The Zeroic Journal

GPT-6.2 Sol vs Astra: which one your product should call

OpenAI shipped two new tiers on November 17. Which product features belong on each, what they cost per user, and what to change before you switch.

On this page 6 sections
  1. GPT-6.2 Sol vs Astra on price
  2. Neither GPT-6.2 tier can switch reasoning off
  3. Four request changes before you switch
  4. Two pricing lines change the GPT-6.2 bill
  5. GPT-6.2 decision guide by feature
  6. How Better Analyst isolated AI inference for 1.5M+ users

GPT-6.2 Sol vs Astra comes down to price and task difficulty. OpenAI released both on November 17, 2026: Sol at $2 per million input tokens and $10 output, the same as GPT-6.1 Sol, and Astra at $8 and $40, four times that. Default your product to Sol, save Astra for the hardest tasks, and keep cheap high-volume calls off both.

Most of the launch talk is about benchmarks. For a founder, the useful questions are narrower. Which of your product’s features should call which tier? What does each one cost per active user? And what in your current OpenAI code stops working when you change the model ID? The prices and rules below come from OpenAI’s own docs as of November 18, 2026.

GPT-6.2 Sol vs Astra on price

OpenAI describes GPT-6.2 Sol as “our go-to model for everyday work, from coding to powering agents at scale” and GPT-6.2 Astra as “our most capable model for your hardest projects, from start to finish.” The pricing page puts numbers on that split:

ModelInput / 1MCached input / 1MOutput / 1M
gpt-6.2-astra$8.00$0.40$40.00
gpt-6.2-sol$2.00$0.10$10.00
gpt-6.1-sol$2.00$0.10$10.00
gpt-6.1-terra$1.00$0.05$5.00
gpt-6-luna$0.10$0.01$0.50

GPT-6.2 Sol is a free upgrade on list price. GPT-6.2 Astra is cheaper than the GPT-6 Astra it follows, which lists at $10 and $50, but it is still four times Sol on every token.

Per-token prices don’t tell you much on their own, so here’s a product. Say you have 10,000 monthly active users and each triggers 30 AI requests a month: 300,000 requests. Each request sends 3,000 tokens of prompt and gets 300 tokens of visible answer. On the two GPT-6.2 models, assume 500 reasoning tokens per request on top, billed as output. That figure is our assumption for a short task at low effort, so measure yours from the usage field.

SetupMonthly billPer active user
GPT-6.2 Astra, low effort$16,800$1.68
GPT-6.2 Sol, low effort$4,200$0.42
GPT-6.1 Terra, effort none$1,350$0.135
GPT-6 Luna, effort none$135$0.0135

These are list prices before caching or batching. The gap between the top and bottom rows is more than 100 times, which is why the model choice belongs to each feature, not to the whole product.

Neither GPT-6.2 tier can switch reasoning off

Both model pages say the same thing: “The API rejects requests that set reasoning effort to none.” OpenAI’s GPT-6.2 migration guide tells anyone on none or minimal to move to low, and recommends starting new integrations at medium. Every GPT-6.2 call therefore pays for some thinking at the output rate.

That matters most for the features that run thousands of times a day: tagging a ticket, routing an intent, pulling three fields out of an email. Those calls gain little from reasoning, and on GPT-6.2 they can’t opt out. GPT-6.1 Terra and GPT-6 Luna still list none as a reasoning effort, so they stay the right home for that work. Our GPT-4o deprecation plan walks through running Luna at effort none.

There’s a second reason to keep those calls where they are. OpenAI’s Responses migration guide says that from GPT-5.4 on, Chat Completions “does not support tool calling with reasoning_effort values other than none.” With none gone on GPT-6.2, a Chat Completions feature that calls functions can’t move to either new model without moving APIs too.

Four request changes before you switch

The migration guide lists the changes. These four are the ones most products hit:

  1. Tool calling needs the Responses API. Both models accept Chat Completions, but “tool calling requires Responses.” Any agent or function-calling feature moves its requests, output parsing and conversation state to /v1/responses first.
  2. Top-level instructions now come after the input. GPT-6.2 appends them after input, so long standing rules there no longer sit in a cacheable prefix. Move stable rules into a leading developer message, and keep instructions for guidance that applies to one request.
  3. Reasoning is dropped when history changes. Starting with GPT-6.2, the Responses API drops earlier reasoning items after mid-conversation edits to history, some model changes and changes of end user or organization. OpenAI warns that answer quality can fall and token use can rise. If your app trims or rewrites old turns, keep the prefix unchanged between responses instead.
  4. Sampling and logprob settings go. Remove temperature, top_p and top_logprobs from Responses requests, and logprobs from Chat Completions. Code that scored confidence from logprobs needs another signal.

If you’re coming from GPT-5.5 or earlier, also swap prompt_cache_retention for prompt_cache_options.ttl. OpenAI also says GPT-6.2 answers are more concise by default, so recheck any “be brief” lines in your prompts against the new output.

Two pricing lines change the GPT-6.2 bill

The 272,000-token line. Both model pages say that for requests with more than 272K input tokens, input and cache rates double and output costs 1.5 times the short rate, for the full request. A retrieval feature that sends 280,000 tokens to Sol pays $1.12 in input for that call instead of $0.56. Cap retrieved context in code so requests stay under the line.

Discounted background windows. OpenAI’s background mode guide adds a public-beta option on GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna: set background_options.completion_window to 2h for 30% off standard rates or 8h for 40% off. In the example above, an 8-hour window would take the Sol bill from $4,200 to $2,520. The catches: it needs the standard service tier, it shares a rate limit with Batch jobs, and standard requests can’t reuse cache entries those background requests create. It fits nightly reports, document backfills and digest emails, never a chat reply.

One more date to know if you use the fastest tier. Ultrafast on Sol costs 3 times standard through February 28, 2027, then 6 times from March 1, 2027.

GPT-6.2 decision guide by feature

  • Classification, routing, tagging and short extraction. Stay on GPT-6 Luna or GPT-6.1 Terra at effort none. GPT-6.2 can’t run without reasoning, so it costs more for the same answer.
  • Support replies, summaries and everyday agent steps on GPT-6.1 Sol. Move to GPT-6.2 Sol at the same price, after the four request changes and a side-by-side eval on a few hundred of your own inputs.
  • Long agent loops, code generation and multi-step analysis where Sol fails your evals. Try GPT-6.2 Astra on that step only, and route the rest of the flow to Sol.
  • Function calling on Chat Completions. Move to the Responses API before you move models, or keep the feature on Terra or Luna at effort none.
  • Prompts near 272,000 tokens. Trim retrieval or price the long-context rate in before you switch.
  • Work nobody waits on. Send it through an 8-hour background window and take 40% off.

Our guide to cutting OpenAI API costs covers caching and batching in more depth, and our AI services overview lists the LLM features we build.

How Better Analyst isolated AI inference for 1.5M+ users

On Better Analyst (formerly FormulaBot), one of the biggest pieces of work was splitting AI inference, data processing and the user-facing API into independently deployable services. That split is what let the platform scale from thousands to 1.5M+ users, with AI agents that answer questions in plain English and show the SQL, Python or R they wrote.

It also decides what a release like GPT-6.2 costs you. When inference lives in one service, picking a tier per feature is a routing table, and the Responses move, the developer message change and the sampling cleanup ship as one deploy with one rollback. When model calls are spread through the codebase, every one of those changes becomes a search across every file that talks to an LLM. If that sounds like your product, an AI audit maps each call to the cheapest tier that passes your evals and lists what to change before you switch.

Frequently asked questions

How much does GPT-6.2 Sol cost compared with GPT-6.2 Astra?
GPT-6.2 Sol costs $2 per million input tokens and $10 per million output tokens, the same as GPT-6.1 Sol. GPT-6.2 Astra costs $8 and $40, four times Sol's rates. Both double the input rate and charge 1.5 times the output rate for the whole request once the prompt passes 272,000 tokens. Batch halves the standard rates for both.
Can you turn reasoning off on GPT-6.2 Sol or Astra?
No. OpenAI's model pages say the API rejects requests that set reasoning effort to none on both GPT-6.2 models, and OpenAI's migration guide says calls on none or minimal should move to low. Every request bills some reasoning tokens as output. If a feature depends on effort none, keep it on GPT-6.1 Terra or GPT-6 Luna, which still accept it.
Do you have to use the Responses API for GPT-6.2?
Only if the feature calls tools. Both GPT-6.2 models accept Chat Completions and Responses requests, but OpenAI says tool calling on GPT-6.2 requires the Responses API. Plain text generation and structured outputs keep working on Chat Completions. Any agent or function-calling feature has to move its requests, response parsing and conversation state to Responses before it switches models.
Is GPT-6.2 Astra worth four times the price of Sol?
Only for the hardest tasks in your product. OpenAI positions Astra for demanding reasoning and coding work and Sol for everyday coding and agent workflows. Run the same set of a few hundred inputs from your own product through both. If Sol's answers pass your checks, the extra spend on Astra buys nothing your users will notice.
What breaks when you move from GPT-6.1 to GPT-6.2?
Four things most products hit. Tool calling needs the Responses API, top-level instructions now land after the input, reasoning items are dropped after mid-conversation edits to history, and temperature, top_p and logprobs settings must be removed. Calls that set reasoning effort to none or minimal also need to move to low, which raises output token counts.

References

  1. OpenAI API - Changelog (GPT-6.2 Astra and Sol, November 17, 2026)
  2. OpenAI API - Pricing
  3. OpenAI API - GPT-6.2 Sol model page
  4. OpenAI API - GPT-6.2 Astra model page
  5. OpenAI API - Using the latest model (GPT-6.2 migration)
  6. OpenAI API - Migrate to the Responses API
  7. OpenAI API - Background mode and discounted completion windows
  8. OpenAI API - Models

More from the journal

Hear from Prashant within 24 hours.

Prashant Abbi Partner @ Zeroic

Deciding which of your LLM calls move to GPT-6.2? An AI audit maps every model call in your product to the cheapest OpenAI tier that passes your evals, and lists the request changes to make before the switch reaches users.

See AI audits, or book a 30-minute call