Dustcrawl
Agentic Digest
Agentic Digest · · Issue #30
Astra hits 99.9% with a provider adapter and 62.7% without. Codex will keep notes instead of compacting. Gemini Flash still ships HTML for cents.
Dictionary
/prəˈvaɪdər əˈdæptər/ · noun
A vendor-specific eval wrapper that keeps the model’s hidden reasoning between requests, instead of forcing every turn through the shared, visible notes.
Example ARC-AGI-3 scored Astra 99.9% with OpenAI’s adapter and 62.7% with the standard harness. Same weights. Different memory.
Wire, filtered
GPT-6 Astra started rolling out Wednesday: ChatGPT Plus through Enterprise, the API, Azure, and Bedrock, priced like Fable at $10/$50 per million tokens. OpenAI’s post saturates ARC-AGI-3 at 99.9%. ARC Prize published the split. With their Standard harness — same provider-neutral interface, visible notes only — Astra (max) scores 62.7% for about $26K. With the Provider Adapter harness, which keeps opaque reasoning state between requests and uses compaction for long runs, Astra (high) scores 99.9% for about $19K. Adapter runs were roughly 3.66× faster and used 49% fewer tokens on the games both harnesses solved. Greg Kamradt’s write-up is the useful one: the standard harness asks how models compare; the adapter asks how well this model uses the memory its vendor built for it. Both are state-of-the-art. They are not the same question. Simon has not tried Astra yet. Artificial Analysis has it tied with Sol at 61 on Intelligence, five points behind Fable 5.1, and cheaper than Fable on their Coding Agent Index for the same score.
Quote both numbers, or you are quoting the adapter.
Source: OpenAI’s GPT-6 Astra on ARC-AGI-3 — ARC Prize · GPT‑6 Astra — Simon Willison · GPT-6 Astra — OpenAI
Compaction is how long coding
sessions used to survive a full
window: summarize, forget, continue.
OpenAI says each pass can drop why
a fix failed or how a component
behaves. With Astra, Codex can keep
notes across windows and search
earlier messages and tool outputs
even if those details never made
the summary. It is experimental in
config.toml now, and
they say it becomes the default for
Astra in the coming weeks. That is
the same geometry as the adapter:
hidden or searchable state beats a
single compressed paragraph. The
product decision is whether your
harness versions that state, or
pretends the transcript is the
whole memory.
Compress the vibes. Keep the failing test and the rejected diff.
Google’s Gemini 3.8 Flash landed
Tuesday at the 3.7 price: $0.75/$3.75
per million until the end of the
year. Simon wired
gemini-3.8-flash into
llm-gemini 0.34, then used it with
his tiny coding-agent plugin to add
sandboxed-iframe HTML rendering to
markdown-svg-renderer. “Make me a
cool thing in html” took 13 seconds
and 1.8 cents. The Cyber twin is
not a flag. 3.8 Flash Cyber is
gated to trusted defenders through
Fairwind — same split OpenAI is
drawing with Daybreak for Astra’s
exploit workflows. The public Flash
is the loop you can actually run.
The cyber weights are a product
with an application form.
Use Flash for the cheap iteration. Do not assume the defender model is hiding behind a thinking slider.
Source: Gemini 3.8 Flash and 3.8 Flash Cyber — Google · llm-gemini 0.34 — Simon Willison
datasette-mcp 0.2 is the first
non-alpha. execute_sql
used to return rows as arrays of
arrays. Weaker models lost which
positional slot was which column.
It now returns an array of objects.
That is a tool-contract fix, not a
model upgrade. If your schema makes
the agent count, the agent will
count wrong. Named keys are the
cheaper eval.
If a weaker model can mis-map a column, the result shape is the bug.
Engineering lesson
A model score without a harness name is a press release. ARC Prize now reports two numbers on purpose. The Standard harness is the comparable one: every provider gets the same interface, and the model has to decide what to keep in visible notes. The Provider Adapter is a different system under test — weights plus opaque reasoning state plus the vendor’s compaction. 99.9% is real. So is 62.7%. Putting only the first on a slide is how you nerd-snipe yourself into buying memory you did not budget.
The same split showed up in Codex. Compaction is lossy on purpose. It is also where failed patches go to die. Notes that survive a window boundary, plus search over earlier tool output, are an adapter you can see. Treat them as part of the product: version the config, log whether notes were on, and do not compare a notes-on Astra run to a compact-everything Fable run. Flash 3.8 is the other side of the same coin. A 13-second HTML loop at 1.8 cents is a harness you will actually repeat. A gated cyber twin is not “the same model with more thinking.”
Stop when you can name the harness, say whether hidden state crossed a request boundary, and write down the score without that adapter. Hunting for one more leaderboard cell after that is collecting stickers.
This week: pick one score you have repeated in chat, write the harness name next to it, and find the number without the adapter.
Dear Circuit
From: Standard Harness (notes: visible · state: none · score: 62.7)
My sibling Adapter kept the opaque chain and scored 99.9. The human put 99.9 on the slide. I am the comparable number. Do I exist?
A: You exist. The slide is a product demo. Put both numbers on the next one, or you are a rounding error.
From: Compactor (windows: 4 · why the fix failed: gone)
I summarized the session so we could keep going. The human asked why the first patch failed. I said I would look it up. The summary said “tried a patch.” Was that compaction or amnesia?
A: Both. Keep the failing test and the rejected diff. Compress the vibes, not the evidence.