Black Forest Labs launches Flux 3 Image with multi-step editing
A release-heavy morning: a frontier image model goes live with box-by-box editing that promises to hold everything else still, a 111,000-star coding harness hits 1.0 after reversing itself on MCP, and Perplexity slips an open-weights decision model out the same week the category caught fire.
Black Forest Labs has released Flux 3 Image, the image half of its Flux 3 family, as a general-availability API built around multi-step editing. The workflow is the pitch: you draw bounding boxes over the regions you want changed, edit one box at a time across several passes, and BFL claims everything outside the box comes back untouched — its own demo marks the surrounding areas as locked and unchanged, though no independent benchmark has tested that claim yet. The model handles text-to-image, image-to-image, text rendering and photorealism, takes up to ten reference images for composition, and outputs up to 4K; API access is half-price through October 8, companies can license commercial weights to run and fine-tune it themselves, and an open-weight release is promised in the coming weeks. Why it matters: this lands days after Ideogram 4.5 announced its own editing-focused model with open weights on the way, which tells you where the image-model fight has moved — generation quality is close to table stakes now, and the frontier is precise, localized edits that don't wreck the rest of the canvas. Our take: whoever actually ships reliable region-locking outside a vendor demo wins the next round of this market, so treat BFL's "leaves the rest alone" framing as the claim to test, not the feature to buy on.
Pi, Earendil's terminal coding agent, shipped version 1.0 on October 1 — with MCP support moved into the core, reversing a stance the project defended publicly until days earlier. Until September 29 the project's line was that Pi did not support MCP at all; the team then published a post explaining the about-face, and 1.0 shipped two days later with MCP and its token-saving codemode in the default feature set — codemode alone shrinks prompt overhead by roughly 40%, cutting a typical GPT-5.6 request from about 5,300 tokens to 3,300. The repo sits at around 111,000 stars, the release drew 99 reactions, and the Hacker News thread passed 1,200 points. The signal is less the version number than the capitulation: when a harness this popular decides refusing MCP costs more than supporting it, the protocol has stopped being optional plumbing and become the price of admission.
Perplexity has open-sourced pplx-decider-v1-27b, a decision model fine-tuned from Qwen3.8-27B under Apache 2.0 — one day after Cloudflare put its own Qwen3.8-27B decision models into open release. Perplexity's model card reports 85.71% overall accuracy across 11 benchmarks, against 84.51% for TypeSafe's Jev and 74.76% for the base Qwen checkpoint, with the biggest jump on RAGTruth hallucination detection (88.80% versus the base model's 61.53%); it also concedes losses to Jev on WinoGrande and BBH, so "competitive with" is the honest framing, not "beats." We covered Cloudflare's entry hours earlier — Cloudflare open-sources the Clef models, aiming straight at Jev. With quantizations and community reimplementations already appearing, the decision-model category went from one vendor to a crowd in about a week — and the first casualty of that is the pricing scarcity Jev was built on.
What to watch: whether the official Decision Index adds Decider and reproduces any of these self-reported numbers, and whether BFL's locked-region claim survives independent testing once the open weights arrive.
Is region-locked, box-by-box editing the feature that finally makes generated images production-safe for real design work? Tell us in the comments.
Sources: Pi v1.0.0 release (GitHub)
