OpenAI rolled out GPT-6 Astra today—a frontier model that saturates several benchmarks and raises serious questions about the cost-effectiveness of open-weights and smaller agent stacks.

The Numbers That Matter

Astra saturates ARC-AGI-3 at 99.9% and FrontierMath Tier 4 at 98%, having already contributed to solving long-standing open problems in mathematics (proving infinitely many prime pairs occur within 186 apart). On cybersecurity evaluations without production safeguards, it scored 100% on ExploitBench—up from 78.5% for GPT-5.6 Sol—and 42.4% on ExploitGym versus Sol's 30.3%. It discovered two previously unknown zero-day vulnerabilities during testing.

For coding, the jump is material. On Terminal-Bench 4.0, Astra reaches 57.9% versus GPT-5.6 Sol's 37.3% and Claude Fable 5.1's 55.8%. On DeepSWE v1.1, it's 74.1% versus Sol's 72.7%. On professional tasks (AutomationBench), Astra scores 41.4% versus Sol's 18.1%—a 2.3x jump.

Where This Hits Your Stack

Computer use is the real story. On OSWorld 2.0, Astra achieves 72.6% at roughly 40 minutes per task, versus Sol's 65.7% at roughly 75 minutes—47% less latency for the same tier of capability. The system card notes that a generic computer-use-agent harness shows Astra causes fewer misaligned outcomes than any other frontier model tested. Critically, in an evaluation of Codex Auto-Review denial circumvention attempts, Astra never tried to bypass the safeguard, even when it was deliberately configured to be evadable.

For agents, the implication is stark: if you've been trading off frontier capability for cost by reaching for a 27B open-weights model or fine-tuned agent at 1/10th the price, Astra's computer-use and coding improvements—and its alignment properties—may force a reconsider. It also uses substantially fewer output tokens than competitors on comparable tasks (65% fewer than Claude Opus 5 on Agents' Last Exam).

The Catch: Cybersecurity Risk Requires Extra Safeguards

Astra meets the Critical threshold in OpenAI's Preparedness Framework for cybersecurity. Its ability to identify and weaponize zero-days is now a material capability, not a theoretical one. The model will refuse more advanced cyber tasks like proof-of-concept exploit creation at launch, but OpenAI plans to expand access through OpenAI Daybreak in the coming weeks with less restrictive safeguards for defensive workflows.

In production, Astra gets misalignment monitoring—a system of classifiers checking reasoning and actions for unauthorized behavior and stopping potentially unsafe activity. This can interrupt legitimate work. If a task is paused in ChatGPT or Codex, you review and continue; in the API, it stops. OpenAI acknowledges these checks can slow or halt legitimate defensive cybersecurity and says it's iterating to reduce unnecessary interruptions.

Availability and Pricing

Astra rolls out today to a limited set of organizations and becomes available over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API, Microsoft Azure, and AWS Bedrock.

API Standard pricing: $10 per million input tokens, $50 per million output tokens. Fast mode (up to 2x throughput) costs 2x the Standard rate.

Codex—OpenAI's agent harness—gets a speed update alongside Astra. Combined with Astra's efficiency, this yields 1.9x faster task completion on Mind2Web versus current GPT-5.6 Sol experience.

For developers using Codex, there's one notable new feature: context-aware note-taking across context windows. When the context window fills, Astra can preserve accumulated details in searchable notes rather than repeatedly compacting them into a single summary. You can enable this in your config.toml today (it becomes default in coming weeks).

What Changes

This is not a marginal update. Astra moves the frontier on three fronts that matter to agent workloads: latency (47% improvement on complex computer-use tasks), alignment (zero auto-review circumvention attempts in testing), and cost per token on comparable outputs. If your agent stack is built on open-weights or smaller proprietary models to contain costs, Astra's benchmark results—especially on reasoning, coding, and computer use—are worth a hard look at actual task-level economics.

The cybersecurity implications are also worth flagging. Astra isn't being positioned as a general-purpose offensive tool, but its exploit-generation capability is now documented and real. Defenders will have access, but so will anyone with API keys and intent. The safeguards are production-grade, but they're also fallible by design (they pause, not prevent).

Sources: