AI Weekly: Gate week — Astra pause, Mythos scans, Sol $4/$20

Between 17 and 22 August 2026 the labs gated access, paused training, and repriced. Not a chatbot launch week. OpenAI paused two weeks of RL on the newest models they still plan to ship, after a Critical cyber threshold call. Anthropic showed protein binders from a real lab, then shipped computer use, Skills, and Files as GA. Mythos 5 landed as a scanner inside Claude Security. Sol got a promotional $4/$20. DeepSeek flipped V4 to peak and off-peak rates. Each section: the idea, then the claim. Sources at the bottom.

1. OpenAI: two-week RL pause, Astra at Critical cyber

Frontier models get better by training harder. This week OpenAI said some of that training pauses — on purpose — while cyber risk is watched.

The pause is two weeks of reinforcement learning on the newest models they still plan to ship. The largest planned frontier RL run stays on hold. Smaller-scale training and evals continue. On 7 August they called a Critical cyber threshold, citing a Hugging Face eval incident plus Astra. Monitoring is required for RL-plus-tools at Sol capability or higher. Astra-plus-tools inference gets extra watch — about 20% of the inference they watch. Some Astra workloads stay paused until they migrate.

So why does it matter? The gate is on training and monitored inference. ChatGPT is not being switched off.

• Two-week RL pause on the newest models they still plan to ship. ChatGPT and API Sol stay up.
• Largest frontier RL still on hold; smaller work continues.
• Astra at a Critical cyber call on 7 Aug; ~20% of watched inference under that watch.

2. Claude protein binders, wet lab

Protein design usually fails a lot. Anthropic’s claim this week is wet-lab hit rates above the usual 10–15% band — with Claude in the loop, not instead of the lab.

Mythos Preview and Opus 4.8 designed for 15 targets. Binders landed on 14. In a multi-target 48-hour window, Opus 4.8 hit 22.6% and Mythos Preview hit 26.7%. Mythos single-target hit 35.1%. Typical campaigns sit at 10–15%. The wet lab ran with Adaptyv Bio and Twist. Opus 5 timed NMR at 23 minutes and LC-MS at 19 minutes. Purity was 96.4% versus the lab’s 96.33%. Life-science work is still blocked on the most capable model. An access program is promised. It is not shipped here.

So why does it matter? Named hit rates from a lab campaign. Not “Claude replaced protein engineers.”

• 14 of 15 targets got binders; typical campaigns sit at 10–15%.
• Multi: Opus 22.6%, Mythos 26.7%; Mythos single 35.1%.
• Most capable model still blocked for life science in this post.

3. Computer use + Skills + Files GA

Agents that click around a desktop used to live in preview. On 20 August Anthropic called computer use, Skills, and Files generally available on the Claude Platform.

GA includes multi-action turns. Browser use sees page structure, not pixels only. HIPAA BAA is included. Skills API lets orgs upload and version skills. Files API adds auto expiry, 5× rate limits, and 1 TB per org. Skills and Files land on Microsoft Foundry. Computer and browser use are coming to Vertex. One customer Anthropic cites says 32 minutes down to 13, and about 30% cheaper. That is a vendor-cited customer story, not an audit.

So why does it matter? The news is GA plumbing — skills, files, a desktop and browser agent — not another demo reel.

• Computer use GA 20 Aug; multi-action turns; HIPAA BAA.
• Files: auto expiry, 5× rate limits, 1 TB/org; Skills/Files on Foundry.
• 32→13 min / ~30% cheaper is a customer quote. Not an independent redo.

4. Mythos 5 in Claude Security

Cyber-capable models keep getting talked about as ChatGPT with a red badge. Anthropic’s Mythos 5 story this week is a scan product, not a prompt box.

Mythos 5 ships inside Claude Security as an Enterprise public beta. Billing is standard tokens — no add-on. Findings come back as CWE, severity, confidence, and a suggested fix. A human must approve patches. Interactive patching uses the org’s existing models, not Mythos as chat. A $35M Defender Advantage Fund (0xDAF) puts credits on the table. The Cyber Verification Program is expanding. This digest is the existence of that gate — not an exploit walkthrough.

So why does it matter? Defenders get a scanner with structured findings. They do not get Mythos in a public chat UI.

• Scan, not chat. Enterprise public beta.
• CWE + severity + confidence + suggested patch; human approves.
• $35M 0xDAF credits; no exploit how-to here.

5. GPT-5.6 Sol $4/$20 promo

Token price is the line on the bill people argue about. On 21 August OpenAI cut Sol’s promotional API rate. The window runs at least through 21 November 2026.

GPT-5.6 Sol API promo is $4 input / $20 output per 1M tokens. That is −20% input and −33% output versus $5/$30. Plus and Pro subscription prices are unchanged in this announcement.

So why does it matter? Cheaper Sol tokens for a dated window. Not a forever list price.

• $4 in / $20 out per 1M; −20% / −33% vs $5/$30.
• Promo at least through 21 Nov 2026.
• Plus/Pro subs unchanged here.

6. DeepSeek V4 peak/off-peak live

Flat token rates are easy to quote. Peak and off-peak make the bill depend on when you call.

DeepSeek’s V4 pricing went live 16 August 2026 at 16:00 UTC — midnight Beijing on 17 August. Peak windows are 01:00–04:00 and 06:00–10:00 UTC. Off-peak is half of peak. Flash output moved from a flat $0.28 to $0.66 off-peak / $1.32 peak. Pro output moved from $0.87 to $1.98 / $3.96.

So why does it matter? V4 is not a small Flash nudge. The clock is half the story.

• Live 16 Aug 16:00 UTC; peak 01:00–04:00 and 06:00–10:00 UTC.
• Off-peak = half of peak.
• Flash out $0.28 → $0.66 / $1.32; Pro out $0.87 → $1.98 / $3.96.