The week's releases were all about what gets read

The week's releases all piled onto the intake path, the unglamorous layer that decides what a model actually reads before it answers. Three document parsers shipped inside the same week, Mistral OCR 4 and Baidu's Unlimited OCR among them, all of them turning scanned pages into text a model can use. A security write-up also renamed prompt injection, the trick where text a model is reading talks it into doing something, as role confusion, which describes the failure better. We made the case a while back that the practical results were coming from small models running beside the data, and document reading is the clearest version of that yet: the assistant is the interface, and something small and specialized does the reading. If the next frontier model swallows the reading job, nobody buys a separate parser. Ask again after somebody runs a thousand scanned invoices through both.

Top Stories#

Three document parsers shipped inside a single week. Mistral published OCR 4, pitched at document intelligence, Baidu released Unlimited OCR with a paper and a repository, and a smaller parser appeared as Infinity-Parser2-Flash. Three releases in a week is what a contested category looks like. Getting a table of numbers out of a PDF has never been one model's job anyway. It is a pipeline, four or five stages long, and the license attached to each stage decides what a product can actually ship. Quality wins the demo, then legal reads the license and picks what ships.

OpenAI previewed GPT-5.6 Sol. The company described it as a next generation model ahead of availability, routine enough now that the interesting question is what a preview does to the layer below it. A frontier preview sells capability, while the rest of this week was selling unit cost on narrow jobs, where a task-specific model can be orders of magnitude cheaper than a general one for pulling the same fields off the same invoice. Capability and unit cost do not really compete until the general model gets cheap enough to take the narrow job too, and nobody has tested Sol on that.

Prompt injection got a more useful name. A write-up recast the problem as role confusion: a model has no reliable way to tell an instruction from content that merely looks like one, so a sentence arriving as data can end up speaking with the authority of the operator. The renaming matters because it points at where a fix has to live. Filtering for bad strings treats the problem as spam, which works about as well as spam filtering ever has. Treating it as roles means carrying provenance through the pipeline, so that everything a parser pulled out of a document stays marked as something the model reads rather than something it obeys.

Recursive calls are being sold as a context strategy. A plug and play library for recursive language models landed for anyone who wants a model to call itself over pieces of a problem instead of holding the whole thing in one window. Calling a small model many times over parts is the opposite instinct to buying a longer context window. Reading everything at once is expensive and tends to blur what matters, while a recursive pass makes the reading order something a program decides. That looks like the more useful primitive for research and knowledge agents, though it trades a context bill for an orchestration problem.

Top Stories#

Top Links#