OpenAI and Rivals Release Four Frontier AI Models

In September 2026, Anthropic, OpenAI, and Google DeepMind released four frontier-class models: Claude Fable 5.1, GPT-6 Astra, GPT-6.1 Sol, and Gemini 4 Argon. OpenAI also canceled GPT-6.1 Astra on September 28 after it failed internal tests. While Astra and Fable 5.1 share a list price of $10 per million input tokens and $50 per million output tokens, Sol and Argon cost just one-fifth of that at $2 and $10. Argon's price is introductory and will eventually double to $4 and $20. For long prompts above 272K tokens, Astra charges $20 for inputs and $75 for outputs, while Sol doubles its input rate and increases outputs by 1.5 times.

These pricing structures diverge sharply for agentic workflows that rely on context caching. Astra charges $1.00 per million cached tokens, compared to $0.25 for Fable 5.1, and just $0.10 for Sol and Argon. In a test run of a 200K-token cached context over 100 steps, Astra costs $20.00, while Fable 5.1 costs $5.00, and Sol and Argon cost only $2.00. Furthermore, Argon is a structural outlier with a massive 1 million token maximum output limit per response, whereas Astra, Sol, and Fable 5.1 cap outputs at 128K tokens.

Performance benchmarks show no single model dominating. On the DeepSWE v1.1 software engineering benchmark, Argon leads at 77.9%, followed by Astra at 74.1% and Fable 5.1 at 67.4%, though the cheaper Sol matches Astra's score. Astra leads frontier engineering on FrontierSWE v2 at 65.5% and computer use on OSWorld-2.0 at 72.6%. On the Vals Index for finance and legal tasks, Argon scores 68.9%, beating Fable 5.1 at 65.8% and Astra at 63.1%. For vulnerability remediation on CWE-bench v1, Argon and Astra tie at 68%, while Fable 5.1 scores 58%. On ARC-AGI-2, Astra scores 95% to Fable 5.1's 90%.

For practitioners, choosing a model depends on access. Argon is currently restricted to cyber defenders in Google's Fairwind Program. While Fable 5.1 and Astra share list prices, Artificial Analysis found Fable 5.1 costs $9.18 per task compared to Astra's $4.72 due to token volume differences. Sol emerges as a highly cost-effective default, matching Astra on DeepSWE v1.1 and scoring 2.2 points above Claude Opus 5.5 on AutomationBench 1.0.6 at medium effort, all at a fraction of the cost.