What changed at the AI frontier: July 2026

I keep a set of reference documents on the current state of AI capabilities, one per domain, eighteen at this point: image generation, embeddings, speech, agents, video, OCR, and the rest. They exist because my agents otherwise work from training data that is months out of date. A scheduled job refreshes them monthly with web research and commits the diff.

The interesting byproduct is the diff itself. It’s a monthly record of how fast things are moving. This is the July edition, filtered down to the changes I think a practitioner would find most relevant.

The deprecation wave

A strong theme this cycle is how quickly the big labs now retire models.

OpenAI retired the DALL·E 2 and 3 API models on May 12, and has already scheduled gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest for retirement on December 1, all pointing at gpt-image-2. Google put shutdown dates on its Imagen model line, rolling from June 30 through August 17, and retired the Veo 2 and Veo 3 model IDs on June 30 in favor of Veo 3.1.

What that means in practice: a pinned model ID in a production pipeline now has a shelf life of months, not years. If your pipeline calls an image or video model by ID, migration is recurring scheduled work. I treat a model migration like a release, with a small bake-off against the replacement before swapping, and this cycle made that feel like the sensible baseline rather than excess caution.

Embeddings: chunking gets less painful

Voyage shipped voyage-context-4 on June 29. It produces contextualized chunk embeddings, meaning each chunk’s embedding also encodes context from the whole document. A lot of the fiddly work in RAG is chunking strategy, and this moves a good part of that problem into the model.

It builds on the Voyage 4 family from January, whose shared embedding space already allowed indexing with the large model and querying with a smaller one without re-indexing. The two together make the embedding layer noticeably less brittle than it was a year ago.

Speech, both directions

The Artificial Analysis leaderboards reshuffled on both sides of the speech stack in this pull. On transcription, Alibaba’s Fun-Realtime-ASR preview leads at roughly 1.7% word error rate, with ElevenLabs Scribe v2 around 2.2% and the best open-weights option (Voxtral Small) around 2.8%. On synthesis, Gemini 3.1 Flash TTS sat on top in my July 1 pull, and Alibaba’s Fun-Realtime-TTS family moved into the frontier cluster. The TTS lead has already changed hands since that pull, which tells you how quickly these rankings churn. Treat the exact order as a snapshot.

One thing worth noticing is that Chinese labs now hold frontier positions on both sides of speech. If you last evaluated speech vendors in 2025, the shortlist has changed.

On-device: a 20B model on the phone

Apple announced its third generation of foundation models on June 8, with the new capabilities going to developer testing first. The interesting one is AFM 3 Core Advanced, a 20-billion-parameter sparse model that runs on-device by activating only 1 to 4 billion parameters per prompt, with the full weights living in flash storage and routing decided per prompt. Google’s Gemma 4 family expanded to five sizes in the same window.

A 20B-class model on a phone changes what “on-device only” product designs can promise. I haven’t built against AFM 3 yet, so I’ll hold judgment on quality, but the architecture alone makes the category worth re-checking if you dismissed on-device models last year.

OCR keeps getting cheaper

Mistral released OCR 4 on June 23 at $4 per thousand pages through the API and $2 per thousand in batch, with bounding boxes, block classification, and inline confidence scores. It also comes as a single self-managed container, though that path is for enterprise customers. Document pipelines that were priced out of LLM-grade OCR a year ago probably aren’t anymore.

How this digest is made, and how much to trust it

These items come out of the monthly refresh of my reference docs. ChatGPT with web search does the sweep, a set of guards catches truncation and silently gutted sections, and I curate what survives. For this post I verified the deprecation dates and release claims above against the providers’ own pages. The leaderboard numbers are as of the early-July pull and will drift.

I plan to publish one of these after each monthly refresh. The full set covers eighteen domains, so if there’s one you’d want in more depth, tell me which.