AI Startups

Perplexity, Tencent, Slack & fal Ship Agent Tools: AI Startups to Watch for August 30, 2026

Perplexity adds managed connectors to its Agent API, Tencent ships a 770B model, Slack launches one-click agent deployment, and fal turns video into live broadcast.

Daniyal

· 5 min read
Futuristic workspace with holographic connectors linking different tech platforms with glowing blue and purple lines

Perplexity shipped managed connectors for its Agent API, letting agents clone repos and open PRs on GitHub without per-request tokens. Tencent released a 770B-parameter model with 1M context. Slack launched one-click AI agent deployment with 10 partners. fal turned a video model into a live broadcast that generates faster than real time. These startups are shaping how AI agents interact with real-world tools. These are not demos. They are products shipping this week.

Perplexity Adds Managed Connectors to Agent API

Perplexity launched managed Connectors for its Agent API, providing integrations for GitHub, Slack, Google Drive, and Datadog (Perplexity docs). API Group admins connect each service once in the API Portal, then agents reference them by ID without needing per-request tokens.

The GitHub connector integrates with Perplexity's Sandbox environment, enabling agents to clone private repositories, modify code, and open pull requests autonomously. This is the infrastructure layer that makes persistent coding agents practical at scale: no token sprawl, no per-session auth setup.

Perplexity has been shipping at an unusual pace this month. The company added Z.ai's GLM 5.3 to its Computer agent platform, updated its Brain memory system with a 9-point correctness improvement, and launched Portable Computer for local-first AI on NVIDIA DGX Spark. The Agent API, launched August 22, now supports 41 models across 9 providers.

Perplexity has been one of the most active AI startups this quarter. See how other AI startups disrupted the market in yesterday's AI Startups roundup.

Why it matters: Managed connectors solve the auth problem that makes enterprise agent deployments brittle. When an agent needs to touch GitHub, Slack, and a database in one session, per-request token management is the first thing that breaks.

Tencent Ships 770B Hy4 Preview With 1M Context

Tencent released Hy4 preview, a 770B-parameter model with 49B active parameters and a 1M-token context window (Tencent AI). Built for productivity, it targets software engineering, data analysis, and scientific research. The model is available as open-source weights on HuggingFace and via Tencent Cloud and OpenRouter, with pricing starting at $0.834 per million input tokens.

Within hours, the community compressed it to 200GiB using a mixed-precision quantization method called MIX-STQ1_0, which uses calibration data to dynamically assign bit-widths across layers. The quantized model maintained near-original accuracy while reducing size by over 85%. Benchmarks show MCP Atlas 83.2, SWE-Bench Pro 65.0, and HLE-agent 53.1 at the compressed size.

Tencent also made GLM-5.3-Flash available on Fireworks AI this week, after a two-day delay to investigate an unexplained reasoning-length discrepancy. The model features 320B total parameters with 18B active per token and a 1M-token context window.

Why it matters: A 770B model that compresses to 200GiB changes the economics of running large models on-premise. The quantization method works for any mixture-of-experts architecture, not just Hy4.

Slack Launches One-Click AI Agent Deployment

Slack introduced "Add to Slack," a one-click deployment bridge that connects AI creation platforms directly to workspace channels (Slack blog). The feature launches with 10 partners including LangChain, Lovable, n8n, OpenAI, and Vercel.

Agents inherit workspace permissions and data boundaries when deployed through this pathway, giving IT teams centralized visibility and governance. The standardization matters more than the convenience: right now, every agent integration is a custom OAuth flow with custom scopes. Add to Slack makes it a single action.

This pairs with Cloudflare's agent wallets and Perplexity's managed connectors as the infrastructure layer for agent deployment. The pattern is the same across all three: reduce the per-integration friction so agents can operate across tools without custom auth for each one.

fal Turns Video Model Into Live Broadcast

fal launched H3 Max Live, an experimental broadcast where video generates faster than real time and scenes are directed by chat prompts (fal blog). Viewers use the !prompt command to choose upcoming scenes without interrupting the stream.

H3 Max is a post-trained variant of MiniMax's H3 open-weight video model. It generates 5 seconds of 768p video with synchronized audio in under 3 seconds, faster than playback. The long-form checkpoint maintains thread consistency across scenes, which is what enables continuous output rather than disconnected 5-second clips.

fal's regular H3 Max ranked number 1 across overall quality, prompt understanding, and aesthetics in the company's human preference evaluations against competing video models. The API is available now, with 5 free 15-second generations per day at 768p.

Why it matters: Generation time dropping below playback time is a threshold moment for video. The bottleneck shifts from "how fast can we render" to "how good is the creative direction."

Exa Cuts Agent Search Token Usage by 95%

Exa released Dynamic Highlights, a search primitive that jointly scans a set of pages to filter redundant information (Exa blog). Instead of extracting query-relevant tokens from each document independently, Dynamic Highlights processes all documents in a single forward pass, allocating more tokens to better content and fewer to irrelevant documents.

The results: 95% average token reduction compared to full page content, with 40% efficiency gains over standard highlights at a 12k character budget. In agentic benchmarks, the reduction translated to 30% lower agent token usage with a 2.1% average quality improvement.

The feature works via the Exa API by setting dynamic: true. It is a research preview.

Why it matters: Token efficiency is the hidden cost layer in agent deployments. A 95% reduction in search tokens changes the economics of any agent that does multi-page research.

OpenAI Retires DALL-E GPT, Three Pricing Changes Hit Tomorrow

OpenAI removes the official DALL-E GPT from ChatGPT today (AIToolsRecap). Image generation stays. The dall-e-2 and dall-e-3 API models were already removed in May. OpenAI has not clarified what happens to images generated through the old GPT after today. Download anything you want to keep.

Tomorrow brings Claude Sonnet 5 repricing ($2 to $3/M input, tokenizer change adding 10-35% tokens on code), GPT-5.4 and GPT-5.4 mini leaving Codex for ChatGPT sign-in users, and two Kimi models sunsetting in favor of kimi-k3.

Claude Code Limit Boost Expires August 31

Claude Code weekly limits have run 50% above the published standard since May 13 (AIToolsRecap), across Pro, Max, Team, and Enterprise tiers. The boost was announced as temporary and extended three times. It expires August 31.

The arithmetic: removing a 50% increase leaves you at two thirds of the boosted ceiling, so capacity falls by a third rather than a half. If your weekly usage bar regularly passes 67%, you would hit the cap without the boost. Check Settings, Usage.

Related reading from earlier coverage, with the numbers behind each one:

More on AI Startups

Browse every AI Startups story →