AI Tools Worth Trying This Week
It’s been an unusually loaded few days for AI releases — several labs pushed out major updates within 72 hours of each other. Here’s what’s actually worth your time, and why.
1GPT-6 Astra— OpenAI
OpenAI’s newest flagship launched on September 3 and is rolling out across ChatGPT Plus, Pro, Business, and Enterprise tiers, plus the API and AWS Bedrock. The headline pitch is “computer use” — Astra is built to actually operate software on your behalf: filling out forms, updating CRM records, running QA checks on a website, or working through a multi-step research task without much hand-holding.
Why try it: If you’ve been frustrated by agents that stall out on long, multi-step workflows, Astra’s context-handling changes are worth a look — it now keeps searchable notes across a session instead of compressing everything into a single summary when the context window fills up, meaning fewer dropped details on long debugging or research sessions.
Worth knowing: Because of its jump in cybersecurity-relevant capability, some of Astra’s more sensitive features ship gated behind OpenAI’s enterprise security program rather than being available to everyone at launch.
2Claude Fable 5.1 & Mythos 5.1— Anthropic
Anthropic’s latest models landed September 1, with Fable positioned as the flagship for coding and knowledge work. Early independent benchmarking has it ahead of the prior generation on several fronts.
Why try it: If your work is mostly writing, research, or coding rather than autonomous computer operation, this is the more natural upgrade path — built around the same conversational and coding strengths Claude has been known for, just sharper.
Worth knowing: Access to these models briefly paused in June due to export-control requirements before being restored; Anthropic has been transparent about the timeline if you want the details.
3Gemini 3.8 Flash— Google DeepMind
Google quietly shipped an update to its fast, low-cost Flash tier on September 2, alongside a separate “Cyber” variant aimed at proactive security use cases for enterprises and governments.
Why try it: Flash models are the ones most people actually build on day-to-day — cheap, fast, good enough for most tasks. If you’re running high-volume workflows like chatbots, summarization, or classification, it’s worth swapping in and checking whether the price-to-performance ratio moved in your favor.
4OpenCode— open source
Not a model — a coding agent. OpenCode runs from a terminal, desktop app, IDE extension, or browser, and lets you pick your own model provider and API key rather than locking you into one vendor.
Why try it: If you like the idea of an agentic coding assistant but don’t want to be married to a single company’s model, this is the more flexible option. It’s grown fast — reportedly past 120,000 GitHub stars this year — with a free core tier and a paid usage-based plan for heavier use.
5Qwen3.8-Max— Alibaba
Alibaba’s Qwen team pushed out an updated Max-tier model in the same window, continuing the pattern of open-weight and semi-open models closing the gap with the closed frontier labs. Worth a look if cost and self-hosting flexibility matter more to you than chasing the very top of the leaderboard.
The bigger picture
A few things are worth noticing across all of these releases:
- Computer-use agents are becoming the actual product, not just a chatbot feature. Astra is the clearest example, but the trend shows up across labs.
- Cybersecurity capability is now a gating factor. Several of this week’s models are powerful enough that vendors are restricting their most sensitive features to vetted enterprise or government programs.
- Fast and cheap tiers matter as much as frontier ones. If you’re building something real, the Flash- and Mini-tier updates are often more relevant to your bottom line than the flagship headlines.
Bottom line
If you only have time to try one thing this week: pick Astra or Fable 5.1 for deep work and coding, Gemini 3.8 Flash if you’re optimizing for cost at scale, and OpenCode if you want an agent that isn’t tied to one vendor.