Mercury 3 ships a working diffusion model, but an outside benchmark says it loses to Haiku 5.5

Good morning 👋 A diffusion model finally shipped with real pricing and a million-token window. Inception's own benchmark says it beats Claude Haiku 5.5. An independent one says it doesn't.

In today's issue:

• 🔭 Mercury 3 ships a real diffusion model, but an outside benchmark says it loses to Haiku 5.5
• 🧠 ChatGPT gets COBOL evals, Pegasus 1.5 lands in Bedrock, NVIDIA's European newsroom stack grows
• 🔬 Agents skip the simulator's answer, vision models hit a thinking-time wall, image models drop language pretraining
• 🏛️ First PAIPA sanctions hit an AI-linked VC, Brussels catalogs platform scam tactics
• 🔗 Chinese chip orders, a robovan raise, Amazon's German bet, and more

🔭 THE ONE THING

🌀 Mercury 3 ships a working diffusion model, but an outside benchmark says it loses to Haiku 5.5

Inception Labs shipped Mercury 3 yesterday: a diffusion-based LLM, non-autoregressive, with a 1M-token context window (up from 260K in Mercury 2.5) and pricing at $0.20/$0.70 per million tokens ($0.10/$0.35 at launch, 90% off cached input), live on the Inception API, OpenRouter, Baseten, and Vercel. Inception's own numbers say it's 40x bigger than Mercury 1 and beats Claude Haiku 5.5; Bito.ai's independent benchmark says otherwise, scoring Mercury 3 at 76% on one-shot coding against Haiku's 79%, and 65% on long agent sessions against 87% for Opus 5.5 and 77% for Sonnet 5.5. The vendor's claimed 1,068 tokens per second hasn't been checked by anyone outside Inception either. A diffusion model that actually ships, with real pricing and a real context window, is worth watching regardless. But on the one number that matters for deciding what to call in production, trust the outside test over the vendor's own scoreboard.

🧠 MODELS & RELEASES

• 🗄️ OpenAI built new ChatGPT evals aimed at low-resource languages, covering general software engineering, SQLite, and COBOL migration specifically. If you're the one stuck rewriting forty-year-old mainframe code, this is the first real signal the model's being tuned for that job, not just benchmarked on Python and JS. link

• 🎥 TwelveLabs' Pegasus 1.5 landed in Amazon Bedrock. Feed it hours of raw footage and it hands back timestamped, structured JSON for whatever moment you defined as mattering, now with finer video segmentation and multimodal prompting via reference images. That's the unglamorous unlock that makes video search actually work at scale. link

• 📡 NVIDIA and the EBU expanded their sovereign AI stack for European public broadcasters, adding Microsoft's Azure Local, Nebius compute, and NVIDIA's open Nemotron models so newsrooms can run agents without routing editorial footage through someone else's cloud. Adobe's slice, GPU-accelerated search in Frame.io, claims 5x faster embedding generation, and UFA is already using it to search 15 years of archive by concept and character expression. link

🔬 RESEARCH HIGHLIGHTS

• Evidence substitution. A new study ran PHREEQC, the USGS's geochemistry simulator, with reports running up to 847,000 characters, and tested what happens when agents fetch those reports from files instead of keeping them in context, the default pattern now for long tool output. GPT-5.1 lost 11.3 points of accuracy doing it that way, GPT-5.2 lost 7.9. Claude Opus 4.6 and Sonnet 4.6 showed no real decline. The failure mode has a name: the model skips the authoritative simulator output and answers from its own generated numbers instead, rounding errors included. arXiv

• More thinking time doesn't fix visual agents. Across 22 benchmarks and eight model backbones, simply raising a vision model's inference budget bought only modest gains. What actually moved the needle was persistent memory, holding onto images and structured data across a task instead of re-deriving them each step, which is what the paper's OpenVLHarness system adds. That got 9.7 to 12.6 points over no-tool baselines, with one variant hitting its gains at a third of the token cost of rivals. arXiv

• Halcyon drops language pretraining from text-to-image. Trained from scratch on 100 million image-text pairs with no internet-scale language pretraining at all, Halcyon's 1.6B model reaches comparable image quality to an autoregressive T2I baseline using 275x less pretraining compute, and a smaller 0.52B version gets there at 18x less. The assumption that a text-to-image model needs a language-pretrained backbone first looks optional, not load-bearing. arXiv

🏛️ POLICY & REGULATION

• State Department Five Chinese entities, including VC firm MiraclePlus and startup Zhisuan, got hit with PAIPA sanctions this week: asset freezes, transaction bans, a $10M-a-year lending cap, the full stack. The thread runs back to Linwei Ding, the ex-Google engineer convicted in January on economic espionage and trade-secret counts, sentenced in September to a year minus a day and ordered to repay about $198,000. It's the first time PAIPA has been aimed at an AI case, and sanctioning the VC for showcasing the founder's pitch is the part that should worry anyone funding China-adjacent AI startups right now. State Dept. fact sheet

• EU Commission A new DSA study catalogs how financial scammers exploit platform features, AI-generated lures included, and lays out countermeasures platforms are expected to adopt. No enforcement teeth yet. Reads like homework assigned ahead of the next compliance push. DSA study

🔗 QUICK LINKS

• ByteDance, Alibaba, Tencent placed over $35B in orders for Nvidia's RTX PRO 5500, the most powerful chip Chinese buyers can get under export rules, and Nvidia is telling them to expect maybe 5 to 10 percent of what they asked for. Reuters
• Neolix The Chinese robovan maker closed a $400M Series E led by Boyu Capital, with Alibaba's AEF NextGen Fund joining, pushing total funding past $1.4B. Bloomberg
• Amazon pledged more than €16B for Germany in 2027, including a €4.8B AWS Frankfurt expansion and a second Berlin lab that doubles its robotics and AI research footprint there. AboutAmazon
• Deepgram ported its Nova-3 speech model to AWS Trainium2 for batch transcription, swapping GPU queues for a custom Rust runtime and actual cost numbers to back it up. AWS
• Hugging Face shipped TorchNeuron, a PyTorch plugin that makes Trainium act like a normal accelerator instead of routing everything through optimum-neuron's lazy tensors and static shapes. Hugging Face
• Itaú rolled out a GenAI assistant called ia.i plus an internal governance layer, IARA, across one of Brazil's most regulated banking operations. OpenAI

Trust the test you can actually check.

Pradeep Perugu