GPT-6.2 Astra and Sol on Amazon Bedrock: Price, Regions, Ultrafast and Which to Use
GPT-6.2 Astra costs $8/$40 and Sol $2/$10 per million tokens on Amazon Bedrock. Model IDs, regions, Ultrafast pricing and what changed from GPT-6.

Short answer: OpenAI's GPT-6.2 Astra and GPT-6.2 Sol became generally available on Amazon Bedrock on 17 November 2026, the same day OpenAI released them. On Bedrock's Global cross-Region inference, Astra costs $8 per million input tokens and $40 per million output tokens, down from $10/$50 for GPT-6 Astra. Sol stays at $2/$10, the same as GPT-6.1 Sol. Astra also gets an optional Ultrafast speed tier, priced at 3x Standard until 28 February 2027 and 6x after that.
If you just want the practical version: Sol is the cheaper everyday model for coding, documents and routine agent work. Astra is OpenAI's top model, and it's now cheaper per token than GPT-6 Astra. Early reviews from power users are mixed, so test both on your own workload before switching.
Below are the full Bedrock prices, the model IDs, where you can run each model, what changed from the previous versions and what developers are saying.
What are GPT-6.2 Astra and GPT-6.2 Sol?
They are the newest versions of OpenAI's two main model lines:
- GPT-6.2 Astra replaces GPT-6 Astra as OpenAI's flagship. AWS's model card describes it as "OpenAI's frontier model for complex, multistep projects such as implementing and testing changes in existing codebases and producing professional deliverables (slides, spreadsheets, documents)."
- GPT-6.2 Sol replaces GPT-6.1 Sol. AWS calls it "a fast, versatile model for everyday work such as implementing features, writing technical content, and automating recurring workflows across business applications."
Both models accept text and images and return text. Both have a 1,050,000-token context window and up to 128,000 output tokens, according to the Bedrock model cards.
AWS's announcement says Astra is aimed at "reconciling competing inputs, investigating complex codebases, reasoning across text and images, and producing professional artifacts", while Sol is for "implementing and debugging software, analyzing documents and data, and automating multistep workflows."
How much does GPT-6.2 Astra cost on Amazon Bedrock?
All prices are in US dollars per million tokens, taken from the AWS model card. "Short context" means 272,000 input tokens or fewer. Above that, the long-context rate applies to the whole request.
| GPT-6.2 Astra (Standard) | Input | Cache write (30 min) | Cache read | Output |
|---|---|---|---|---|
| Global CRIS, short context | $8.00 | $10.00 | $0.40 | $40.00 |
| In-Region / US Geo CRIS, short context | $8.80 | $11.00 | $0.44 | $44.00 |
| Global CRIS, long context | $16.00 | $20.00 | $0.80 | $60.00 |
| In-Region / US Geo CRIS, long context | $17.60 | $22.00 | $0.88 | $66.00 |
In-Region and US geographic (Geo CRIS) prices include the 10% premium AWS charges over OpenAI's rates. Global CRIS matches OpenAI's own list price.
How much does Ultrafast mode cost?
Ultrafast is a speed tier. AWS says OpenAI reports "API speeds up to six times faster than Standard." It's a launch promotion: 3x the Standard rate through 28 February 2027, then 6x from 1 March 2027.
| GPT-6.2 Astra Ultrafast | Input | Output |
|---|---|---|
| Global CRIS, short context (promo) | $24.00 | $120.00 |
| In-Region / US Geo CRIS, short context (promo) | $26.40 | $132.00 |
| Global CRIS, long context (promo) | $48.00 | $180.00 |
Here's what that means for a budget. Today, Ultrafast Astra output costs $120 per million tokens. From March the same tokens would cost about $240, which is our calculation from AWS's 6x note. If you build a product around Ultrafast now, plan for that change.
How much does GPT-6.2 Sol cost on Amazon Bedrock?
| GPT-6.2 Sol (Standard) | Input | Cache write (30 min) | Cache read | Output |
|---|---|---|---|---|
| Global CRIS, short context | $2.00 | $2.50 | $0.10 | $10.00 |
| In-Region / US Geo CRIS, short context | $2.20 | $2.75 | $0.11 | $11.00 |
| Global CRIS, long context | $4.00 | $5.00 | $0.20 | $15.00 |
| In-Region / US Geo CRIS, long context | $4.40 | $5.50 | $0.22 | $16.50 |
These match GPT-6.1 Sol's Bedrock prices exactly. One difference: the GPT-6.2 Sol model card lists Standard as the only service tier. GPT-6.1 Sol does offer Ultrafast on Bedrock. If you rely on Sol Ultrafast, stay on 6.1 for now or move that traffic to Astra Ultrafast.
GPT-6.2 vs GPT-6 Astra vs GPT-6.1 Sol: what changed?
This comparison uses short-context Standard prices on Global CRIS, from each Bedrock model card.
| GPT-6 Astra | GPT-6.2 Astra | GPT-6.1 Sol | GPT-6.2 Sol | |
|---|---|---|---|---|
| Bedrock launch | 8 Sep 2026 | 17 Nov 2026 | 29 Sep 2026 | 17 Nov 2026 |
| Input / output per 1M | $10 / $50 | $8 / $40 | $2 / $10 | $2 / $10 |
| Cache read per 1M | $1.00 | $0.40 | $0.10 | $0.10 |
| Ultrafast | 6x Standard | 3x (promo), then 6x | 6x Standard | Not listed |
| Context window | 1.05M | 1.05M | 1M | 1.05M |
| Max output | 128K | 128K | 131,072 | 128K |
The biggest saving is in the cache-read price for Astra, down 60% from $1.00 to $0.40. Agents that resend a long system prompt or codebase context on every turn will feel that more than the 20% headline cut.
OpenAI also says Astra does more for less. According to The Decoder's summary of OpenAI's results, GPT-6.2 Astra scores 5.2 points higher than GPT-6 Astra on OpenAI's Agents' Last Exam benchmark "at under a third of the cost", and uses 22% fewer output tokens on its internal GDPval-v2 test. These are OpenAI's own numbers. We haven't tested the models.
Where is GPT-6.2 available on Bedrock?
Both models work on two Bedrock endpoints, and that's where most first-day errors will come from.
| Endpoint | Model ID | How it runs |
|---|---|---|
bedrock-mantle (OpenAI-compatible) |
openai.gpt-6.2-astra / openai.gpt-6.2-sol |
In-Region in us-east-1 (N. Virginia) only |
bedrock-runtime |
us.openai.gpt-6.2-astra / us.openai.gpt-6.2-sol |
US geographic cross-Region inference |
bedrock-runtime |
global.openai.gpt-6.2-astra / global.openai.gpt-6.2-sol |
Global cross-Region inference |
On bedrock-runtime there is no direct in-Region invocation, so you must name a us. or global. profile. According to the model cards:
- US Geo profile source Regions: us-east-1, us-east-2, us-west-1, us-west-2, ca-central-1 and ca-west-1. Requests are served from US Regions (plus the Canadian source Region for Canadian callers).
- Global profile source Regions include Frankfurt, Zurich, Stockholm, Milan, Spain, Ireland, London, Paris, Tokyo, Seoul, Osaka, Mumbai, Hyderabad, Singapore, Sydney, Jakarta, Melbourne, Malaysia, New Zealand, Thailand, Taipei, Tel Aviv, Mexico, Cape Town and São Paulo.
- Not listed: me-south-1 (Bahrain) and the GovCloud Regions.
That matters for EU and APAC teams. There is no EU or APAC geographic profile, so from Europe or Asia the only option is the Global profile, which can route your request anywhere in the world. If your data must stay in a specific country or region, read our guide to Bedrock data residency and in-country inference before you switch.
How do I use GPT-6.2 on Amazon Bedrock?
- Create a Bedrock API key in the Amazon Bedrock console (AWS's sample uses a long-term key).
- Install the OpenAI SDK (
pip install openai). Bedrock exposes the OpenAI Responses and Chat Completions APIs. - Pick your endpoint. Use
https://bedrock-mantle.us-east-1.api.aws/openai/v1for in-Region use, orhttps://bedrock-runtime.{region}.amazonaws.com/openai/v1with aus.orglobal.model ID. - Send a request. This is AWS's own sample from the GPT-6.2 Astra model card, using Ultrafast on the mantle endpoint:
from openai import OpenAI
client = OpenAI(
base_url="https://bedrock-mantle.us-east-1.api.aws/openai/v1",
)
response = client.responses.create(
model="openai.gpt-6.2-astra",
service_tier="ultrafast",
input="Explain how Amazon Bedrock cross-Region inference works.",
)
print(response.output_text)
Remove service_tier="ultrafast" (or set it to "default") for Standard pricing.
Known gotchas
- Converse API and documents: on
bedrock-runtime, Converse does not acceptdocumentcontent blocks for either model. AWS says to send files through the Responses API with aninput_filepart. - Guardrails: for Astra, Bedrock Guardrails work only through the Converse API. For Sol, they work through Chat Completions, InvokeModel and Converse, but not Responses. On
bedrock-mantle, the Sol model card lists Guardrails and invocation logs as not supported. - Quotas: each output token counts as 10 tokens against your tokens-per-minute quota, so long answers use up quota fast.
- No count-tokens API and no intelligent prompt routing for either model.
What are developers saying about GPT-6.2?
Early reactions are split. Most of them are about OpenAI's own apps and API, not Bedrock specifically.
- VentureBeat ran its launch story under the headline "'Essentially unusable'", quoting investor Matt Shumer's post on X. It also quoted developer Theo Browne calling Astra "the worst model OpenAI has ever released" and Ethan Mollick describing it as "very uneven & weird" (VentureBeat).
- On Hacker News, the reactions were more practical. One commenter (chriddyp) reported that on their data-analysis exam "Astra went from ~$20 (V6) to ~$4 (V6.2)" and "Sol went from $2.6 (V6.1) to $1.2 (V6.2)" with similar accuracy. Another (Mazer23) found the 6.2 models "faster and cheaper than their 6.1 counterparts". A third (theanonymousone) wrote: "Happy for the new Sol. It's my workhorse on company account." (Hacker News)
- Independent leaderboards are mixed too. The Decoder reports that Artificial Analysis ranks Astra second on its Intelligence Index (62, behind Claude Fable 5.2 at 65), while Vals AI ranks Astra 7th and Sol 9th on its index (The Decoder).
We didn't find any Bedrock-specific community reports yet. The questions we're seeing are about Sol's missing Ultrafast tier, whether EU teams can avoid Global routing, and how Astra compares with Claude Fable 5.2 at the same $8/$40 list price.
Which should you use: Astra, Sol or Claude Fable 5.2?
| If you need… | Start with | Why |
|---|---|---|
| Cheap, high-volume coding and document work | GPT-6.2 Sol | $2/$10, same price as 6.1 Sol, and OpenAI says it's close to Astra on its FrontierCode benchmark |
| Hardest multi-step agents, research, professional deliverables | GPT-6.2 Astra | OpenAI's top model, now 20% cheaper per token than GPT-6 Astra |
| Low-latency chat or live agents | GPT-6.2 Astra Ultrafast | Only 6.2 model with Ultrafast on Bedrock; promo pricing ends 28 Feb 2027 |
| Sol at Ultrafast speed | GPT-6.1 Sol | 6.2 Sol model card lists Standard tier only |
| A non-OpenAI alternative at the same list price | Claude Fable 5.2 | Also $8/$40; see our Fable 5.2 on Bedrock guide |
| The cheapest OpenAI model for simple tasks | GPT-6.1 Terra | $1/$5; see our GPT-6.1 Terra guide |
Editor's take
The headline is the price. Astra now matches Claude Fable 5.2's $8/$40 list price, and its cache-read rate dropped 60%, which is a big deal for agent workloads. Sol 6.2 is a free upgrade on paper, with the same price, a slightly larger context window and, if early Hacker News reports hold up, fewer tokens per task.
We'd be more careful about the quality claims. Most benchmark numbers so far are OpenAI's own, and several well-known testers were openly negative on day one. Sensible next steps: replay a sample of real prompts against global.openai.gpt-6.2-sol and global.openai.gpt-6.2-astra, compare cost per task rather than per token, and check that your Region and Guardrails setup is supported before moving production traffic. AWS re:Invent starts on 30 November (see our re:Invent 2026 guide), so more Bedrock model news is likely soon.
Also today on 9to5Amazon: Amazon Grocery's new prepared meals line, explained.
FAQ
Is GPT-6.2 available on AWS Bedrock?
Yes. AWS made GPT-6.2 Astra and GPT-6.2 Sol generally available on Amazon Bedrock on 17 November 2026. You can use them through the bedrock-mantle endpoint (us-east-1) or through us. and global. cross-Region profiles on bedrock-runtime.
How much does GPT-6.2 Astra cost?
On Bedrock's Global cross-Region inference, Astra costs $8 per million input tokens and $40 per million output tokens for requests up to 272,000 input tokens. In-Region and US Geo pricing is 10% higher ($8.80/$44).
How much does GPT-6.2 Sol cost?
$2 per million input tokens and $10 per million output tokens on Global CRIS, or $2.20/$11 In-Region and on the US Geo profile. That's the same as GPT-6.1 Sol.
What is GPT-6.2 Astra Ultrafast mode?
It's a faster service tier for Astra. OpenAI reports speeds up to six times faster than Standard. On Bedrock it costs 3x Standard until 28 February 2027 and 6x from 1 March 2027. You turn it on with service_tier="ultrafast" in the Responses API.
What is the GPT-6.2 Astra model ID on Bedrock?
openai.gpt-6.2-astra on bedrock-mantle, and us.openai.gpt-6.2-astra or global.openai.gpt-6.2-astra on bedrock-runtime. For Sol, replace astra with sol.
GPT-6.2 Sol vs GPT-6.1 Sol: should I upgrade?
The price is the same, so the main question is quality and token use on your workload. Early Hacker News reports say 6.2 is cheaper per task. One catch on Bedrock: 6.2 Sol doesn't list an Ultrafast tier, while 6.1 Sol has one.
What is the GPT-6.2 context window?
Both models have a 1,050,000-token context window and up to 128,000 output tokens on Bedrock. Requests over 272,000 input tokens are billed at the long-context rate.
Can I use GPT-6.2 on Bedrock in Europe?
Yes, but only through the Global cross-Region profile, which can route requests outside Europe. There's no EU geographic profile for these models at launch.
Sources
- AWS What's New: OpenAI GPT-6.2 Astra and GPT-6.2 Sol are now generally available on Amazon Bedrock (17 Nov 2026)
- Amazon Bedrock User Guide: GPT-6.2 Astra model card
- Amazon Bedrock User Guide: GPT-6.2 Sol model card
- Amazon Bedrock User Guide: GPT-6 Astra model card
- Amazon Bedrock User Guide: GPT-6.1 Sol model card
- Amazon Bedrock User Guide: OpenAI models
- The Decoder: OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains
- VentureBeat: 'Essentially unusable': OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users
- Hacker News: GPT-6.2 Astra and Sol discussion
Featured image: Server racks inside an AWS data center, photo published by Amazon (illustrative of AWS infrastructure; not a specific Bedrock or OpenAI facility). Source