AI tools for creatives: July 2026

In July 2026 the conversation around AI moved from what the tools can generate on screen to how the tools are run, priced, and regulated behind the scenes. OpenAI and Anthropic pushed deeper into IPO territory, DeepSeek finally shipped V4 in full, and the EU AI Act's transparency rules stopped being theoretical and started being enforced (on 2nd August to be precise).

In my studio testing this month, I ran these updates against live client briefs to separate the genuinely useful from the merely loud. The headline finding: cost and governance are as important as raw capability. Model tiers got cheaper, usage limits got renegotiated, and the labelling of AI-generated content became a legal requirement rather than a nice-to-have.

Meanwhile, the video model wars intensified, Framer's agents matured rather than expanded, and Google folded NotebookLM into a much bigger ambition it now calls Gemini Notebook.

Each month I work through what has actually shipped and filter it for what matters in a creative practice. You can find the June 2026 edition and May 2026 edition in the archive.

Main points

• OpenAI launched ChatGPT Work on 9 July alongside three restructured model tiers (Sol, Terra, Luna); this turns the assistant into an autonomous workplace agent rather than a standard chat window.
• Anthropic released Claude Opus 5 on 24 July at roughly half the cost of Fable 5; this provides creative studios with a near-frontier default for daily workflows.
• DeepSeek released V4-Flash on 31 July under an open MIT licence while retiring legacy API aliases on 24 July; this forces automated pipeline updates across commercial applications.
• The European Commission began enforcing the AI Act's core transparency rules on 2 August; this makes watermarking and synthetic media disclosures an immediate compliance requirement.
• Midjourney launched V8.2 on 24 July with a redesigned personalisation engine and expanded reference pools; this makes locking in consistent brand styles significantly easier.
• Google rebranded NotebookLM as Gemini Notebook on 16 July while adding secure cloud compute; this expands the platform from a summarisation tool into an active data analysis engine.

New this month

These tools or integrations entered the market explicitly during the July cycle:

• ChatGPT Work – OpenAI's workplace agent, launched 9 July for handling multi-step office tasks
• Claude Opus 5 – Anthropic's proactive, near-frontier model at roughly half the cost of Fable 5
• Grok 4.5 – xAI's flagship coding and agentic model, released broadly on 8 July on the V9 foundation
• DeepSeek V4 (GA) – general availability release on 31 July, replacing the preview-era Flash and Pro aliases
• Midjourney V8.2 – aesthetics and personalisation-focused update, live from 24 July
• Gemini 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber – three new efficiency-tuned models for scaling agentic workflows
• Lyria 3.5 / Flow Music – Google's next-generation music model with stronger vocal quality
• Gemini Notebook – NotebookLM's rebrand and expansion into Search and the Gemini app with a secure cloud computer

General AI tools

ChatGPT

[Major update] OpenAI's biggest move of the month was ChatGPT Work, launched 9 July, which shifts the product from a chatbot into a workplace agent capable of completing tasks across connected tools. The desktop app was also redesigned to separate Chat and Work modes clearly. Towards the end of July, OpenAI cut pricing on the lower tiers considerably: GPT-5.6 Terra now costs 20% less and Luna 80% less, with a new Fast mode replacing the old Priority Processing add-on.

Beyond Work, OpenAI quietly tightened social features: from 9 July, users on web, iOS, and Android could no longer create new group chats or join one via invite link, a rollback that suggests the company is refocusing away from casual social use and back towards productivity. A new Activity view was also added to the desktop sidebar to track recently engaged chats.

Claude

[Major update] Anthropic's headline release was Claude Opus 5, a proactive model that Anthropic positions as achieving performance close to Fable 5 at roughly half the cost. It is now available through Amazon Bedrock and the Claude Platform on AWS, with particular gains in coding, agentic workflows, and cybersecurity tasks.

Access to Claude Fable 5 and the newer Claude Mythos 5 was formally restored on 1 July, closing out the export-control suspension that began in June. Anthropic also extended a 50% higher weekly usage limit promotion on Claude Code through 19 July across Pro, Max, and Team plans, and expanded Voice Mode with support for additional languages and deeper Gmail and Slack integration for paid users.

Perplexity

[Updated] Perplexity unveiled iOS pre-orders for a Comet device and introduced new agent model options, including Claude Sonnet 4.6 and Gemini 3.1 Pro, giving users more routing choice for research-heavy tasks. Comet remains free across iOS, Android, Windows, and Mac following its March paywall removal, and continues to be my preferred starting point for scanning competitive landscapes or breaking down long creative briefs.

Gemini

[Major update] Google's July slate was unusually dense. Three new efficiency-tuned models, Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, arrived for developers scaling agentic workflows at lower latency and cost. Gemini Robotics ER 2 launched as Google's most capable embodied reasoning model, aimed at developers building physical-world agents rather than creative tooling directly, but it signals where the underlying research is heading.

Closer to daily creative use, Gemini Spark was expanded to more regions and gained the ability to handle complex web errands, using your logged-in accounts and saved passwords, with permission, to do things like schedule viewings or research and start flight bookings. Google Vids also picked up Gemini Omni and personal avatars, letting you generate and edit clips with everyday language and create a custom digital avatar to appear in your own videos.

Microsoft Copilot

[Still relevant] No major standalone Copilot announcements landed in July, but the Claude access for Microsoft 365 licence holders introduced in June continues to bed in. Copilot remains most useful for teams already embedded in Microsoft infrastructure.

DeepSeek and Grok

[Major update] DeepSeek's V4 family went generally available on 24 July, with the DeepSeek-V4-Flash-0731 checkpoint and MIT-licensed open weights hitting Hugging Face a week later on 31 July, alongside a live API. The full V4-Pro remains in preview, with general availability unconfirmed but widely expected in August. Crucially, DeepSeek retired its legacy deepseek-chat and deepseek-reasoner API aliases on 24 July, so anyone still calling those endpoints needs to migrate immediately.

xAI shipped Grok 4.5 broadly on 8 July, its flagship coding and agentic model built on the V9 foundation, priced at $2 per million input tokens and $6 per million output tokens. It is a meaningful step up for real-time social monitoring and coding-adjacent creative tooling on the X platform.

Creative and visual tools

Adobe Firefly

[Updated] After June's major Creative Agent expansion, July was a quieter consolidation month for Adobe. The public beta of AI Assistant across Photoshop, Premiere, Illustrator, InDesign, and Frame.io continued to bed in, with the assistant's chatbot connectors now reaching ChatGPT, Claude, and Microsoft Copilot, alongside Nano Banana 2 Lite as a selectable generation model inside Firefly. Google Gemini and Slack integrations remain on the roadmap rather than live.

Studios already testing the private-beta Firefly studio with persistent Elements and Projects folders should expect continued refinement rather than a fresh feature drop this month; treat July as a bedding-in period before the next major Adobe announcement.

Midjourney

[Major update] Midjourney released V8.2 on 24 July, its most aesthetics-focused update of the year. The release brings a substantially reworked personalisation system built around a new profile approach and a much larger reference image pool, alongside a dedicated Style Explorer and Style Creator for Niji users. Background training on V9 continues, aimed at resolving lingering body coherence issues that have dogged the platform since V8.

Canva

[Updated] Canva's July release focused on brand governance rather than generation. Style Match now transforms uploaded photos into on-brand graphics instantly, Brand Kits gained smarter approval workflows and dedicated brand video storage, and Brand Templates can now be published in a single click. A new AI Post Writer generates social captions directly from a finished design, and the editor now supports attaching images to comments for clearer client feedback loops.

Ideogram and other design tools

[Still relevant] Following June's MCP integration, Ideogram's V4 model continues to be adopted by third-party platforms; Recraft added Ideogram V4, Krea 2, and Cosmos 3 as selectable models in its own changelog this month, underlining how quickly these text-accurate generation models are becoming interchangeable middleware rather than standalone destinations. Ideogram remains my top choice for any visual generation where clear, readable typography is essential.

The remaining specialised layout utilities maintain their positions in the workspace:

• Recraft for precise, controlled commercial vector artwork, now with Ideogram V4 and Krea 2 available as selectable models
• Krea for real-time visual experimentation and fluid canvas-based conceptual prototyping
• Khroma for generating custom colour palettes based on trained human aesthetic selections
• AutoDraw for converting quick digital drawings into clean, predictable icons
• Fontjoy for balancing typographic contrast to build instant font pairings

Video AI tools

To keep the studio hierarchy current, here is how the major video model updates from July stack up:

Runway

[Major update] Runway shipped Agent Skills, letting you build an entire ad campaign, generate a commercial, or localise existing ads with a single slash command rather than a chain of separate prompts. The company also launched an AI model router, positioning itself as an infrastructure layer that sits above individual generation models rather than competing purely on its own model quality. On the API side, Gen-3 Alpha Turbo and Gen-4 Aleph were retired on 30 July in favour of Gen-4.5 and Aleph 2.0, and Nano Banana 2 Lite was added to paid plans for fast image generation.

Google Flow

[Major update] Google introduced Lyria 3.5, its newest music generation model, bringing significant gains in musicality, lyrics, and vocal quality to Flow Music. Separately, Google Vids picked up Gemini Omni for natural-language video generation and editing, plus personal avatars that let creators appear as a custom digital double in their own clips.

Kling and Luma

[Still relevant] Neither platform issued a major model version in July, with Kling 3.0 continuing to hold its position as the benchmark for physical motion consistency and multi-shot sequencing, and Luma's Ray 3.2 remaining the strongest option for multi-keyframe control. Both are worth revisiting once their next major versions land, expected later in Q3.

Pika and Seedance

[Still relevant] Pika continues to serve short-form, social-native production well, with no major update this month. Seedance 2.0 Mini remains the fastest option for high-volume workflows accepting mixed reference inputs, and continues to rank strongly against Kling and Runway on independent Elo benchmarks for joint audio-visual generation.

Music and audio AI tools

Suno and Udio

[Major update] The single biggest story in music AI this month landed on 31 July, when the Munich Regional Court ruled against Suno in a case brought by GEMA, Germany's music rights society. The court found that Suno's models had "memorised" copyrighted songs closely enough to reproduce them, ruling this an infringement of reproduction rights under German law and ordering Suno to disclose revenue so damages can be calculated. Suno disagrees with the ruling and is expected to appeal to the Munich Higher Regional Court. The judgment is not yet final or enforceable, and the exact damages figure will only be set after appeal routes are exhausted.

In the US, Suno separately asked a federal court to reject a bid by Universal Music Group and Sony Music to add over 61,000 recordings to their existing infringement case, while the Udio case saw a New York judge deny a similar expansion attempt by Sony. Both sides in the Suno case are expected to move for summary judgment on fair use imminently, a ruling that will likely set precedent across the entire AI music sector.

I would treat any commercial use of Suno or Udio output for client work as higher risk than it was a month ago. Keep monitoring the Suno and Udio legal tracker at AI Vortex closely over August, since the US summary judgment ruling could land at any time.

ElevenLabs

[Updated] ElevenLabs continued its infrastructure-first approach in July. The older scribe_v1 transcription model was fully removed on 9 July as planned, and the API gained a Music Finetunes system, letting you train a custom music model from uploaded audio via five new endpoints. Backup LLM cascade latency was also cut from 8 to 4 seconds, and single-use batch authentication tokens were added for speech-to-text workflows.

Hume AI

[Still relevant] Hume AI published new research on human-quality voice AI in July, continuing to refine its Empathic Voice Interface and Octave text-to-speech engine. Pricing remains transparent and tiered, from a free plan through to a $500 per month Business tier, and it stays my top recommendation for character-driven audio products where emotional nuance genuinely matters.

UI design AI tools

Figma

[Major update] Figma's big July release closes a long-standing gap between design and code: auto layout now behaves much more like CSS, removing the workarounds developers previously needed at handoff to make built versions match the canvas. Existing frames stay on the legacy behaviour unless you opt in, while new frames adopt the updated version automatically; the legacy option remains available until January 2027.

Figma also improved how code-backed screens import onto the canvas: colours, type, and spacing now bind to existing variables rather than landing as hardcoded values, and more imported frames arrive with auto layout applied automatically, cutting down on manual cleanup after bringing coded or legacy screens back into Figma.

Google Stitch and v0

[Updated] Stitch picked up a useful addition in July: it can now analyse an existing design and generate its missing states automatically, useful for quickly filling out hover, empty, and error states without manually specifying each one. The platform remains free and Gemini 3-powered. The v0 platform had no major announcement in July and continues to be an industry standard for turning design intent into production-ready React and Next.js code.

No-code builders

Framer

[Updated] Rather than a big-bang release, Framer spent July maturing the agents introduced in June's 3.0 launch. The July update gives Agents support for Page Effects, accessibility layer attributes, richer overlays, CMS video insertion, text shadows, and draggable interactions, alongside better handling of longer conversations and code changes. Claude Opus 5 is now available inside Framer, completing navigation-building benchmarks around 1.5 minutes faster than Opus 4.8 while using half the credits of Fable 5 for the same task.

Replit, Cursor, and Lovable

[Updated] Replit went through a significant pricing overhaul in July, though the practical trade-offs against Cursor remain similar to prior months: Cursor's Composer 2 model continues to read up to 15 files in parallel before making edits, useful for larger codebases. Lovable and Cursor both continue to ship monthly, incremental improvements to prompt understanding and deployment options rather than headline features this month.

Local and open-source models

DeepSeek V4-Flash (Open Weights)

[New] DeepSeek-V4-Flash-0731 became officially available on 31 July 2026 with open weights on Hugging Face under an MIT licence, alongside a live API. It shares the same architecture and size as the earlier preview build but was re-post-trained, giving a meaningful quality bump without a parameter count change. This is now the most permissively licensed frontier-adjacent open model available for commercial studio use.

Kimi K3

[Updated] Moonshot AI's Kimi K3 is now the largest open-weight model publicly released, with free weights published on Hugging Face in late July under a modified MIT licence. It runs at roughly 2.8 trillion parameters with a 1M-token context window, and the main BF16/MXFP4 checkpoints are in the ~600GB–1.4TB range, putting it firmly in enterprise-GPU territory rather than anything remotely laptop-friendly.

Local tools infrastructure

Studio engineering frameworks continued their steady incremental improvement:

• Ollama reached v0.4.20 in July, with new memory scheduling supporting mismatched multi-GPU setups
• LM Studio continues to connect directly to Hugging Face's full model catalogue, showing VRAM estimates before download
• Hugging Face remains the default distribution point for open weights, including the new DeepSeek and forthcoming Kimi K3 releases
• A new companion app called Bionic shipped mid-July, turning a local open model into an agent that edits files, runs scripts, and transcribes voice locally via Mistral's Voxtral, all from a sandboxed, zero-retention workspace

Novel and niche creative tools

Rather than chasing generic text-to-image features, several specialised tools continued to gain traction by executing distinct creative tasks unusually well:

• Storique builds complete narrative books illustrated in a distinct watercolour aesthetic
• Higgsfield provides an experimental video platform used by independent animators for surreal motion concepts
• Kittl AI transforms rough sketches into complex vector layouts with texture overlays for print production
• Workbeaver takes multi-step control of desktop and browser environments to automate asset sorting and metadata migration

Research and education utilities

Gemini Notebook (formerly NotebookLM)

[Major update] On 16 July, Google renamed NotebookLM to Gemini Notebook and gave every notebook a secure cloud computer, letting it write and execute code natively for complex data analysis grounded directly in your own sources. This unlocks entirely new output formats, including interactive charts and spreadsheets built from your research materials, and notebooks now sync across the Gemini app and Search.

Khanmigo and Cursor for students

[Still relevant] Both platforms held steady in July with no major announcements. Khanmigo remains a carefully designed AI tutoring product built to guide self-study rather than replace it, while Cursor's educational programme continues offering verified students a full year of Pro access.

Worth checking out

The following developments are worth keeping on your radar heading into August:

• Runway Agent Skills – slash-command production of full ad campaigns and localised assets
• ElevenLabs Music Finetunes API – train a custom music model from your own uploaded audio
• Figma's CSS-aligned auto layout – closes a long-standing design-to-code handoff gap
• DeepSeek V4-Flash open weights – MIT-licensed, now generally available for commercial use

July 2026 industry trends

Four distinct patterns emerged across the creative technology landscape this month:

• Governance becomes a live cost: the EU AI Act's transparency enforcement and the GEMA ruling against Suno both turned compliance and licensing risk into immediate, practical concerns rather than distant regulatory noise.
• Consolidation of infrastructure, not just canvases: Runway's model router and DeepSeek's aggressive open licensing show platforms competing to become the layer other tools are built on top of, rather than standalone destinations.
• Cheaper frontier access: price cuts across ChatGPT's lower tiers and the arrival of Claude Opus 5 at roughly half of Fable 5's cost both point to intensifying competition on cost-per-token, not just capability.
• Maturation over reinvention: several major platforms, including Framer and Figma, spent July refining agents and workflows introduced in prior months rather than launching entirely new categories of tool.

Ethical AI and sustainability reports

EU AI Act enforcement begins

From 2 August 2026, the European Commission's AI Office and national authorities began enforcing the AI Act's core provisions, alongside new transparency rules requiring AI systems to disclose when users are interacting with AI and when content has been generated or altered by it. Providers of generative AI systems already on the market have a four-month grace period to implement the required detectability measures, though the broader compliance deadlines for high-risk systems have been pushed out to December 2027 and August 2028 under the AI Omnibus regulation approved by the Council of the EU on 29 June. For commercial design studios, the practical takeaway is that AI disclosure and content labelling are no longer optional extras from August onward. Non-compliance with the AI Act's duties, including transparency, can trigger fines of up to €15 million or 3% of worldwide annual turnover, with even higher ceilings for prohibited practices, so disclosure and labelling need to be treated as core compliance work rather than a cosmetic overlay.

The Suno ruling and music licensing risk

Beyond the direct commercial risk outlined above, the Munich court's finding that specific songs are "reproducibly contained" within Suno's models raises a broader question for any studio using generative audio commercially: if training data can be shown to persist inside a model's parameters, the fair-use defence that much of the generative AI industry has relied on becomes considerably shakier. Worth watching closely as the US cases progress through summary judgment.

Environmental audits continue

Following June's UN University research on data centre resource consumption, no new major environmental audit landed in July specifically for creative AI workflows, though the underlying projections, up to 945 terawatt-hours of annual electricity consumption by 2030, continue to inform how studios choose between cloud rendering and local processing. The Jisc AI and environmental sustainability report (PDF, 1.2MB) (opens in a new tab) from June remains the most relevant detailed reference for creative and education sector workflows.

Previous editions

• AI Tools for Creatives – June 2026
• AI Notes – May 2026
• AI Tools for Creatives – April 2026
• AI Tools for Creatives – January to March 2026
• AI Notes – September, October, November 2025
• AI Notes – August 2025
• AI Notes – July 2025

Researched and written by John Luba. If you spot any changes, get in touch.