Weekly AI Tools Roundup: July 21, 2026

What's New This Month

OpenAI: ChatGPT Agent 2 — autonomous multi-step tasks arrive

OpenAI launched ChatGPT Agent 2 (July 9), a major step up from the original Agent Mode. Where Agent 1 was primarily a research-and-summarise assistant, Agent 2 can autonomously execute multi-step workflows: booking flights, comparing products across sites, building simple apps, and managing recurring tasks without prompt-by-prompt guidance. It uses an improved chain-of-thought planner with self-correction — if a step fails, it retries with a different approach rather than handing back to the user.

For marketers and small business owners: Agent 2 can now handle end-to-end workflows like "research competitors, draft a comparison table, and export to Sheets" in one session. The free tier gets 10 agent runs/day; Plus gets unlimited. This is the first ChatGPT feature that genuinely replaces a VA for structured research tasks. Test it with a clear brief and see if the output is usable as-is or needs editing.

Runway: Act-Two — performance capture from any driving video

Runway released Act-Two (early July), a performance capture model that animates a character from a source video — film yourself waving, and a 3D character mirrors the motion. Pricing is 5 credits/second with a 3-second minimum. The model handles facial expressions, body motion, and lip-sync in a single pass. Early community tests show it works best with clearly lit faces and front-on camera angles; profiles and extreme motion degrade quality.

For content creators: Act-Two dramatically reduces the cost of animated explainer videos and social content. You no longer need a motion capture stage or a 3D animator for basic character animation. Combine with Runway Gen-4.5 for background scenes and you have a near-complete production pipeline inside one platform. If you're on the Runway Standard plan, Act-Two credits are included.

Anthropic: Claude Sonnet 5 becomes the default model

Anthropic made Claude Sonnet 5 the default model across all Claude plans effective June 30, 2026. Sonnet 5 was previously an opt-in model; now every Claude.ai user and API customer defaults to it. The launch pricing is $2/M input tokens and $10/M output tokens until August 31, 2026 — the cheapest frontier model for multi-step agentic work. Sonnet 5 shows measurable improvement in long-running coding tasks, tool-use accuracy, and context retention over extended sessions.

For developers and AI-assisted workers: If you've been defaulting to Claude Haiku for cost, Sonnet 5 at this pricing is the better value for anything beyond simple Q&A. The agentic quality improvement is real — fewer hallucinations in multi-step workflows, better code debugging. The August 31 pricing window makes this the cheapest time ever to adopt Sonnet for production workloads.

Adobe: Firefly 4 — Style Match, longer video, commercial-safe training

Adobe shipped Firefly 4 in July with two headline features for professional creative workflows. Style Match lets you upload a reference image and generate new content that closely matches its visual style — useful for brand systems where consistency matters more than novelty. Firefly Video now supports clips up to 60 seconds with improved temporal consistency (fewer glitches across the duration). All Firefly output retains Content Credentials (C2PA metadata) by default.

For enterprise design teams: Firefly 4 is Adobe's answer to "our legal team won't sign off on Midjourney/Runway." The commercial-safe training claim (Adobe Stock + public domain only) is the differentiator. Style Match reduces the need for brand guideline documents — designers can now generate on-brand assets from a single reference. If you're already on Creative Cloud, Firefly 4 credits are included in most plans; no separate subscription needed.

Udio: v3 — stems export, better vocals, genre control

Udio released v3 with three practical upgrades: stems export (download individual instrument tracks — vocals, bass, drums, melody — separately for mixing and mastering), significantly improved vocal quality (fewer artifacts, better language support), and enhanced genre control via a new "style DNA" system where you blend two reference tracks to get a hybrid output.

For music producers and podcasters: Stems export is the feature that makes Udio v3 usable in a real production pipeline. Previously generated songs were locked mixes — now you can layer your own mastering, add live instruments, or remix the stems. The vocal quality improvement removes the main barrier to using Udio for client projects. At $10/month for the Plus plan, this is now a legitimate production tool rather than a demo.

Notion: AI 2.5 — Workspace Mode spans the entire instance

Notion significantly upgraded its AI integration in July with Workspace Mode. Previously, Notion AI could only query within a single page or linked database. Mode 2.5 can now pull context from your entire Notion workspace — all databases, documents, calendar entries, and project pages — to answer cross-project questions with sourced, linked answers. The model also handles multi-database joins (e.g., "show me all tasks assigned to people in the marketing team whose projects are overdue").

For teams using Notion as a knowledge base: This is the upgrade that makes Notion AI feel like a real second brain. Before Workspace Mode, AI answers were shallow because they couldn't see outside the current page. Now you can ask "what's our Q3 content plan and who owns each piece?" and get a sourced answer drawn from your actual project database. If your team lives in Notion, this alone justifies the $8/user/month AI add-on.

Zapier: AI Agents — GA release, natural language workflow building

After a year in beta, Zapier AI Agents hit General Availability in July. The agent layer lets you describe a workflow in plain English and Zapier builds the multi-step automation — including conditional logic, error handling, and multi-branch paths. It integrates with Zapier's full catalog of 7,000+ apps. Pricing: AI Agents are an add-on to existing Zapier plans starting at $29/month for 1,000 agent runs.

For non-developers who need automation: This is the most accessible multi-step agent builder available. If you've been manually copying data between Notion, Sheets, and Gmail, describe the workflow to Zapier AI in natural language and it builds it. The conditional logic and error recovery that previously required a developer are now one-prompt. Start with a high-frequency manual task and test whether the agent handles edge cases reliably before expanding to more complex workflows.

Fable 5 returns globally after 19-day suspension

Anthropic's Fable 5 model returned to all users worldwide on July 1, 2026, following a US Department of Commerce export control order imposed June 12 after Amazon researchers discovered a jailbreak technique. Fable 5 is available again on Claude.ai, the Claude Platform API, Claude Code, and Claude Cowork for users in every country. Two structural changes accompanied the return: a new safety classifier trained specifically to block the jailbreak technique (stops it in over 99% of attempts), and a billing change — Fable 5 requires usage credits after July 7 for most subscription tiers.

For developers who were locked out: Fable 5 is back and the safety classifier is an improvement. The billing cliff (July 7) means API costs have changed — check your Anthropic console for updated rates. If you were using Opus 4.8 as a fallback during the suspension, be aware that some legitimate coding queries are now being rerouted to Opus due to false positives on the new classifier. Test your production workflows to confirm outputs are still correct.

Kimi K3 — Chinese frontier model matches GPT-5.6 and Fable 5

A Chinese lab shipped Kimi K3 in early July — the first Chinese frontier model to match the performance of Fable 5, GPT-5.6, and Kimi's own prior releases. Independent benchmarks show Kimi K3 competitive on reasoning, math, and code generation tasks. The model is available via API and in the Kimi consumer app. The launch is notable because it ends the period of absolute Western frontier model dominance and adds a credible third option for API customers who care about latency in Asian markets.

For API customers: Kimi K3 is not yet as widely integrated as OpenAI or Anthropic offerings, but if you serve an Asian user base, the latency advantage is real. Watch for third-party wrappers (LiteLLM, OpenRouter) to add Kimi K3 support in the coming weeks. Pricing appears competitive with Claude Sonnet 5 at launch rates.

Why This Matters for Creators

• AI agents are crossing the research threshold. ChatGPT Agent 2, Zapier AI Agents, and Claude Sonnet 5's agentic improvements all arrived within the same month. Autonomous multi-step execution is moving from demo to usable — test one of these on a real recurring task this month.
• Video is splitting into "visual generation" and "performance capture." Runway Act-Two doesn't generate video from text — it generates animation from performance. That's a different workflow entirely, and it opens motion content to creators who can't animate or hire animators.
• Audio and music production are getting stems-level control. Udio v3's stems export means AI music is now part of a professional pipeline, not just a curiosity. Combined with ElevenLabs for voice, you have a full audio production suite for under $30/month.
• Enterprise compliance is becoming a feature. Adobe Firefly 4's commercial-safe training + C2PA credentials, plus Fable 5's return with a new safety classifier, show that compliance is a competitive differentiator — not an afterthought.
• Notion AI is finally useful for teams. Workspace Mode transforms Notion from a page-level summarizer into a genuine knowledge-base query engine. If your team lives in Notion and pays for the AI add-on, this update makes that spend justifiable.

What to Watch Next

• OpenAI Operator web agent update — expected imminently after July's model launches.
• Hugging Face Open Video Generation Leaderboard v2 human evaluation tier — now live, use it to shortlist open-source video models.
• Google Gemini 3.5 Flash adoption — launched at I/O 2026, seeing wider integration across Google products.
• Claude Sonnet 5 pricing — launch rate $2/$10 per M tokens expires August 31. If you're evaluating, test now.
• Adobe Firefly video public beta — expected H2 2026. Enterprise teams: start the application process early.

StigStack reviews AI tools independently. Some links may be affiliate links — we only recommend tools we've personally evaluated. Last updated July 21, 2026.