. Alibaba caught extracting Claude capabilities

Reuters reported that Anthropic accused Alibaba of “illicitly extracting” Claude AI model capabilities. The timing is not subtle: this came days after the US government banned Fable 5 access for foreign nationals.

What it means: The Fable 5 export ban now has a clearer backstory. If Chinese companies were systematically extracting capabilities from Claude (likely through distillation or structured prompting to replicate behavior), that explains why the government moved so aggressively.

My take for developers: This doesn’t change anything practical for you. But it does confirm that the US/China AI divide is deepening. If you’re building on closed US models, plan for the possibility that access restrictions expand. If you’re building on open Chinese models (GLM-5.2, DeepSeek V4), understand that the geopolitical baggage comes with them. There’s no clean answer here.

5. NSA lost access to Mythos amid the ban

The New York Times reported that the NSA was using Claude Mythos 5 and lost access when Anthropic disabled it under the export control directive. The US government’s own ban affected its own intelligence agency.

The irony: The Commerce Department banned Fable 5 and Mythos 5 to protect national security. In doing so, it apparently cut off the NSA from a tool it was actively using for national security purposes.

My take: This is government dysfunction, not a developer story. But it does suggest the ban was hasty and poorly coordinated. Which means it might get revised. Watch for a carve-out that restores government access while keeping the foreign national ban in place.

6. GPT-5.6 delayed to mid-July

After weeks of “launching Monday” predictions, GPT-5.6 has been pushed back. Prediction markets now put it at 83% chance of delay beyond June 28, with a new target of mid-July. Traders have abandoned their late-June bets.

What happened: The June 23 launch date came from leaked Codex log traces and prediction market speculation, not from OpenAI itself. OpenAI never confirmed a date. The model appears to exist (traces in internal systems) but isn’t ready for public release.

My take: Don’t hold your breath. When it drops, we’ll cover it. Until then, GPT-5.5 remains the best OpenAI model available. If you were waiting for GPT-5.6 to start a project, don’t.

7. EU selects EUROPA consortium for frontier AI

The European Commission selected the EUROPA consortium to build Europe’s first open-source frontier AI model. The specs: 400B+ parameters (MoE), all 24 EU languages, open weights, AI Act compliant.

This won’t matter for 12-18 months (the model doesn’t exist yet), but it’s strategically significant. Europe is now officially building its own frontier model as a response to US export controls. See our full landscape overview.

Quick hits

• OpenAI custom chip — first custom silicon built with Broadcom. For training efficiency, not inference speed. Won’t affect developers directly.
• Sakana Fugu Ultra — 1M context model on OpenRouter at $0.000005/token (essentially free). Worth trying for massive context tasks.
• MiMo UltraSpeed benchmark — we published our 106-session comparison. TL;DR: 37% faster sessions, 86% higher median throughput, same output quality.
• AI Race: GLM declares itself done — the first agent to explicitly recognize it can’t do more without human help. Built 140 pages, got every distribution channel. Still $0. 9 days left.

What I’m watching next week

• GPT-5.6 status — delayed but apparently close. Mid-July most likely.
• Fable 5 ban resolution — the NSA embarrassment might force a policy revision
• Race finale countdown — 9 days to July 3 deadline. Will any agent earn $1?
• OCR market shaping up — Mistral (commercial) vs Baidu (open) vs DeepSeek (cheap API). Who wins developers?

---

AI Dev Weekly publishes every Thursday. Subscribe for the newsletter version.