14 noviniek
Mistral released OCR 4, a structure-aware document AI model with bounding boxes, block classification and confidence scores across 170 languages. Pricing is $4 per 1000 pages, available via Mistral API, Amazon SageMaker, and same-day on Microsoft Foundry.
Google has quietly pushed Gemini 3.5 Pro general availability from June to July 2026, citing tester feedback on token efficiency and long-horizon task performance. The delay puts the model head-to-head with GPT-5.6 and Claude Opus 4.7 in mid-July.
Google released DiffusionGemma, an experimental 26B MoE model (3.8B active) with a diffusion head that generates text in parallel blocks rather than token-by-token. NVIDIA optimized it for RTX GPUs and DGX Spark.
OpenAI on June 22 moved GPT-5.5-Cyber from preview to full release (85.6% on CyberGym), expanded Codex Security, opened a partner program with CrowdStrike/Sophos/Fortinet and launched the open-source 'Patch the Planet' initiative.
OpenAI partnered with Trail of Bits on 'Patch the Planet,' using Codex Security tools to help open-source maintainers find and patch vulnerabilities at scale.
New arXiv paper combines supervised fine-tuning with GRPO and claims a 3B model beats Claude Opus 4.5 on reasoning benchmarks. HN top: 305+ score, 162 comments.
Prosus built ToqanClaw on its in-house Toqan platform — data stays under user control and isn't used to train third-party models. Targets Prosus's 5M+ merchants and restaurants.
Japan-based Sakana AI launched Fugu, a multi-agent orchestrator that coordinates GPT, Claude and Gemini behind a single OpenAI-compatible API and claims to beat Fable 5 in benchmarks.
Apertus, the fully open foundation model from EPFL, ETH Zürich, and CSCS, hit the top of Hacker News over the weekend. It is Apache 2.0, trained on 15T tokens across 1,000+ languages, and engineered to meet EU AI Act requirements including PII removal and anti-memorization controls.
Sakana AI unveiled Fugu — a multi-agent system exposed as a single API that dynamically orchestrates top frontier models per task. Reflects the shift from 'one big model' to routing/orchestration.
Agents can now ship a Worker without a human sign-up. A temporary account spins up automatically, lives 60 minutes, and the human can later claim it. Cloudflare is stripping OAuth and dashboard friction out of agentic pipelines.
Author benchmarks GPT-5.5 against the MIT-licensed open GLM-5.2 on the AA-Omniscience hallucination benchmark and finds GPT-5.5 hallucinates roughly 3× more (86% vs. 28%) — both tested under identical OpenRouter conditions.
At Build in San Francisco, Microsoft unveiled MAI-Code-1-Flash — its first in-house coding model, aimed at reducing dependency on OpenAI and lowering developer costs.
Hard safety blocks in Fable 5 refuse legitimate security work; researchers say the model is too restrictive even for benign red-team tasks.
