G'day! Five releases from the last day or so, why each one matters, and how to try it. OpenAI's Images 2.6 is out in ChatGPT and the API, and it can turn a prompt into a looping GIF. Cognition has launched SWE-2 Flash, a faster, cheaper coding model inside Devin, and Ricursive has released Ripple-65B Preview, a downloadable US open-weight model built for agentic coding. There's also a soothing simulator where you restore battered old game cartridges, and Kino, a free AI video editor where your teammates and coding agents share one timeline. About a five-minute read.

Image: OpenAI

OpenAI has released Images 2.6 in ChatGPT, Codex and the API. It says images generate about 50% faster than before, and for the first time you can animate them as GIFs: brand mascots, reaction loops, product shots and storyboard frames from a single prompt. In the API, developers get two flavours: GPT-Image-2.6 Flare for quick drafts (OpenAI says as little as six seconds on the low quality setting) and GPT-Image-2.6 Sunburst for higher detail and tighter consistency across edits. Markup tools that were already in ChatGPT now come to the API, so you can point at the bit you want changed, and designs can export as SVG for websites and print. OpenAI says people now create about 4 billion images a week across ChatGPT, Codex and the API.

Why it matters: image generation is becoming as everyday as chat, and GIFs plus faster edits make it useful for social posts, product mock-ups and early storyboards without jumping to a full video tool. Consistency across edits is the part brands care about most.

TRY IT → free to try in ChatGPT and Codex on every plan (including Free), on desktop, mobile and web, available globally. Developers use GPT-Image-2.6 Flare or Sunburst in the Images API: OpenAI lists both at US$8 per million image tokens in and US$30 per million out (about A$11.42 and A$42.83), with text prompts billed separately; use OpenAI's image cost calculator for a per-picture estimate. OpenAI: Introducing Images 2.6 · OpenAI API pricing (image models)

Image: Cognition

Cognition, the company behind the Devin coding agent, has launched SWE-2 Flash, a smaller American coding model aimed at everyday engineering work such as automations, code review, features and docs. It builds on Thinking Machines' Inkling-Small (276 billion total parameters, 12 billion active). Cognition says a new mid-training stage before reinforcement learning lifted its Terminal-Bench 4.0 score from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%, and that it beats larger US open-weight models on several of its tests. Flash runs at about 200 tokens a second versus about 60 for full SWE-2 in Cognition's demo, and costs one-fifth as much. Weights are not public yet: Cognition says it is building a partner program first, partly because the model was also trained for cybersecurity tasks.

Why it matters: most agent jobs are routine, not frontier research, so a faster, cheaper model inside a product people already use is more useful day to day than another giant scoreboard champ. The cyber training cuts both ways: helpful for defenders, risky if misused, which is why the weights stay gated for now.

TRY IT → available now in Devin Desktop and the Devin CLI. Devin has a Free plan (US$0, light quota and limited models), Pro at US$20 a month (about A$28.55) with higher quotas and free use of SWE-2, and Max at US$200 a month (about A$285.53). Flash is billed at one-fifth the price of SWE-2 inside that usage. Cognition: Introducing SWE-2 Flash · Devin pricing

Image: Ricursive Intelligence (screenshot of the Ripple launch page)

Ricursive Intelligence, a US lab that co-designs AI models and chips, has released Ripple-65B Preview. It is a 65-billion-parameter dense model trained from scratch on 12 trillion tokens, with a context window of up to a million tokens, and it is aimed at agentic coding (long jobs where the model drives a terminal and tools). Ricursive says it scores 80.52 on Terminal-Bench 2.1, edging larger US open-weight models such as Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4), and that the full R&D cost was about US$32 million. The architecture is tuned to run efficiently on NVIDIA hardware, including smaller boxes such as an NVIDIA DGX Spark. Weights are out under Apache 2.0 on Hugging Face, and the model is also on Together AI, GMI Cloud and OpenRouter.

Why it matters: a strong, downloadable US coding model at this size is rare, and the price on hosted APIs is low enough to wire into everyday coding agents. All the headline scores are Ricursive's own, and the Preview name means more post-training is still to come.

TRY IT → free to download (Apache 2.0) from Hugging Face, or use via OpenRouter at US$0.20 per million tokens in and US$0.60 per million out (about A$0.29 and A$0.86). You will need a capable GPU (or a hosted endpoint) to run the full weights locally. Ricursive: Introducing Ripple-65B Preview · Hugging Face: Ripple-65B Preview · OpenRouter: Ripple-65B Preview

Image: Parasol Corp (Steam)

Cozy Game Restoration is a quiet cleaning and restoration simulator from solo developer Parasol Corp (Adam). You unwrap sealed packages, find battered old video game cartridges covered in stickers, markers and grime, and gently bring them back to life with tweezers, swabs, scrapers and cloths. Each cart is a bit different, and the pleasure is in the small physical jobs done well. It supports mouse-only play and controllers, and it ships with a long list of community translations. It unlocked yesterday morning Australian time after sitting on "coming soon" the evening before.

Why it matters: not every new game needs to be loud. This is a cheap, low-stress palate cleanser for anyone who likes ASMR-style cleaning games or has a soft spot for old plastic and faded labels.

TRY IT → out now on PC (Steam) for A$19.99. Windows and Mac, single-player. No Australian classification was listed at check time; overseas age ratings are very low. Steam: Cozy Game Restoration

Image: Kino

Kino (YC S23) relaunched on Hacker News yesterday as an AI media manager and video editor. Footage lives in one library, every clip is transcribed and described, and edits happen on a shared timeline that both humans and coding agents (Claude Code, Codex and similar) can drive. Under the hood, timelines are code built on Remotion, so agents can cut, add motion graphics and answer review comments without a separate tool chain. Review links let teammates leave notes in a browser with no account. Apps are out for Mac, Windows and Ubuntu.

Why it matters: most "AI video" tools generate clips from a prompt. Kino instead puts agents inside a real editor next to your actual footage and your team's notes, which is closer to how edit suites already work.

TRY IT → Free plan forever (US$0): full editor, unlimited imports/exports, collaboration, a monthly Kino chat allowance and 100 MCP tool calls a week; your own Claude/Codex usage is billed by that provider. Pro is US$25 a month (about A$35.69), or US$20 a month billed yearly (about A$28.55), with unlimited MCP calls and higher allowances. Business is US$100 a month per editor (about A$142.76). Kino · Kino pricing · Launch HN: Kino AI

Know someone who'd like this? Forward it on. They can sign up free at https://fivetotry.beehiiv.com/subscribe

Five to Try is curated with AI help and checked by a human. Hit reply and tell us what you tried.

Want your product in front of curious Aussie tech readers? Sponsor Five to Try at [email protected].