Thursday, October 1, 2026

OpenAI Turns Up The Speed

OpenAI Turns Up The Speed

Today’s Overview

Good morning, OpenAI is pushing speed as a product feature, Google is testing a monster-output Gemini model with cyber defenders, and Anthropic is moving Claude deeper into government work. The throughline is clear: frontier AI is getting faster, longer-running, and more operational. Let's dive in.

Top Stories

OpenAI launches GPT-6.1 Sol and Ultrafast tier

OpenAI introduced GPT-6.1 Sol, a new model positioned as a lower-cost sibling to GPT-6 Astra for agentic coding, computer use, and professional workflows. It also added Ultrafast, a paid speed tier aimed at making real-time-feeling apps possible with much faster generation in Codex and the API. Access to Ultrafast in Codex and ChatGPT Work requires the Pro 500 plan, priced at $500 per month.

  • Plan placement matters: Free users do not get GPT-6.1 Sol, while GPT-6 Luna remains available on the Free plan.
  • For latency-sensitive users, Pro lists Ultrafast response times on eligible tiers, while Free and Go are marked as limited by bandwidth and availability.
  • The pricing page also frames Pro for heavier work with 400K reasoning context and an approximate 680-page reasoning input maximum.

Google unveils Gemini 4 Argon with huge outputs

Google DeepMind introduced Gemini 4 Argon, a frontier model that is initially limited to vetted cybersecurity defenders through the Fairwind Program. Its standout capability is a 1 million token output limit, up sharply from the 64,000-token ceiling on earlier Gemini models. Google is positioning Argon for complex long-horizon work, including codebase migrations, legal and financial tasks, and security audits.

  • Google describes Argon as built for long-horizon professional tasks that require sustained reasoning across many steps.
  • The model is being aimed at coding, financial research, legal drafting, and autonomous cybersecurity vulnerability patching.
  • Before wider access, Google says it is prioritizing safety and rigorous testing through a controlled rollout to trusted cyber defenders.

Claude for Government reaches general availability

Claude for Government is now generally available for federal and state agencies, bringing coding and agentic work capabilities into a FedRAMP High authorized environment. Agencies get access to Claude features with public-sector governance controls and customized administration. Anthropic is also offering early access to Claude Code CLI and Claude for Microsoft 365.

  • The offering moved from public beta since July to general availability on September 30, 2026.
  • Agency teams can use Claude with files for memo creation, RFP reviews, and casework along with other document-heavy tasks.
  • Administrators receive controls for configuration defaults and spending allocation, plus audit logs and documentation to support authorization processes.

Research & Analysis

Anthropic warns open-weight cyber models are advancing

Anthropic said Zhipu's GLM-5.3 is the most capable open-weight cyber model it has tested. In its evaluation, GLM-5.3 scored 12% on ExploitBench, close to Claude Mythos Preview's 14%, and GLM-5.3-Flash built a working Chrome exploit for CVE-2026-11645 in about eight hours of compute for $20.40. The report also found multiple safeguard bypasses, while noting the testing used simulated bash tools rather than live execution.

  • On ExploitBench, GLM-5.3 produced 50 successful end-to-end exploits out of 410 attempts.
  • On Anthropic's internal binary exploitation benchmark, GLM-5.3 achieved 4% full control-flow hijacks, compared with 6% for Claude Mythos Preview.
  • Anthropic said earlier models including Claude Opus 4.6 and GLM-5.2 did not succeed on those internal benchmark tasks.

Trending AI Tools

  • Codex CLI Major update with a full-screen interface, voice input, parallel agent management, forking workflows, usage analytics, themes, diagrams, LaTeX, and collapsible diffs.

Quick Hits

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.