LLMs

OpenAI Launches GPT-6.2 Astra and Sol With 20% Cost Reductions and New Benchmark Leads

OpenAI releases GPT-6.2 Astra for complex work and GPT-6.2 Sol for everyday tasks, cutting input costs 20% while advancing on scientific and professional benchmarks.

Last verified:

OpenAI’s Cost-Reduction Strategy in GPT-6.2 Astra and Sol

OpenAI released two new models on November 17, 2026: GPT-6.2 Astra for complex professional and scientific work, and GPT-6.2 Sol for everyday multi-step tasks. According to the OpenAI Blog, both models reduce token costs while maintaining or advancing performance on key benchmarks, signaling a shift toward efficiency as a primary competitive lever alongside capability.

20% Price Cut and Benchmark Gains in GPT-6.2 Astra

According to the OpenAI Blog, GPT-6.2 Astra‘s API pricing drops 20% compared to GPT-6 Astra: input tokens cost $8/MTok (down from ~$10), output tokens cost $40/MTok, and cached input tokens drop to $0.40/MTok—a 60% discount on reused context. Performance gains accompany the cost reduction. On Agents’ Last Exam, which evaluates long-horizon professional workflows, Astra scores 5.2 points higher than GPT-6 Astra while costing less than one-third as much per task. The same model also scores 5.5 points higher than Claude Fable 5.2 at just 5% of Fable’s cost on the same benchmark, according to OpenAI’s testing.

For scientific work, Astra sets new highs on Terminal-Bench-Science 0.1 (a 9-percentage-point lead over GPT-6 Astra at one-third the cost) and the Vals Science Index, which measures model performance on interpreting biological data and running experiments to infer system behavior. OpenAI’s blog emphasizes Astra’s strength in professional document and code generation, stating the model “reads less like AI” in spreadsheets and slides, and produces cleaner, more functional code for websites and 3D design.

GPT-6.2 Sol: Efficiency for Everyday Tasks

GPT-6.2 Sol maintains GPT-6.1 Sol’s API pricing while improving capability. According to OpenAI, Sol performs within 0.5 points of Astra on Cognition’s FrontierCode 2.0—a benchmark measuring whether code changes meet maintainer standards for merging—at approximately one-fifth Astra’s cost (both at maximum settings). OpenAI’s internal team has adopted Sol as the default for both technical and non-technical tasks, the blog notes, citing improvements in speed, multistep agentic workflows, and natural-language clarity.

Why This Matters

The simultaneous release of a frontier model (Astra) with sharply reduced pricing and an efficient tier (Sol) with held-price-but-improved-capability signals OpenAI’s bet that cost-per-task and latency, not raw capability alone, now drive customer switching. Teams currently evaluating Claude Fable or running cost-sensitive agentic workloads face a re-evaluation decision: Astra’s 5% cost advantage on Agents’ Last Exam, combined with its 20% input-price reduction, may shift multi-model strategies favoring higher-token-volume tasks (scientific compute, long-context retrieval, batch processing). Sol’s FrontierCode parity at one-fifth the cost targets developers choosing between Astra and cheaper alternatives—collapsing the value gap without requiring a tier downgrade. If these benchmarks hold under independent reproduction and real-world usage, the pricing floor for frontier-class models shifts materially downward.

Frequently Asked Questions

How much cheaper is GPT-6.2 Astra compared to GPT-6 Astra?

GPT-6.2 Astra's input and output token prices are 20% lower than GPT-6 Astra's, and cached input drops to $0.40/MTok (a 60% discount).

How does GPT-6.2 Sol compare to GPT-6.2 Astra on code tasks?

According to OpenAI, Sol scores within 0.5 points of Astra on FrontierCode 2.0 at approximately one-fifth Astra's cost (both at maximum settings).

What benchmark gains does GPT-6.2 Astra show over prior versions?

On Agents' Last Exam, Astra scores 5.2 points higher than GPT-6 Astra and 9 percentage points higher on Terminal-Bench-Science 0.1, both at reduced cost.

#gpt-6 #pricing #benchmarks #frontier-models #efficiency