OpenAI Launches GPT-6 Astra: 99.9% ARC-AGI 3, $10/$50 Pricing, and the Battle With Claude Fable 5.1

OpenAI officially commenced the rollout of GPT-6 Astra on September 3, 2026, introducing its most capable reasoning and autonomous execution model to date. Initially deployed to a restricted cohort of enterprise defense, software architecture, and research partners, Astra will expand to ChatGPT Plus, Pro, Business, and Enterprise subscribers over the coming week.

Astra arrives as a direct, aggressive counter to Anthropic’s Claude Fable 5.1 and Claude Mythos 5, mirroring Anthropic’s exact flagship pricing of $10.00 per million input tokens and $50.00 per million output tokens. However, OpenAI has tilted the competitive battlefield from raw per-token sticker prices to token efficiency, autonomous system reliability (SRE-Bench), and breakthrough exploit remediation capabilities.

+-----------------------------------------------------------------------------+
|               GPT-6 ASTRA VS. CLAUDE FABLE 5.1 SPECIFICATION MATRIX         |
+-----------------------------------------------------------------------------+
|                                                                             |
|   METRIC / CAPABILITY         OPENAI GPT-6 ASTRA        ANTHROPIC FABLE 5.1 |
|   -----------------------------------------------------------------------   |
|   Input Token Pricing         $10.00 / 1M tokens        $10.00 / 1M tokens  |
|   Output Token Pricing        $50.00 / 1M tokens        $50.00 / 1M tokens  |
|   Prompt Cache Discount       50% ($5.00 / 1M)          75% ($2.50 / 1M)    |
|   Context Window              1,000,000 tokens          1,000,000 tokens    |
|   Artificial Analysis Rating  61 (Tied)                 61 (Tied)           |
|   ARC-AGI 3 (General Intel)   99.9% (Adapter Harness)   96.4% (Standard)    |
|   SWE-bench Verified          73.8% (Preliminary)       74.2% (Verified)    |
|   Cost per Resolved Task      ~$0.20 per issue          ~$0.38 per issue    |
|   SRE-Bench (Infra Triage)    99.2%                     92.8%               |
|   ExploitBench (Cybersecurity)100.0% (Gated Access)     88.5% (Red-Teamed)  |
+-----------------------------------------------------------------------------+

Technical & Strategic Deep Dive

1. Deconstructing the 99.9% ARC-AGI 3 Score & the “Provider Adapter”

The headline benchmark of the Astra release is its 99.9% score on ARC-AGI 3, the challenging visual reasoning evaluation designed to resist memorization and test true out-of-distribution reasoning. However, technical teams must note the critical asterisk accompanying this claim: OpenAI ran this benchmark using a specialized Provider Adapter harness.

Unlike zero-shot or raw chain-of-thought evaluations, the Provider Adapter harness maintains cross-turn state across spatial grid transformations, executing recursive hypotheses and pruning incorrect logic branches before committing output tokens. When evaluated on vanilla zero-shot test harnesses without intermediate adapter state retention, early community replications estimate Astra’s baseline score between 89.4% and 92.1%—still exceptional, but illustrative of how harness design influences reported results.

2. The ExploitBench 100% Score and Frontier Cyber Access

Astra is the first model to solve 100% of challenges in ExploitBench, demonstrating autonomous capability to reverse-engineer stripped binaries, identify heap memory corruption bugs, construct Return-Oriented Programming (ROP) exploit chains, and verify payloads inside air-gapped sandbox environments.

Recognizing the acute national security and enterprise infrastructure risks of unconstrained offensive cyber tooling, OpenAI placed Astra’s vulnerability synthesis tools behind the Frontier Cyber Access program. Organizations seeking full access to these features must undergo rigorous Know-Your-Customer (KYC) vetting, sign legal attestation agreements, and implement dual-key authorization protocols that log all testing sessions cryptographically.

3. Real-World Coding Economics: Astra vs. Claude Fable 5.1

While Claude Fable 5.1 maintains a slight edge in SWE-bench Verified resolve rate (74.2% vs. Astra’s 73.8%), Astra dramatically alters the unit economics of autonomous coding agents. By optimizing internal reasoning token generation during multi-file codebase navigation, Astra completes tasks with fewer redundant loops.

At an average cost of $0.20 per benchmark task compared to Fable’s $0.38, enterprise engineering teams running thousands of continuous integration agent evaluations per day stand to reduce their monthly API spend by up to 47% without sacrificing resolution accuracy.

Real-World Utility & Limitations

High-Utility Strengths

• Site Reliability & Incident Management: With a 99.2% SRE-Bench score, Astra excels at analyzing multi-service log dumps, isolating cascading network faults, and generating immediate Kubernetes patching manifests.
• Cost-Effective Agentic Coding: Delivers near-parity with Claude Fable on code generation while significantly cutting the token overhead required for test verification loops.
• Structured API Call Reliability: Demonstrates zero schema hallucination across complex nested tool-call definitions.

Operational Limitations & Trade-offs

• Prompt Cache Economics: Anthropic Claude Fable retains an advantage for long-context workloads with its 75% cache discount ($2.50/1M cached tokens) versus OpenAI’s 50% discount ($5.00/1M cached tokens).
• Compliance Bottlenecks: Accessing Astra’s full capabilities requires navigating OpenAI’s Frontier Cyber Access portal, creating procurement delays for commercial penetration testing teams.