Menu

Notes

Free-to-download models catch up: three releases in three days

Three free-to-download models in three days: Meta's Muse Spark Open on 15 October, Reflection AI's Beam on 16 October under the Apache 2.0 licence, and DeepSeek V4.1 Flash on 17 October under the MIT licence. On Vercel's AI Gateway, open models handled more than seven in ten tokens in September.

The capability gap to paid models is now small for routine work. The cost gap is not what it looks like: a free model that needs three times the tokens to finish a job can cost more than a paid one that finishes in one pass. That is what AI for Products' own six-model test showed.

For product people: Open models now handle most everyday jobs. Before you assume they are cheaper, measure the cost of the finished job, as the test above did.

Source: Meta (15 October), Reflection AI (16 October, Apache 2.0) and DeepSeek (17 October 2026, MIT), via AI for Products' weekly brief issue 01.

Notes

More notes

  • The Agentic AI Foundation now holds ten projects, from three

    The Agentic AI Foundation took in four more projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. It now holds ten projects, up from three.

    For product people: Build an agent skill once and expect it to work across assistants. Standards are how a team's prompts and skills become an asset rather than a lock-in.

    Source: The Agentic AI Foundation, October 2026, via AI for Products' weekly brief issue 01.

  • OpenAI models come to Google Cloud, next to a managed Claude

    OpenAI's models are coming to Google Cloud, and Anthropic now offers a way to run Claude on Google Cloud that Anthropic itself manages.

    For product people: On Google Cloud you can now try both without a new contract. Choose the Claude option by who handles your data, not by the logo.

    Source: Google Cloud and Anthropic, 22 October 2026, via AI for Products' weekly brief issue 01.

  • Decision models: an answer without the essay

    Three decision models arrived in one day: OpenAI's Decisions API became generally available, Liquid AI released d1 and Ollama released d3. They answer a question with a choice rather than a paragraph.

    For product people: A surprising share of the steps in a workflow are choices, not prose. Those steps suit a decision model: faster, cheaper and easier to test.

    Source: OpenAI, Liquid AI and Ollama, 22 October 2026, via AI for Products' weekly brief issue 01.

All notes