DEV Community

Cover image for GPT-6.2 Astra cuts cache reads 60%: what the rate card changes
AIHubMix
AIHubMix

Posted on Edited on

GPT-6.2 Astra cuts cache reads 60%: what the rate card changes

The engineering question after OpenAI's GPT-6.2 Astra and GPT-6.2 Sol launch is not whether the context window grew. It did not. The official community announcement appeared at 02:00 on 18 November 2026, Beijing time; the launch itself is dated 17 November. What changed for anyone already calling GPT-6 Astra (gpt-6-astra) is the Standard rate card: GPT-6.2 Astra (gpt-6.2-astra) is 20% lower on input and output, and 60% lower on cached input.

On a cache-heavy workload, that larger percentage cut can make the total saving exceed 20%. How much depends on the token mix. GPT-6.2 Sol (gpt-6.2-sol) keeps GPT-6.1 Sol's list price; its case rests on task quality at the same rates.

The envelope did not change

GPT-6.2 Astra keeps the same limits as GPT-6 Astra: a 1,050,000-token context window, 922,000 maximum input tokens, 128,000 maximum output tokens, text and image input, and text output. The GPT-6.2 Astra model page still rejects a reasoning effort of none. Those limits also apply to Sol.

The knowledge cutoff moved. GPT-6 Astra and GPT-6.1 Sol both stop at 30 April 2026. GPT-6.2 Astra's cutoff is 18 May 2026. GPT-6.2 Sol's is later still, 1 June 2026. If you pick the flagship for "fresher knowledge," check the dates; Sol's listed cutoff is later, although that does not guarantee knowledge of any particular event.

The other change is a migration rule, not a capability: starting with GPT-6.2, the API drops reasoning items when it sees a mid-conversation prefix edit, a model downgrade, or a change of end user or organization. Treat an ID swap as a contract change, not a no-op. The migration post in this series covers the request shape.

OpenAI also describes GPT-6.2 Astra as more concise by default and better at professional documents, frontends, and long-horizon work. These are vendor claims.

Official rates, and one assumed bill

The figures below are OpenAI's Standard rates per million tokens, from the official pricing page, for short-context requests (272,000 input tokens or fewer). Comparisons in this series use official prices on every side.

Model Input Cached input 30m cache write Output
GPT-6 Astra $10.00 $1.00 $12.50 $50.00
GPT-6.2 Astra $8.00 $0.40 $10.00 $40.00
GPT-6.1 Sol / GPT-6.2 Sol $2.00 $0.10 $2.50 $10.00

A 12-hour cache write is twice the 30-minute write rate on each of these models. Requests with more than 272,000 input tokens are billed at double the input and cache rates and 1.5 times the output rate for the full request. For the 6.2 models, OpenAI lists Ultrafast at 3× Standard through 28 February 2027 Pacific time and 6× starting 1 March Pacific time. The calendar boundary corresponds to 1 March at 16:00 Beijing time; the docs do not give a more precise activation timestamp. Those tiers are OpenAI's; this post does not assume they are available through a gateway.

Both GPT-6.2 Astra and GPT-6.2 Sol are now available on AIHubMix. The public index and resolved schemas list Chat Completions, Responses, and Anthropic Messages for both IDs. Messages is provided through AIHubMix gateway translation, not an OpenAI-native endpoint. This was checked at approximately 23:04–23:06 Beijing time on 18 November 2026. The example below uses official prices; no inference request or gateway bill was measured for this article.

For a deliberately simplified ten-request billing example, assume 80,000 cache-write tokens once, 720,000 cache-read tokens across nine later requests, 18,000 ordinary uncached input tokens in total, and 30,000 billed output tokens in total, including reasoning. Assume every reuse hits within the 30-minute retention window, no other tokens or tool fees, and no request exceeds 272,000 input tokens. These are assumed billing categories, not a measured, cumulatively growing conversation. Cache-write rates replace, rather than add to, the ordinary input rate for those tokens.

  • GPT-6 Astra: $1.00 write + $0.72 cache reads + $0.18 extra input + $1.50 output = $3.40.
  • GPT-6.2 Astra: $0.80 write + $0.288 cache reads + $0.144 extra input + $1.20 output = $2.432.
  • GPT-6.2 Sol: $0.20 write + $0.072 cache reads + $0.036 extra input + $0.30 output = $0.608.

On this workload GPT-6.2 Astra is $0.968, or about 28%, cheaper than GPT-6 Astra — more than the 20% list cut, because cached input fell from $1.00 to $0.40. Sol, at the same token counts, is exactly one-fourth of GPT-6.2 Astra, because every Standard rate in the table is 4×. For a separate no-cache baseline, bill all 818,000 input tokens at ordinary input rates, with no cache-write charges, and keep the same 30,000 billed output tokens: $9.68 on GPT-6 Astra and $7.744 on GPT-6.2 Astra, exactly 20% less. That is a different billing assumption, not a claim that cache misses avoid write charges.

Keep the effort setting in the regression check

These results come from OpenAI's launch charts and text, not tests run for this article.

On Agents' Last Exam, OpenAI reports GPT-6.2 Astra at xhigh scores 78.3% at $2.89 per task versus GPT-6 Astra at max, 73.1% at $8.99: a 5.2-point gain at less than one-third the cost. This is not a same-effort comparison. With both at max, 6.2 Astra scores 77.6% at $4.47: a 4.5-point gain at about half the cost. On Terminal-Bench-Science 0.1, both at max, the gain is 9 percentage points at less than one-third the cost. Internal GDPval v2, also both at max, improves by 6.2 percentage points with 22% fewer output tokens.

The low-effort rows show why those gains cannot stand in for a regression check. On LifeVal at low effort, GPT-6 Astra scores 67.5%, versus 46.3% for GPT-6.2 Astra; GPT-6.1 Sol scores 66.8%, versus 25.3% for GPT-6.2 Sol. On GDPval v2 at low, Sol falls from 28.8% to 23.1%. These substantial drops apply to those tasks and effort settings; they do not establish a regression on every low-effort workload.

Retest existing low-effort jobs before switching IDs. For a new integration, OpenAI recommends starting explicitly at medium. That recommendation does not document the default when the field is omitted. Compare completion quality, latency, and the whole bill at the effort you actually use. Fewer output tokens could reduce task cost further if the vendor's efficiency claims hold on your traffic, but that saving is not included in the fixed-token example above.

Sol at the same price

GPT-6.2 Sol keeps GPT-6.1 Sol's $2 / $0.10 / $10 input / cached-input / output rates, as well as its context and output limits. The upgrade case is therefore quality at the same rate, not a list-price cut. OpenAI reports that 6.2 Sol improves over 6.1 Sol on its internal GDPval v2. That is a direct predecessor comparison, but still vendor-reported: the 22% output-token reduction cited above belongs to Astra versus GPT-6 Astra, not Sol versus 6.1 Sol. Sol's knowledge cutoff also moves from 30 April to 1 June 2026.

For existing 6.1 Sol users, evaluate 6.2 Sol on a few previously difficult multistep jobs, with explicit effort settings and the new prefix rule in place. There is no price increase in the official Standard token rates, but equal rates do not guarantee equal task cost or eliminate regressions. Separately, OpenAI reports that Sol comes within 0.5 points of 6.2 Astra on Cognition's FrontierCode 2.0 at about one-fifth the cost, both at max. That sibling comparison is the tier post's subject, not evidence of the size of the gain over 6.1 Sol.

Astra is also the wrong step if your traffic already belongs on a cheaper GPT-6 tier. Official Standard rates for GPT-6.1 Terra are $1 / $5, and for GPT-6 Luna $0.10 / $0.50. Even after the 20% cut, a $40 output rate does not make the flagship a volume model. If the work is everyday agent coding and the official FrontierCode claim is even roughly right, Sol is the first ID to try, not Astra.

What to do

Prioritize an upgrade evaluation if you are already on GPT-6 Astra and your bill has a substantial cache-read component. The list price is lower, the cache-read price is much lower, and the context window is the same. Revisit prompts that say "be concise" — OpenAI warns that Astra is more concise by default — and keep conversation prefixes stable so you actually get the cache rate in the table.

Start with a small regression run on AIHubMix. Both IDs are in the OpenAI models list. Select the documented protocol your client uses, set effort explicitly, and compare task quality and actual spend before moving production traffic.

If GPT-6.1 Sol is already your default, compare it with 6.2 Sol before changing tiers. The 6.2 Sol upgrade is a same-price quality bet, not an automatic bill reduction. Keep 6.1 Sol as a fallback until your regression checks pass. Stay on Terra or Luna if those tiers already clear the work; 6.2 Astra will still be several times the price.

Read the migration post before moving a multi-turn agent to a 6.2 ID: prefix changes can cause the API to drop earlier reasoning items.

Keep reading: the GPT-6.2 series

Sources

  1. Introducing GPT-6.2 Astra and Sol — official launch claims and interactive charts, rechecked in-browser at approximately 03:38–03:39 UTC on 18 November 2026; these remain vendor-reported evaluations.
  2. OpenAI API pricing — Standard, long-context, and Ultrafast rates.
  3. GPT-6.2 Astra model page and GPT-6.2 Sol model page — context, output limit, cutoff, and rejected none effort.
  4. GPT-6 Astra, GPT-6.1 Sol, and Using GPT-6.2 — predecessors and the explicit medium starting recommendation.
  5. GPT-6.2 Astra on AIHubMix, GPT-6.2 Sol, and parameter index — available models and documented protocols, checked approximately 15:04–15:06 UTC on 18 November 2026; no inference test.

Top comments (0)