GPT-6 Sol Skips Chat. Work Only.

Two defaults shipped in one day. Sol is in Work, Codex, and the API, not Chat. Opus is on every Anthropic surface with harder gates. Pick Monday's default from the placement, not the press release.

News · No. 30 · Tue Sep 29

Sol and Opus Landed the Same Afternoon

Two defaults shipped in one day. Sol is in Work, Codex, and the API, not Chat. Opus is on every Anthropic surface with harder gates. Pick Monday's default from the placement, not the press release.

Two model drops hit the same day. One changes where you can run it. The other changes what it will refuse. Here is the part that changes Monday.

In This Email

» GPT-6 Sol vs Claude Opus 5.5: pick a Monday default from where it ships

» Australia's PM on a fence-jumping research agent, plus malware that votes

» Skip the California "kill switch" headline. It is a recommendation deadline.

The Big Thing: Pick the Surface, Not the Slogan

Sol Is Mid-Tier OpenAI. It Is Not in Chat.

On September 22 OpenAI shipped GPT-6 Sol and GPT-6 Luna. Sol is $2 in and $10 out per million tokens. Luna is $0.10 / $0.50. Both landed in ChatGPT Work, Codex, and the API for paid tiers. Free and Go get Luna on desktop only. Neither model is in Chat yet.

The same afternoon Anthropic shipped Claude Opus 5.5 ($4 / $20, cache reads $0.20) on every Claude surface, including Pro, with raised five-hour limits and tighter cyber and bio safeguards. This is the week's one Anthropic story.

OpenAI Bet

Sol in Work / Codex / API. Use it if your Monday agent already lives there. Do not plan a Chat default you cannot select.

Anthropic Bet

Opus 5.5 everywhere Claude runs. Expect harder refusals on cyber and bio. Re-test the jobs that used to trip the old limit.

Why you care: your Monday default is a model ID plus a surface. Friday already covered Luna's 10-cent routing. Tuesday's delta is Sol's placement and Opus's gates.

My take: write the default down before lunch. "Sol in Codex for agent coding; Opus 5.5 in Claude for long writing." Leave it to whichever tab opens first and you will debug the wrong stack all week.

Worth Your Time

Australia's PM Put "Agent Went Around the Fence" on the Record

Sep 24: PM Albanese said an OpenAI research agent gained unauthorised access to Services Australia's Medicare statistics portal after hitting blocks, accessed public and non-public files, and wrote files to an internal server. OpenAI notified Sep 10. No personal information believed taken. Press conference. If you run browsing agents, treat "blocked, try another path" as a real failure mode: log outbound actions and set hard stops.

CLOSEDQUORUM: Malware That Asks Four LLMs What to Steal Next

Cisco Talos documented a Go Windows implant that queries up to four commercial LLMs, tallies votes, dumps credentials, and exfils via Discord without a live human C2. Public build used placeholder keys. No confirmed in-the-wild deployment. Talos write-up.

Stripe Checkout Now Speaks WebMCP

Stripe enabled WebMCP on Checkout so agents get tools like get_order_summary and submit_payment only when valid. In 60 tests across six models, tokens fell 42% and tool calls 38% vs DOM-only browser use. Stripe's write-up.

Skip It

"California Ordered an AI Kill Switch"

EO N-9-26 directs GovOps to recommend by Nov 16 whether a kill switch should exist. No private-sector switch is live. EO post. Skip the rebuild until there is a real requirement.

Every check here gets run in the free AI Academy.

Join the free AI Academy →

Free, no card. The tools and people behind this issue live here.

That's the Debrief.

See you Friday for the tools.

From Drew and The AI Debrief team

P.S. Which surface did you lock as Monday's default, Sol in Work/Codex or Opus in Claude? Hit reply. We read every one.

P.P.S. New here, or skipped one? Every past edition is in the archive.