Blog
News & Insights
Guides, model updates, and perspectives on AI-powered development.
Bring Better Context with Tarsk File Workflows
Copy files into a Tarsk project, attach them to chat, and review previews before you send your next request.
Project Task Automations in Tarsk
Turn project tasks into repeatable work that Tarsk can start on demand or on a schedule.
Kimi K3 and the Rise of Open-Weight AI
Moonshot AI released Kimi K3 as the largest open-weight model ever at 2.8T parameters. The model's benchmarks, architecture, and what open weights mean for developers.
Tarsk's macOS Sandbox: Safe by Default
When you tell an AI agent to run a bash command, you're handing it a shell. Tarsk's macOS Seatbelt sandbox provides kernel-level containment for bash and skill scripts, and it's on by default.
Cut Your Token Costs by Splitting Planning From Code
Orchestrate mode uses a smart model to plan your task and cheap models to build it. Most users see 40 to 60 percent lower token spend on multi-file features.
Connect MCP Servers with OAuth in Tarsk
Tarsk now supports OAuth for remote MCP servers, so your agent can sign in to services with your account instead of static tokens. Here is how to add Sentry from the marketplace.
Microcompaction: Cut Mid-Turn Token Cost
Older tool outputs used to ride every later model call in an agent turn. Microcompaction digests that stack once context gets tight, so you pay less and finish longer tasks.
Tarsk's Recommended Models: July 2026
Four major model launches in ten days. Claude Sonnet 5, GPT-5.6 Sol, GPT-5.6 Luna, and Grok 4.5 each changed the cost-to-capability math. Here's how to pick.
Save Money and Tokens With Compression
The tool calls in your coding session account for up to 55% of your token spend. Here is how you can reduce that cost and speed up your responses using compression.
Tarsk updates: dual chat, voice, skills
Tarsk updates this week include dual chat, faster voice input, a new skills catalog, an integrated terminal, browser element selection, and a long list of workflow fixes.
New faster streaming voice input
Tarsk now uses Vosk for voice input. It streams your text in real time as you speak -- no pause, no wait. Runs locally, works offline, supports 20+ languages.
GLM-5.2: Open-Weights Model Challenges Claude Opus on Coding
Zhipu AI's GLM-5.2 is a 744B open-weights model with MIT license and 1M context that beats GPT-5.5 on coding benchmarks at one-quarter the cost of Opus 4.8.
What's New in Tarsk: Mid-June 2026
MCP server editing for local and remote, model selector grouped by provider with price sorting, long-running agent confirmation, folder projects, and a floating review chat in the diff view.
Claude Fable 5 in Tarsk
Anthropic's Mythos-class model ships with 1M context and long-horizon agent skills. Eleven Tarsk providers list claude-fable-5 today, from Anthropic and Bedrock to OpenRouter and GitLab Duo.
NVIDIA Nemotron 3 in Tarsk
Nemotron 3 ships as Nano, Super, and Ultra: open MoE models for agent workloads. Nine Tarsk providers already list them, including OpenRouter and Together AI for Ultra.
What's New in Tarsk: June 2026
Settings got a navigation overhaul, context usage now breaks down by category, chat drafts persist across tabs, and you can preview markdown or open your repo in Cursor with one click.
Local Voice-to-Text: Privacy in Your Browser
Tarsk now supports voice-to-text. It runs Whisper entirely locally in your browser using WebAssembly, keeping your spoken words completely private.
NVIDIA Nemotron-3 Ultra: A Very Large Model
NVIDIA launched Nemotron-3 Ultra at Computex 2026. It has 550 billion parameters, uses Mamba-2 layers, and does not cost as much as you might think.
Introducing Claude Opus 4.8
Anthropic's most capable model yet brings adaptive thinking, production-ready code generation, and long-running agentic workflows to developers everywhere. Here's what it means for your workflow.