OpenAI released GPT-6.2 on November 17, 2026, cutting Astra input and output prices by 20% and lowering cached reads to $0.40 per million tokens. Lower prices do not necessarily mean lower bills: once conversation prefixes are rewritten, existing reasoning content is discarded, and the model may consume more tokens by repeating the reasoning.
What this is
GPT-6.2 includes Astra for high-difficulty tasks and Sol for everyday work. Both support context windows of 1.05 million tokens and maximum outputs of 128,000 tokens. Astra's standard pricing is $8 per million input tokens and $40 per million output tokens. Once usage exceeds 272,000 tokens, input and cached-read rates double, while the output rate becomes 1.5 times higher.
The real cost driver is how the API is used. The Responses API retains reasoning items—the intermediate reasoning records that can be reused across requests. If a system compresses history, rewrites opening messages, or adjusts safety identifiers, old records are removed. Top-level instructions also no longer qualify for caching; long-lived system instructions should be placed in the first developer message within input.
Tool calls must also use the Responses API, reasoning effort cannot be set to none, and parameters such as temperature and top_p must be removed. For legacy code still running on the Chat Completions API, this is not a simple model-name replacement.
Industry view
We believe this release moves competition from model unit prices toward total cost of use. Cheaper cached reads benefit enterprise applications with stable prompts and highly repetitive context. But systems that frequently compress conversations or inject instructions dynamically may consume more tokens when reasoning items expire. Buyers should compare total per-task costs before and after migration rather than focusing only on the advertised price per million tokens.
The risks are equally clear. Prices rise sharply once context exceeds 272,000 tokens; without segmentation and caching strategies, a larger context window can become a budget problem rather than a capability advantage. Another objection is that migration now concentrates usage on the Responses API and imposes tighter parameter restrictions, so short-term adaptation and testing costs may offset the savings. We view the official price cut primarily as a signal to rebuild calling pipelines, not as an unconditional cost benefit.
Impact on regular people
For enterprise IT: Teams responsible for AI integration should first audit cache hit rates, prompt stability, and the share of workloads using long contexts before deciding whether to migrate. Otherwise, a “20% price cut” may exist only on paper.
For individual professionals: People using coding assistants or building workflows may notice more coherent answers, but old configurations can stop working after the upgrade. Settings related to cost, privacy, and auditability should also be reviewed again.
For the consumer market: If enterprise applications redirect the savings into broader service coverage, the effective price of AI features may fall. Conversely, migration and adaptation costs may be passed on through subscription prices.