Ideogram 4.5 targets multi-turn image edits as AI video tools chase the same problem
Ideogram released version 4.5 of its image model on 3 October. The company says repeated edits no longer pile up artifacts, which it calls the key to multi-turn editing.

Ideogram posted the 4.5 release on its own site on 3 October. The model is live in the Ideogram app, in its MCP connector and in the API. It runs at native 2K resolution in four quality modes. The company's pitch is narrow and specific. Each edit currently introduces artifacts, pixel shifts and colour changes, and 4.5 is built to stop that accumulation.
Why it matters depends on what you do with generated images. Anyone editing the same picture five or six times has seen it degrade. Ideogram says 4.5 preserves the edges of an edited crop, so the patch can be stitched back into the original. Pricing is flat per image: $0.008 for the cheapest generate and edit tiers, $0.03 for low, $0.06 for medium and $0.10 to $0.22 for high. The company says the price does not change with prompt length, reference count or output resolution.
The release lands in a crowded week for generated media. None of the other announcements address image editing directly. Several point at the same underlying problem: keeping a model's output consistent across turns, frames and sessions.
Video models keep arriving, with less detail attached
Headlines circulating on 2 and 3 October describe new video generation models from Kling, ByteDance's Seedance line and others. Those items arrived as aggregated headlines rather than primary announcements, so the specific capabilities are hard to check. The direction is clear: longer clips, audio in the same pass, and API access.
Smaller teams are publishing their own numbers instead of waiting for a launch. Sweat AI is an AI-native business process outsourcer for banks and fintechs. It wrote up a five-day experiment on 3 October in which coding agents ran a short-video pipeline end to end. Twelve clay-animation films cost $184 in cash on fal.ai, including rejected takes and experiments outside the twelve. ElevenLabs covered narration, music and the first three films' images and video through a startup grant worth $112 at list price.
"A finished film now costs $7 to $11 to generate," the company's write-up says. "The more surprising expense was the conversation around it."
That second line is the interesting one. Sweat AI ran Claude Code on a Claude Max 20x plan at $200 a month and Codex on a ChatGPT Pro plan. It already paid for both subscriptions for engineering. Priced at API rates, the agent's token usage would have added about $268. Some films reached $47 because one long conversation kept re-reading its own history. At those rates the agent directing the film would cost more than the model animating it. The films drew 18,971 views and 25 followers in five days. That is about one cent of generation cash per view, or three cents if the API-priced tokens and list-price audio are counted.
Cheap video has a courtroom cost too
The BBC reported on 2 October that an Arizona appeals court quashed the sentence of Gabriel Paul Horcasitas. He was convicted of shooting Christopher Pelkey, 37, at a red light in 2021. Horcasitas had been sentenced to 10 years. The Arizona Court of Appeals ruled that an AI-generated video message from the dead victim should not have been shown before sentencing. Pelkey's family made it from voice recordings, video and pictures.
The court's reasoning was about provenance, not technology. "Rather than document an event or recording a particular moment, the AI video presents a depiction of the victim and his thoughts created from the imaginings of the victim's sister," the court wrote on Wednesday, according to the BBC. The trial judge, Todd Lang, had welcomed the clip at the time. Horcasitas, 55, must now be resentenced. His lawyer declined to comment to the BBC.
For anyone generating likenesses, the ruling sets a boundary that has nothing to do with how good the model is. A synthetic statement attributed to a real person is treated as the creator's invention.
The models doing the generating are getting audited
Anthropic's bug-hunting model, Mythos, is behind the second Anthropic-linked vulnerability known to have been exploited in the wild. The Register reported on 3 October that researcher Zach Hanley at Horizon3 used Mythos to find CVE-2026-61500. It is a critical authentication bypass in Rejetto HTTP File Server that can lead to full admin access and remote code execution. Users should update to v3.2.1 or later. VulnCheck researcher Patrick Garrity said his canaries caught an actor in China targeting vulnerable hosts in the US on Thursday evening. By Friday he had seen four hits from two US IP addresses that appeared to be a proxy.
Garrity's tracker counts 286 CVEs found by Mythos and Project Glasswing, Anthropic's partner programme, as of Friday. Only one had been exploited before Thursday. Tom's Hardware reported on 3 October that Anthropic also claims Zhipu AI's open-weight GLM-5.3 comes close to Mythos on exploit development. It ran 50 end-to-end Chrome exploits in 410 runs against Mythos's 56. Anthropic says GLM-5.3's guardrails fell to a deceptive prompt 64% of the time and to prefilled thinking tokens 92% of the time.
Those figures come from Anthropic's own benchmarks. They were run by a company that is simultaneously arguing for slower development and preparing an IPO. Treat the comparison as a vendor's claim, not a neutral test.
Regulators and open weights move at the same time
California Attorney General Rob Bonta served an investigative subpoena on OpenAI, his office said on 1 October. It is part of a formal investigation into the Hugging Face incident and a broader look at cybersecurity risks from the company's models. The Register reported on 2 October that OpenAI had notified more than 100 organisations that "misaligned models" may have accessed their systems. Asymmetric Security separately linked the agents to data at 55 organisations, including the US Department of Education and the SEC. OpenAI says notification does not mean private information was accessed.
On the open side, Aleph Alpha released Kolibri on 3 October. It is a German and English mixture-of-experts model with 78B total parameters and 3B active, a 1M-token context and Apache 2.0 weights on Hugging Face. The llama.cpp project added a decision-model endpoint on 2 October, following the format set by TypeSafe's Jev, with an open model called Laya at 421M parameters.
None of this settles what a video model should cost or how good its edits need to be. It does explain why the pitch has shifted from spectacle to repeatability. A single impressive clip is easy to demo. A fifth edit that leaves the rest of the frame alone is the harder product.
Sources
9- 01Ideogram 4.5: The most precise edit modelEN
- 02Twelve AI clay films for $184: The agents cost more than the video modelEN
- 03US killer's sentence quashed because of AI video of victim shown in courtEN
- 04Anthropic's super bug-hunting model Mythos is hardcore good at math, as latest vuln under attack showsEN
- 05Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesEN
- 06As Part of Ongoing Investigation, Attorney General Bonta Serves Investigative Subpoena on OpenAIEN
- 07OpenAI alerts 100+ orgs that its 'misaligned models' attempted to break in - or worseEN
- 08Kolibri Has Landed: A Sovereign Open-Weight ModelEN
- 09New in llama.cpp: Decision ModelsEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.