OpenAI Cuts GPT-6.2 Astra API Price 20% to $8/$40

Published • news
OpenAIPricingBenchmarks

OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, two closed-weight models with the frontier tier's API price dropping to $8 per million input tokens and $40 per million output tokens. OpenAI reports that is 20% below GPT-6 Astra; cached input falls to $0.40 per million. GPT-6.2 Sol keeps GPT-6.1 Sol's rates at $2 input and $10 output. Both models are live as gpt-6.2-astra and gpt-6.2-sol, and MarkTechPost reports availability in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu plans—not in regular ChatGPT.

Both models share a 1,050,000-token context window, a 922,000-token maximum input, and a 128,000-token maximum output. They accept text and image input and return text; neither accepts a reasoning_effort of none. Astra's knowledge cutoff is May 18, 2026, and Sol's is June 1, 2026. Requests above 272,000 input tokens double input and cached-input rates and set output at 1.5x standard; batch is 50% of standard, and ultrafast runs at 3x standard through February 28, 2027, then 6x. OpenAI recommends keeping prompt prefixes unchanged to preserve cache reuse and reasoning continuity across turns.

Model AA Intelligence Index API input / output per 1M tokens Cached input per 1M Max output Context window
GPT-6.2 Astra 62 $8.00 / $40.00 $0.40 128,000 1,050,000
GPT-6.2 Sol 57 $2.00 / $10.00 $0.10 128,000 1,050,000
Claude Fable 5.2 65 $8.00 / $40.00 $0.20 128,000 1,000,000
Gemini 4 Carbon 60 $2.00 / $10.00 until Nov 30, then $4.00 / $20.00 $0.10 intro Not disclosed 1,048,576

OpenAI's benchmark claims are explicitly cost-adjusted. On Agents' Last Exam, Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost and 5.5 points ahead of Claude Fable 5.2 at about 5% of the cost; the Fable comparison excludes nine tasks that used a fallback model. On Terminal-Bench-Science 0.1, OpenAI reports a 9-point gain over GPT-6 Astra at under one-third the cost. On Harvey LAB-AA v1.1, OpenAI says Astra more than doubles Fable 5.2's pass rate at about one-sixth the cost; MarkTechPost lists the absolute pass rates as 15.0% for Astra and 20.8% for the Astra for Law configuration. On internal GDPval v2, Astra gains 6.2 points over GPT-6 Astra while using 22% fewer output tokens, and Sol gains 5.3 points over GPT-6.1 Sol.

Independent Artificial Analysis results, summarised in MarkTechPost's explainer, add friction. Astra scores 62 on the AA Intelligence Index v5.0.1 behind Claude Fable 5.2's 65, with a measured cost of $4.00 per Intelligence Index task at max effort; Sol scores 57, Gemini 4 Carbon 60, and Claude Opus 5.5 59. MarkTechPost notes Sol matches GPT-6 Astra's score at about 15% of its cost per task. Coding is flat: Astra posts no gain on the AA Coding Agent Index, and Sol drops 3 points, driven by a 14-point fall on SWE-Atlas-QnA-AA. On AA-Omniscience hallucination tests, Astra registers 59% and Sol 55% on very hard questions. Epoch's Capabilities Index puts Astra at 174, tied with Fable 5.2. On AA-3D, Astra leads at 1,318 Elo, ahead of Fable 5.2 at 1,271 and Sol at 1,259.

In MarkTechPost's comparison table, Astra's $8/$40 list price matches Claude Fable 5.2's. Sol's $2/$10 is the more aggressive economic move. AI Mastery's pricing data, as of October 22, 2026, lists the outgoing GPT-6 Astra at $10 input and $50 output, confirming the 20% cut OpenAI claims.

AI Mastery analysis

Engineering teams should treat this launch as a three-way model-selection problem, not a quality race. Astra buys OpenAI's strongest professional outputs, but its value depends on vendor-reported output-token reductions and higher pass rates on long-horizon tasks. Sol is the more consequential default: at one-quarter of Astra's list price, it lands within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the cost, according to OpenAI. This extends the token-efficiency repricing pattern AI Mastery has tracked across recent releases. The constraint is trustworthiness: 59% and 55% AA-Omniscience hallucination rates mean compliance, legal, and financial workflows still need verification layers. Flat coding-agent scores are a separate warning; teams whose primary workload is repository editing should test Sol's 14-point SWE-Atlas-QnA-AA drop against internal baselines before migration.

The broader signal is that OpenAI is competing on cost per completed task rather than benchmark leadership alone. A second-place Intelligence Index finish paired with a 20% price cut shows frontier work being repriced around usable output, but the high hallucination rates and stagnant coding scores are the hidden costs engineers need to price into evaluations.

Sources

Frequently asked questions

How much does GPT-6.2 Astra cost per million tokens?

Astra costs $8 per million input tokens, $40 per million output tokens, and $0.40 per million cached input tokens. OpenAI reports this is 20% below GPT-6 Astra. Requests above 272,000 input tokens double input and cached-input rates and charge output at 1.5x.

Does GPT-6.2 Sol cost less than Astra?

Yes. Sol holds GPT-6.1 Sol's rates at $2 input and $10 output per million tokens, one-quarter of Astra's list price. OpenAI reports Sol comes within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the cost.

What are the context and output limits for GPT-6.2 Astra and Sol?

Both models share a 1,050,000-token context window, a 922,000-token maximum input, and a 128,000-token maximum output. They accept text and image input and return text. Neither accepts a reasoning_effort of none.

Did GPT-6.2 improve coding benchmarks?

No. Astra posts no gain on the AA Coding Agent Index, and Sol drops 3 points, driven by a 14-point fall on SWE-Atlas-QnA-AA, as reported by MarkTechPost from Artificial Analysis. Teams doing repository editing should test against internal baselines.

Where are GPT-6.2 Astra and Sol available?

Both are live in the OpenAI API as gpt-6.2-astra and gpt-6.2-sol. MarkTechPost reports rollout in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu plans, but not regular ChatGPT.

Free interactive tools for the decisions this piece raises.

Related Reading