GPT-6.2 Astra and Sol launched 17 Nov 2026: prices and Responses API changes

GPT-6.2 Sol matches GPT-6.1 Sol on price. Move instructions into a developer message first, or cached prompts can cost 84% more than gpt-5.

Read time
17 min
Word count
2.5K
Sections
9
FAQs
8
Share
Two engineers review architecture diagram cards at a laptop, with the caption GPT-6.2 API Changes: update prompts and caches
On this page · 9 sections
  1. What OpenAI released on 17 November
  2. Prices: Sol is unchanged, Astra is 20% cheaper
  3. What changes in the Responses API
  4. What the changes cost: three worked examples
  5. Should you skip GPT-6.1 Sol?
  6. A 21-day plan to 11 December
  7. India-specific considerations
  8. How eCorpIT can help
  9. References

Summary. OpenAI released gpt-6.2-astra and gpt-6.2-sol on 17 November 2026, for both the Responses API and Chat Completions. On OpenAI's pricing page, checked on 20 November 2026, GPT-6.2 Astra costs $8 input, $0.40 cached input and $40 output per million tokens, down from $10, $1 and $50 on GPT-6 Astra. GPT-6.2 Sol costs $2, $0.10 and $10, exactly the same as GPT-6.1 Sol. The launch came 24 days before 11 December 2026, when gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro shut down; the deprecations page still names gpt-6.1-sol as the replacement for gpt-5 and o3. Two Responses API changes come with 6.2: reasoning items are dropped after edits to earlier conversation turns, and top-level instructions now sit after the input, outside the prompt cache. GPT-6.2 also has no Flex tier, which OpenAI is winding down; discounted background windows of 30% and 40% take its place. Our verdict: teams still on gpt-5 or o3 should target GPT-6.2 Sol directly, because it costs the same, has a later knowledge cutoff and cheaper Ultrafast, unless they depend on Flex pricing or cannot move cached instructions into a developer message before 11 December.

About this article. This is for engineering leads, platform teams and finance owners in India, the UK, the EU and the US who run OpenAI API workloads, especially those part-way through the move off gpt-5 and o3. We checked every model ID, price, date and parameter on 20 November 2026 against OpenAI's changelog, pricing page, deprecations page, model pages and its GPT-6.2, reasoning, prompt caching, background and Flex guides, plus the ECB reference rates. OpenAI's announcement page did not load for us, so launch details beyond the docs are attributed to press reports. The cost models are our own illustrations at list prices; your token counts will differ, and vendors change prices often. This is not financial advice, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.

What OpenAI released on 17 November

OpenAI's changelog entry for 17 November positions GPT-6.2 Astra for demanding reasoning and coding work and GPT-6.2 Sol for everyday coding and agent workflows. Both support the Responses API and Chat Completions, but tool calling requires the Responses API. A second entry the same day added two-hour and eight-hour background completion windows in public beta for GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna.

Press coverage reported that both models also rolled out in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and relayed OpenAI's claim that GPT-6.2 Sol comes within half a point of Astra on a coding benchmark at about a fifth of the cost. Those are OpenAI's figures, not results we have reproduced.

The model pages give the hard specifications.

Specification GPT-6.2 Astra GPT-6.2 Sol GPT-6.1 Sol GPT-6 Astra
Model ID gpt-6.2-astra gpt-6.2-sol gpt-6.1-sol gpt-6-astra
Context window 1,050,000 1,050,000 1,050,000 1,050,000
Maximum input tokens 922,000 922,000 922,000 922,000
Maximum output tokens 128,000 128,000 128,000 128,000
Knowledge cutoff 18 May 2026 1 June 2026 30 April 2026 30 April 2026
Input modalities Text, image Text, image Text, image Text, image
Batch API Supported Supported Supported Supported
Flex processing Not offered Not offered 50% off 50% off
Background windows 30% or 40% off 30% or 40% off Not listed Not listed

Two details matter for planning. First, neither 6.2 model has a dated snapshot: each page lists only its alias as a model ID, so there is nothing to pin. Second, rate limits on the Build, Launch and Grow tiers are identical across GPT-6.2 Astra, GPT-6.2 Sol and GPT-6.1 Sol, from 5,000 requests and 1,000,000 tokens per minute on Build to 15,000 requests and 40,000,000 tokens per minute on Grow. Moving between them will not change your throughput ceiling.

One inconsistency is worth knowing: the gpt-5 model page now recommends GPT-6.2 Astra, while the deprecations table still names GPT-6.1 Sol.

Prices: Sol is unchanged, Astra is 20% cheaper

All rates below are Standard processing for prompts of 272,000 input tokens or fewer, in US dollars per million tokens, from OpenAI's pricing page.

Model Input Cached input 30-minute cache write 12-hour cache write Output
gpt-6.2-astra $8.00 $0.40 $10.00 $20.00 $40.00
gpt-6-astra $10.00 $1.00 $12.50 $25.00 $50.00
gpt-6.2-sol $2.00 $0.10 $2.50 $5.00 $10.00
gpt-6.1-sol $2.00 $0.10 $2.50 $5.00 $10.00

Astra's input and output rates fall by exactly 20%, and its cached read falls by 60%, from $1.00 to $0.40. That is because cache reads on GPT-6.2 Astra and Sol (and on GPT-6.1 Sol and Terra) cost 5% of the input rate, against 10% on most other recent models, including GPT-6 Astra. Long prompts follow the same rule on all four models: above 272,000 input tokens, input and cache rates double and output costs 1.5 times as much for the whole request.

The other tiers are where 6.2 and 6.1 diverge.

Processing option GPT-6.2 Sol GPT-6.1 Sol
Standard $2 input, $10 output $2 input, $10 output
Batch API 50% off: $1, $5 50% off: $1, $5
Flex Not available 50% off: $1, $5
Background, 8-hour window 40% off: $1.20, $6 Not available
Background, 2-hour window 30% off: $1.40, $7 Not available
Fast mode 2x: $4, $20 2x: $4, $20
Ultrafast 3x to 28 Feb 2027, then 6x 6x: $12, $60

Ultrafast on the 6.2 models costs three times Standard until 28 February 2027, 100 days from today, and six times from 1 March 2027 Pacific time; on GPT-6.1 Sol it is already six times. Our GPT-6.2 Ultrafast budget note covers that step-up. Regional processing endpoints add a 10% uplift for models released on or after 5 March 2026, which covers every model in these tables.

Rupee and sterling reference

The ECB reference rates of 19 November 2026 cross to ₹93.30 and £0.76 per US dollar, rounded to two decimals. At those rates GPT-6.2 Sol costs about ₹186.60 input and ₹933.00 output per million tokens, and GPT-6.2 Astra about ₹746.40 and ₹3,732.00. In sterling, Sol is about £1.52 and £7.60, and Astra about £6.08 and £30.40.

What changes in the Responses API

OpenAI's GPT-6.2 guide lists three changes that affect existing integrations.

1. Prefix edits drop reasoning

From GPT-6.2 onwards, the Responses API drops earlier reasoning items after mid-conversation edits to the conversation history, after certain model changes (including downgrades), or when the end user or organisation changes. OpenAI warns that answer quality can fall and token use can rise as the model repeats earlier work.

Common patterns that trigger this:

  • Rewriting an earlier developer message, for example to inject a fresh timestamp or a user's updated profile.
  • Trimming or summarising old tool results in place to save context.
  • A router that sends turn one of a conversation to Astra and turn two to Sol.

The fix is to keep the prefix unchanged and put anything dynamic at the end. The same discipline keeps earlier input eligible for prompt caching, so it pays twice.

2. Top-level instructions move after the input

This is the change most likely to raise a bill without anyone noticing. Models starting with GPT-6.2 append the top-level instructions field after input, beyond any cache breakpoint. Those tokens are billed at the uncached input rate on every request, even when the instructions never change, and they do not carry over with previous_response_id. To keep a long system prompt cacheable, put it in a developer message as the first item in input; that message stays in the conversation history.

Top-level instructions are still useful for guidance that applies only to the current request, such as asking for a summary on the final turn. You can change them between turns without disturbing the cached prefix.

3. Background windows replace Flex

GPT-6.2 Astra and Sol launch without Flex processing. For long-running Responses API work, OpenAI points to background mode with a completion window: set background to true, service_tier to default, and background_options.completion_window to 2h or 8h, with streaming off. The discount applies to each eligible request, including follow-ups; reusing the same window, continuation ID, model and project earns higher queue priority. Requests in these windows share the Batch rate limit, and Standard requests cannot reuse prompt cache entries created by 2h or 8h requests, even for identical prompts. Our guide to the 2h and 8h windows covers which workloads fit.

Flex itself is on the way out. OpenAI says it is winding down the Flex beta, will not add Flex to future models, and will publish a retirement date in the Flex guide once confirmed.

Smaller parameter changes

The migration quickstart in the same guide lists the remaining edits.

  • Reasoning effort. Both 6.2 models reject none. If you used none or minimal on gpt-5, start with low; for new work, start with medium.
  • Sampling parameters. Remove temperature, top_p and top_logprobs from Responses requests (and logprobs as well in Chat Completions), and remove message.output_text.logprobs from include.
  • Cache retention. Coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl set to 30m. A 12h value is available at 2.5 times the input rate for the write.
  • Pro mode. The reasoning guide shows reasoning.mode set to pro on gpt-6.2-sol. Pro mode bills its extra model work at the selected model's standard token rates.
  • Data residency. Both models support US and EU residency with Standard, Fast and Batch processing; Ultrafast is US-only.

A minimal Responses call that applies these points looks like this. The extra_body wrapper covers SDK versions that do not yet type the newer fields.


            from openai import OpenAI

client = OpenAI()
STABLE_RULES = open("system_prompt.txt").read()   # long, unchanging prompt

resp = client.responses.create(
    model="gpt-6.2-sol",                          # was "gpt-5"
    input=[
        {"role": "developer", "content": STABLE_RULES},  # cacheable prefix
        {"role": "user", "content": "Summarise ticket 4821 for the on-call engineer."},
    ],
    instructions="Reply in under 120 words.",     # this turn only, billed uncached
    reasoning={"effort": "low"},                  # was "minimal" on gpt-5
    extra_body={"prompt_cache_options": {"ttl": "30m"}},  # replaces prompt_cache_retention
)

u = resp.usage
print(u.input_tokens, u.input_tokens_details.cached_tokens, u.output_tokens)
          

For an overnight job, the background version adds three settings.


            job = client.responses.create(
    model="gpt-6.2-sol",
    input=[{"role": "developer", "content": STABLE_RULES},
           {"role": "user", "content": "Classify the attached 500 support tickets."}],
    background=True,
    service_tier="default",                       # discount needs Standard
    extra_body={"background_options": {"completion_window": "8h"}},  # 40% off
)
print(job.id, job.status)                         # poll or use a webhook
          

What the changes cost: three worked examples

Example 1: a chat assistant moving off gpt-5

Assume 300,000 requests a month. Each sends a 3,000-token system prompt that is cached, 1,000 new input tokens and 600 output tokens including reasoning. gpt-5 charges $1.25 input, $0.125 cached and $10 output per million tokens. GPT-6.2 Sol charges $2, $0.10 and $10, with cache writes at $2.50. Rupees use ₹93.30 per dollar.

Setup Cost per request Monthly, USD Monthly, INR Change vs gpt-5
gpt-5 today $0.007625 $2,287.50 ₹2,13,424 baseline
GPT-6.2 Sol, prompt in developer message $0.0083 $2,490 ₹2,32,317 +8.9%
Same, new tokens also written to cache $0.0088 $2,640 ₹2,46,312 +15.4%
GPT-6.2 Sol, prompt left in instructions $0.014 $4,200 ₹3,91,860 +83.6%

The workings for the last row: 3,000 tokens at $2 is $0.006, 1,000 at $2 is $0.002 and 600 at $10 is $0.006, so $0.014 a request, times 300,000 is $4,200. Against $2,287.50, that is 83.6% more. Moving the same 3,000 tokens into a developer message cuts that line from $0.006 to $0.0003. GPT-6.1 Sol has the same token rates, so the first two 6.2 rows equal the 6.1 Sol figures in our GPT-5 and o3 retirement cost comparison. The same shape on o3, at $2 input, $0.50 cached and $8 output, also comes to $0.0083 a request and $2,490 a month, so for o3 users GPT-6.2 Sol is cost-neutral at this mix.

Example 2: an asynchronous job that uses Flex today

Assume 4,000 million input tokens and 800 million output tokens a month of enrichment work that can wait. Rates are from the pricing page and background guide.

Route Calculation Monthly, USD Monthly, INR
GPT-6.2 Sol, Standard 4,000 x $2 + 800 x $10 $16,000 ₹14,92,800
GPT-6.2 Sol, 2-hour window 4,000 x $1.40 + 800 x $7 $11,200 ₹10,44,960
GPT-6.2 Sol, 8-hour window 4,000 x $1.20 + 800 x $6 $9,600 ₹8,95,680
GPT-6.2 Sol, Batch API 4,000 x $1 + 800 x $5 $8,000 ₹7,46,400
GPT-6.1 Sol, Flex 4,000 x $1 + 800 x $5 $8,000 ₹7,46,400

A Flex workload that moves to the best background window on 6.2 pays 20% more, $9,600 against $8,000. If the job tolerates the Batch API's file-based flow, Batch on 6.2 matches Flex on 6.1 exactly. The background route is the easier code change, because it stays on the Responses endpoint and supports follow-ups; Batch is cheaper. Since Flex is being retired on a date OpenAI has not yet set, staying on it only buys time.

Example 3: an Astra workload

Assume 50 million uncached input tokens, 150 million cached input tokens and 20 million output tokens a month, ignoring cache writes. On GPT-6 Astra: 50 x $10 + 150 x $1 + 20 x $50 = $1,650 (₹1,53,945). On GPT-6.2 Astra: 50 x $8 + 150 x $0.40 + 20 x $40 = $1,260 (₹1,17,558). That is 23.6% less, more than the headline 20%, because cached reads fall by 60%. Press reports also relay OpenAI's claim that the 6.2 models use fewer output tokens for the same work, which would widen the gap, but measure that on your own prompts before you budget for it.

Should you skip GPT-6.1 Sol?

For most teams still on gpt-5 or o3, yes. The token rates are identical, so the decision rests on features, and on balance they favour 6.2. The table draws on the model pages, pricing page, deprecations page and GPT-6.2 guide.

Factor Points to GPT-6.2 Sol Points to GPT-6.1 Sol
Token price Same Same
Knowledge cutoff 1 June 2026 30 April 2026
Ultrafast 3x until 28 Feb 2027 Already 6x
Async discount 30% or 40% background windows 50% Flex, being wound down
Cached system prompt in instructions Billed uncached Stays in cached prefix
Named in the deprecation notice No Yes, for gpt-5 and o3
Prefix edits Drop reasoning items No change documented

Skip to 6.2 Sol if your requests already keep a stable prefix, your system prompt can move into a developer message, and you have no Flex workloads. You do one migration instead of two, and you get the later cutoff and cheaper Ultrafast.

Take 6.1 Sol first if your cost model depends on Flex at 50%, if your agent framework rewrites earlier turns and you cannot change it before 11 December, or if your change-control process needs the exact model OpenAI named as the replacement. GPT-6.1 Sol is not on the deprecations list today, so it is a safe landing point for now.

Consider Astra only by evaluation. At four times Sol's rate, GPT-6.2 Astra makes sense where Sol falls short on your own evals. If you run GPT-6 Astra today, moving to 6.2 Astra is a straight price cut once the instructions change is handled.

Whichever you choose, do not mix them inside one conversation, because model changes can drop reasoning items on 6.2. If you route through a gateway, make the model a routing alias per workload.

A 21-day plan to 11 December

  1. By 24 November. List every gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro call from your usage dashboard, not only from code search. Tag each as interactive, Flex, Batch or pro. Our GPT-5 and o3 shutdown migration guide covers the inventory step.
  1. By 26 November. Move stable system prompts from instructions into a first developer message. Remove temperature, top_p and logprob settings, and map none or minimal effort to low.
  1. By 1 December. Run a golden set of 100 to 300 real prompts on GPT-6.2 Sol and GPT-6.1 Sol at the same effort. Record quality, output tokens and cached tokens from the usage object.
  1. By 4 December. Ship behind a flag, send 10% of traffic for 24 hours, then 100%. Move Flex jobs to an 8-hour window or the Batch API, or keep them on 6.1 Sol for now.
  1. By 9 December. Delete the old model strings so nothing falls back to them before the 11 December shutdown.

India-specific considerations

Indian teams pay OpenAI in US dollars, so the rupee figures above move with the exchange rate as well as with token counts. At ₹93.30 per dollar, the gap between the cached and uncached setups in Example 1 is ₹1,59,543 a month ($1,710), before any GST treatment your finance team applies to imported services.

Prompts that carry customer or employee data fall under the Digital Personal Data Protection Act 2023, so a migration is a good moment to check that each workload's prompts match its notice and purpose. The GPT-6.2 guide lists US and EU data residency only, and regional endpoints carry the 10% uplift. Our guide to OpenAI regional processing and India covers what that means for storage and processing.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our LLM migration and cost optimisation service can inventory your gpt-5 and o3 calls, move cached prompts into developer messages, and test GPT-6.2 Sol against GPT-6.1 Sol on your own prompts before 11 December. Talk to us at /contact-us/.

Last updated: 20 November 2026.

References

  1. API changelog, entries of 17 November 2026 — OpenAI
  1. API pricing, checked 20 November 2026 — OpenAI
  1. Deprecations, checked 20 November 2026 — OpenAI
  1. Using GPT-6.2 — OpenAI
  1. Prompt caching guide — OpenAI
  1. Background mode guide — OpenAI
  1. Flex processing guide — OpenAI
  1. OpenAI launches GPT-6.2 Astra and Sol on ChatGPT and APIs, 17 November 2026 — TestingCatalog
  1. GPT-6.2 Astra model page — OpenAI
  1. GPT-6.2 Sol model page — OpenAI
  1. GPT-6.1 Sol model page — OpenAI
  1. GPT-6 Astra model page — OpenAI
  1. GPT-5 model page — OpenAI
  1. Euro foreign exchange reference rates, 19 November 2026 — European Central Bank
  1. Reasoning models guide — OpenAI
  1. o3 model page — OpenAI

Frequently asked

Quick answers.

01 When were GPT-6.2 Astra and GPT-6.2 Sol released?
OpenAI's API changelog dates both models to 17 November 2026. They are available as gpt-6.2-astra and gpt-6.2-sol on the Responses API, Chat Completions and the Batch API, though tool calling requires the Responses API. Press reports say both also rolled out in ChatGPT Work and Codex for paid plans that day.
02 How much do GPT-6.2 Astra and GPT-6.2 Sol cost?
At Standard rates for prompts up to 272,000 input tokens, GPT-6.2 Astra costs $8 input, $0.40 cached input and $40 output per million tokens. GPT-6.2 Sol costs $2, $0.10 and $10. Batch halves both. Longer prompts double input rates and raise output by half for the whole request.
03 Is GPT-6.2 Sol cheaper than GPT-6.1 Sol?
No. Standard, cached, cache-write, Batch and Fast rates are identical. The differences are elsewhere: GPT-6.2 Sol has no Flex tier but offers background windows at 30% or 40% off, and its Ultrafast mode costs three times Standard until 28 February 2027, against six times on GPT-6.1 Sol.
04 What changes in the Responses API with GPT-6.2?
Two behaviours change. Editing earlier turns, switching models or changing the end user now drops stored reasoning, which can lower quality and raise tokens. Top-level instructions are appended after the input and billed uncached every time. Put a reusable system prompt in a developer message as the first input item instead.
05 Should a team migrating off gpt-5 or o3 skip GPT-6.1 Sol?
Usually yes, because the token rates are the same and GPT-6.2 Sol has a later knowledge cutoff and cheaper Ultrafast. Stay on GPT-6.1 Sol for now if you rely on Flex at half price, cannot stop rewriting earlier turns before 11 December, or must use the exact replacement OpenAI named.
06 What happens on 11 December 2026?
OpenAI's deprecations page lists 11 December 2026 as the shutdown date for gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro, including their dated snapshots. Fine-tuned versions follow the same date as their base models. After that the models stop serving API requests, so any hard-coded model string will return errors.
07 Does GPT-6.2 support Flex processing?
No. OpenAI says new models from GPT-6.2 Astra and Sol onwards launch without Flex, and it is winding down the Flex beta on older models. Use background mode with a two-hour window for 30% off or an eight-hour window for 40% off, on Standard processing. The Batch API still gives 50% off.
08 Is GPT-6.2 Astra worth the price over Sol?
Only where your own evaluations show Sol falls short. Astra costs four times as much per token. For teams already on GPT-6 Astra, the move is a clear saving: input and output rates fall by 20% and cached reads by 60%, so a cache-heavy workload can save more than 20% overall.

About the author

Manu Shukla

Founder & Director

Founder of eCorpIT. Hands-on engineer leading senior-only delivery for AI apps, custom software, and cloud systems for global clients.

Subscribe

One engineering note a week. No fluff, no spam.

Senior-architect playbooks on AI agents, mobile apps, cloud, security, data, and marketing — delivered every Wednesday.

Past the reading

Read enough. Let's build something.

A senior architect responds in 24 working hours with scope, indicative cost, and a timeline. NDA before any technical conversation.

Loading search…

Browse all guides