OpenAI has officially launched GPT-6.2 Astra and its sibling model, GPT-6.2 Sol, introducing a pricing structure that aims to undercut previous frontier models while penalizing specific usage patterns. The headline numbers for the Standard tier of GPT-6.2 Astra are $8 per million input tokens and $40 per million output tokens. This represents a 20% reduction from the previous GPT-6 Astra pricing of $10/$50. However, the more significant shift lies in the cached input rates, which have dropped to $0.40 per million tokens, a 60% decrease from the $1.00 rate of the prior generation.

The Hidden Cost of Reasoning and Long Contexts

While the base rates look attractive, developers must navigate two critical cost multipliers. First, reasoning tokens are billed at the same rate as output tokens ($40 per million). Unlike previous models, GPT-6.2 Astra rejects requests that set reasoning effort to 'none', meaning every API call incurs mandatory reasoning overhead. Second, context window size dictates pricing tiers. For prompts under 272K tokens, the standard rates apply. Once input exceeds 272K tokens, the price doubles to $16 for input and $60 for output per million tokens. This penalty applies to the entire request, not just the overflow, making retrieval-augmented generation (RAG) strategies that keep prompts under 272K crucial for cost control.

Cache Economics and Agentic Workflows

The 0.05x multiplier for cache reads is a game-changer for agentic sessions. Simon Willison noted this as 'good news for long-running agentic sessions.' For a typical 50-turn coding agent, utilizing cache can reduce session costs by 74% compared to uncached calls on the new model. A cached session on GPT-6.2 Astra costs approximately $7.70, compared to $11.13 for the same workflow on GPT-6 Astra. The cache write costs are tiered: $10 per million for 30-minute retention and $20 per million for 12-hour retention. Crucially, cached prefix tokens are billed at the write rate, not the input rate, and there is no double-charging for the same tokens.

Competitive Landscape and Sibling Model Sol

GPT-6.2 Sol offers a budget-friendly alternative at $2 input and $10 output per million tokens, maintaining the same 1,050,000 token context window and 128,000 token output cap as Astra. When compared to competitors, GPT-6.2 Astra matches Anthropic's Claude Fable 5.2 on input/output rates ($8/$40), but Claude's cache reads are cheaper at $0.20 per million. However, early benchmarks suggest Astra may not outperform Fable 5.2 in all metrics, according to Willison. OpenAI also offers Batch processing at a 50% discount and a new 'Background' tier in public beta, providing 30-40% discounts for 2-hour or 8-hour completion windows, respectively.

Key Takeaways

  • GPT-6.2 Astra costs $8/$40 per million tokens (Standard tier), a 20% drop from GPT-6 Astra.
  • Cached input reads are now $0.40 per million, a 60% reduction that heavily favors agentic workflows.
  • Inputs exceeding 272K tokens trigger a 2x price increase for input and 1.5x for output.
  • Reasoning tokens are mandatory and billed at output rates, making 'none' reasoning effort unavailable.
  • GPT-6.2 Sol is priced at $2/$10 per million, offering identical context limits at one-quarter the cost.
  • Background processing is in public beta with 30-40% discounts, but shares rate limits with Batch jobs.

The Bottom Line

GPT-6.2 Astra is a cost-efficient powerhouse for developers who can rigorously manage context windows under 272K tokens and leverage aggressive caching strategies. However, the mandatory reasoning overhead and steep long-context penalties make it a risky choice for naive implementations. For teams prioritizing cost-per-token over raw benchmark scores, GPT-6.2 Sol or Anthropic’s cheaper cache options may offer better value.