GPT-6.2 vs GPT-6 Astra and GPT-6.1 Sol: Should You Switch?
Switch new work to official GPT-6.2 only if you can call OpenAI directly. SandBase still serves GPT-6 Astra at $9/$45 and GPT-6.1 Sol at $2/$10.

Start a new OpenAI-API project on GPT-6.2 when you can send gpt-6.2-astra or gpt-6.2-sol to OpenAI. Keep the SandBase client on GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, or GPT-6 Luna when that client only has a SandBase key. On the morning of 18 November 2026, SandBase does not list either 6.2 ID. A switch that assumes it does will 404.
The IDs, the official prices, and the 404 list are in GPT-6.2 Astra and Sol Are Officially Out. The cross-vendor price and context table is GPT-6.2 vs Gemini 4 Carbon and Claude Fable 5.2. The evening of 12 November, when none of this was official, is Is GPT-6.2 Out?.
Key takeaway
- New work that must use the new models: OpenAI’s API,
gpt-6.2-astraorgpt-6.2-sol. Not SandBase.- Official short-context Standard: Astra $8 / $0.40 cached / $40 output. Sol $2 / $0.10 cached / $10. Per 1M tokens. The API rejects reasoning effort
none.- Stay on GPT-6 Astra (
openai/gpt-6-astra) when you need the previous flagship on SandBase today: $9 / $45, 10% off $10 / $50, cache read $0.90, cache write $11.25.- Stay on GPT-6.1 Sol (
openai/gpt-6.1-sol) when the $2 / $10 workload already runs there. The sticker matches official GPT-6.2 Sol. The host does not.- GPT-6.1 Terra remains $1 / $5 on the SandBase hero, cache read $0.05, cache write $1.25. Do not treat a background-window discount on OpenAI as a SandBase price change.
The decision, one job at a time
| You are about to… | Send this | Do not send |
|---|---|---|
| Call the new flagship from an OpenAI key | gpt-6.2-astra | A SandBase path that 404s, or a bare gpt-6.2 |
| Call the new everyday model from an OpenAI key | gpt-6.2-sol | openai/gpt-6.2-sol on SandBase |
| Keep a SandBase flagship request working today | openai/gpt-6-astra | gpt-6.2-astra on that same key |
| Keep a $2 / $10 SandBase agent on the listed Sol | openai/gpt-6.1-sol | An unlisted 6.2 Sol string |
| Cut unit price again on SandBase, visual and writing work | openai/gpt-6.1-terra | A guessed 6.2 Terra ID. None was in the catalog |
| High-volume extraction on SandBase | openai/gpt-6-luna | A 6.2 ID “because it is newer” |
Official GPT-6.2 Sol’s short-context input, cached input, 30-minute cache write, and output ($2 / $0.10 / $2.50 / $10) match the live GPT-6.1 Sol hero. Official GPT-6.2 Astra ($8 / $0.40 / $10 write / $40) is cheaper on the sticker than both official GPT-6 Astra ($10 / $1 cached / $50) and the SandBase Astra hero ($9 / $0.90 / $11.25 / $45). Cheaper per token is not cheaper per task if the new model spends more output. OpenAI’s launch post says the opposite on the evals they cite, fewer output tokens, and those evals are OpenAI’s. They are not a SandBase probe. There is no TTFT, TPOT, throughput, or cache-hit number to put in this table.

Official Using GPT-6.2 guide typography from https://developers.openai.com/api/docs/guides/latest-model.md, fetched 18 November 2026, Asia/Shanghai. Plain rendering of published official markdown, not a live HTML screenshot (HTML 403 from this environment). The opening line points demanding work at GPT-6.2 Astra and lower-cost complex work at GPT-6.2 Sol. The “What’s new” paragraph says that, starting with GPT-6.2, prefix edits, a model downgrade, an end-user change, or an organization change cause the API to drop reasoning items. The migration block says to set model to gpt-6.2-astra or gpt-6.2-sol, start new integrations at medium, use low where you used none or minimal, and use the Responses API for tool calling.
That prefix rule is a reason not to switch a client that rewrites history. If your agent edits earlier instructions, swaps the model mid-conversation, or changes the end user, GPT-6.2 can lose the reasoning items, answer worse, and spend tokens repeating work. Keep the prefix stable, or stay on the model the client was validated against.
The other migration lines that break a naive swap: start new integrations at reasoning effort medium; if the client sends none or minimal, send low instead, because the API rejects none. Tool calling on these models requires the Responses API. From GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl set to "30m".
What the SandBase heroes charge today

Live SandBase GPT-6 Astra card at /models/openai/gpt-6-astra, captured 18 November 2026, Asia/Shanghai (HTTP 200). Hero is $9.00 / $45.00 (10% off the $10.00 / $50.00 list), with cache read $0.90 and cache write $11.25 per 1M tokens, 1.1M context, 128K max output. This is a predecessor listing, not a GPT-6.2 card. A site-wide first-top-up bonus banner is visible at the top; it is not a model discount. Only the Astra card carries a 10% OFF badge.

Live SandBase GPT-6.1 Sol card at /models/openai/gpt-6.1-sol, captured 18 November 2026, Asia/Shanghai (HTTP 200). Hero is $2.00 / $10.00, cache read $0.10, cache write $2.50 per 1M tokens, 1.1M context, 128K max output, no discount badge. This is a live predecessor listing, not a GPT-6.2 Sol card. A site-wide first-top-up bonus banner is visible at the top; it is not a model discount. Only the Astra card carries a 10% OFF badge.

Live SandBase GPT-6.1 Terra card at /models/openai/gpt-6.1-terra, captured 18 November 2026, Asia/Shanghai (HTTP 200). Hero is $1.00 / $5.00, cache read $0.05, cache write $1.25 per 1M tokens, 1.1M context, 128K max output, no discount badge. This is a live lower-tier predecessor, not a GPT-6.2 card. A site-wide first-top-up bonus banner is visible at the top; it is not a model discount. Only the Astra card carries a 10% OFF badge. The 17 November changelog includes this model, with GPT-6 Luna, on the background-window discount. That discount is an OpenAI Responses beta. It is not printed on this hero. GPT-6 Luna stays the volume card at $0.10 / $0.50. This crop does not show Luna.
Arithmetic only, short-context Standard, 100,000 uncached input tokens and 10,000 output tokens:
| Host and model | Input rate | Output rate | This example |
|---|---|---|---|
| Official GPT-6.2 Astra | $8 | $40 | $1.20 |
| SandBase GPT-6 Astra | $9 | $45 | $1.35 |
| Official GPT-6 Astra list | $10 | $50 | $1.50 |
| Official GPT-6.2 Sol, or SandBase GPT-6.1 Sol | $2 | $10 | $0.30 |
| SandBase GPT-6.1 Terra | $1 | $5 | $0.15 |
If a real turn thinks far past 10,000 output tokens, every row is low. Ultrafast at 3×, through 28 February 2027, turns the official Astra example into $3.60 before the prompt changes. The 6× rate starts 1 March 2027, Pacific time. Long context, above 272K input, doubles input and cache and multiplies output by 1.5 on the official 6.2 pages. None of those multipliers were read off a SandBase invoice.
GPT-6 Astra usage limits and API cost still owns the older card’s limit notes. It is not a 6.2 bill.
When to switch, and when the old ID is the correct ID
Switch the new project to official GPT-6.2 Astra if the work is the “hardest project” class the Astra page names, you have an OpenAI API key, and you can accept medium as the starting effort. Switch that project to official GPT-6.2 Sol if you want the everyday coding and agent tier at the $2 / $10 sticker and you are allowed to call OpenAI directly. The announcement’s ChatGPT side is Work and Codex for Plus, Pro, Enterprise, and Edu, not regular Chat. Pro 500 gets Ultrafast on either model. Pro 200 gets Ultrafast for the first time, starting with Sol. Enterprise and Edu still need an admin enable. Those sentences are the 17 November announcement, not a SandBase plan.
Keep GPT-6 Astra on SandBase if the production key is a SandBase key, the prompt stack was validated on openai/gpt-6-astra, or you rewrite prefixes between turns and do not want the new drop-reasoning behavior. The 10% on that hero is real. It does not follow you onto gpt-6.2-astra.
Keep GPT-6.1 Sol on SandBase if the bill is already $2 / $10 and the only reason to move is the version number. Official Sol’s sticker matches. The behavior, the cutoff (Jun 01, 2026 on the Sol page, versus the older Sol’s cutoff on its own page), and the host do not. Move when you have rerun the agent on gpt-6.2-sol through OpenAI, not when the changelog date looks newer.
Keep Terra or Luna when the task fits those cards. Terra’s hero is a fifth of Sol’s output price. Luna’s hero is $0.10 / $0.50. Neither became a 6.2 model on 17 November. The background-window row mentions them beside the new IDs. That is a billing option on OpenAI’s Responses API, shared with 6.2, not a rename.
Do not switch if you need audio in. Both 6.2 model pages mark audio unsupported. A SandBase card that shows audio input is a different model. Do not switch a fine-tune onto 6.2. Fine-tuning is unsupported. Do not switch a client that depends on reasoning effort none.
Two older pro-mode cards were still HTTP 200 this morning and are easy to confuse with a new SKU. openai/gpt-6-astra-pro is $10 / $50. openai/gpt-6.1-sol-pro is $2 / $10. They are pro mode of the current SandBase models, not GPT-6.2. This check did not look up a gpt-6.2-astra-pro ID, so it does not claim one exists or does not.
curl -X POST https://api.sandbase.ai/v1/chat/completions \
-H "Authorization: Bearer $SANDBASE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-6.1-sol",
"messages": [
{"role": "user", "content": "This request stays on GPT-6.1 Sol because SandBase has no GPT-6.2 card."}
]
}'Sign in, then use the first API call shape if the key is new. The model field is a live ID. GPT-6 Astra’s card also lists Responses. /v1/run is the wrong route for these text models.
FAQ
Should every Astra app move to GPT-6.2 Astra?
No. Move a new OpenAI-API integration if you want the new ID and the $8 / $40 list. Leave a SandBase integration on openai/gpt-6-astra until a card exists and you have rerun the prompts. The unit-price gap ($9 / $45 on the card versus $8 / $40 official) is not available inside SandBase today.
Is GPT-6.2 Sol a drop-in for GPT-6.1 Sol?
The short-context token prices match. The ID, the host, the knowledge cutoff on the new page (Jun 01, 2026), and the prefix-edit rule do not. Test before you retarget production.
What if my client sends reasoning effort none?
GPT-6.2 rejects it. The migration guide says to use low where you used none or minimal, and to start a new integration at medium.
Will SandBase bill the official 6.2 price if I guess the path?
No path returned HTTP 200 on 18 November. A 404 is not a discounted route. There is also no measured latency to compare, so a switch for speed cannot be justified from SandBase data. That measurement waits on a live card.
What to open
Open GPT-6.1 Sol if the $2 / $10 SandBase agent should stay put. Open GPT-6 Astra if the SandBase flagship should stay put. Open OpenAI’s docs for gpt-6.2-astra and gpt-6.2-sol only on a key that can actually call them.
SandBase heroes captured 18 November 2026, Asia/Shanghai. The Using GPT-6.2 frame is a plain rendering of published markdown from https://developers.openai.com/api/docs/guides/latest-model.md, fetched the same morning, not a live HTML screenshot (HTML returned HTTP 403 from this environment). Official 6.2 prices come from the published model markdown the same morning. No SandBase call to a 6.2 ID was sent, because the cards 404.
Found this useful? Share it