Tuesday is mostly plumbing, the good kind. Anthropic shipped a way to stop re-uploading a million tokens every agent turn, Simon Willison's CLI got a cheaper default model, and NVIDIA put out an open semantic layer your agent can query over MCP. Then a policy change that will bite anyone who asks Claude Code for a remake of a game they loved.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Claude's Message Threads beta sends only the new turn, and cuts time to first token about 35% on very long chats
Until now every Messages API call carried the entire conversation. With a thread you send the full history once with "thread": {"type": "create"}, then only new messages, continuing from the last response id. Anthropic says that cuts time to first token by roughly 35% for requests between 500,000 and one million tokens. Token billing and rate limits don't change, since Claude still reads everything. A thread expires five minutes after its last request, so you keep your own copy for replay.
Link: platform.claude.com/docs
Try it now
Turn 1, straight from the docs:
curl -sS https://api.anthropic.com/v1/messages -H "x-api-key: $ANTHROPIC_API_KEY" -H "anthropic-version: 2023-06-01" -H "anthropic-beta: message-threads-2026-08-12" -H "content-type: application/json" -d '{"model":"claude-opus-5-5","max_tokens":1024,"messages":[{"role":"user","content":"Hi"}],"thread":{"type":"create"}}'
On turn 2 send only the new message with "thread": {"type": "continue", "previous_message_id": "<id from turn 1>"}. Build the 404 replay path first.
Story II
llm 0.37 adds GPT-6.1 Sol and Terra and switches the default to GPT-6 Luna, OpenAI's cheapest frontier model
Simon Willison's command-line tool picked up gpt-6.1-sol, gpt-6.1-terra and chat-latest. If you never set a default, you're now on GPT-6 Luna, which he says costs half what GPT-5.6 Luna did. There's a new llm openai endpoint command group for OpenAI-compatible servers, llm chat -u prints token usage after each turn, and embeddings accept typed attachments for multimodal models. Python 3.10 support is gone.
Link: github.com/simonw/llm
Try it now
uv tool upgrade llm (or pip install -U llm)
llm models default to see what you're on now
llm chat -m gpt-6.1-sol -u and watch the token count per turn
Want the old default back? llm models default gpt-5.6-luna
Story III
NVIDIA open-sources Auto Ontology, a semantic layer that makes your agent write SQL against certified definitions
Auto Ontology builds a starting ontology from your warehouse metadata, lets business people certify it, then serves it to agents through nine read-only MCP tools plus a CLI and API. Sources are Databricks, Postgres, Snowflake and DuckDB, and any OpenAI-compatible endpoint can stand in for the default Nemotron models. NVIDIA reports 90% faster ontology creation in its own testing, but hasn't published text-to-SQL accuracy numbers, per MarkTechPost. Apache 2.0, about 1.8k stars.
Link: github.com/NVIDIA/auto-ontology
Try it now
Point it at a DuckDB or Postgres copy, never production:
git clone https://github.com/NVIDIA/auto-ontology && cd auto-ontology
cp .env.example .env then add an NVIDIA API key from build.nvidia.com
docker compose up -d --build
Then have your agent call check_answerable before ask_question. The companion repo NVIDIA/auto-ontology-eval lets you measure accuracy yourself.
Story IV
Claude Code now declines to recreate existing video games, even under a new name
Anthropic's Addy Osmani confirmed on X that two sentences in Claude Code's system prompt for Opus 5.5 block cloning other people's games, citing the usage policy on IP. Renaming the clone or calling it original doesn't get around it. The change follows a run of AI-built browser ports of Halo and GTA Vice City. Original games, genre games and projects with your own assets are still fine, though The Decoder notes some users report refusals on original work too. Osmani says the rules will keep changing.
Link: the-decoder.com
Try it now
Describe the mechanics you want, not the game you're thinking of:
"Build an original browser roguelike in a single HTML file: grid movement, procedurally generated rooms, three enemy types with distinct behaviors, a fog-of-war radius that shrinks each floor, and placeholder art drawn with canvas shapes. Name it Saltmarsh."
If you get a refusal on something that really is yours, say where the assets and design come from.
— The Vibe Gate news desk. We read the firehose so you can keep building.