Astra Token Rates Against Sol Predecessor

Pricing page origin

Source 2 records the figures on the OpenAI API pricing page dated after the September 3 2026 limited preview. The same page places both models in the flagship models table and applies identical column headers for standard short context and long context. Wikipedia notes the predecessor relationship between GPT-6 Astra and GPT-5.6 yet supplies no pricing data of its own.

Source 1 repeats the same rates on the model detail page for gpt-6-astra and lists the knowledge cutoff of April 30 2026 together with the 1 050 000 token context window. Both sources are maintained by OpenAI and updated on the same schedule as the model rollout to the Trusted Access Program.

Source 3 states that GPT-6 Astra entered limited preview on September 3 2026 for trusted partners with public release scheduled for September 7 2026. The page records the model as the direct successor to GPT-5.6 and notes the largest training run to date on more than 100 000 GPUs at the Stargate site.

OpenAI vice president of research Aidan Clark told reporters that development involved by far their largest training run: It is the first time we have pretrained on more than 100 000 GPUs at our Stargate site in Texas.

Token category definitions

The quoted numbers cover only text token billing under the standard processing tier for prompts below the 272 000 token threshold. Source 2 defines input tokens as those sent in the request, cached input as previously stored tokens billed at a lower rate, cache writes as new entries charged at 1.25 times the uncached input rate, and output tokens as those generated by the model.

Source 1 adds that prompts exceeding 272 000 input tokens receive 2 times input and cache rates plus 1.5 times output for the entire request. The rates therefore measure only the four named token categories on the short context standard column and nothing else.

Reasoning effort settings listed in source 1 range from low through max and affect token consumption without changing the published per token prices. The 128 000 maximum output tokens per response also remain outside the headline pricing line. Source 1 further specifies that the model supports text and image input only with no audio or video modalities recorded.

Excluded cost factors

Source 2 states that batch and flex processing receive a 50 percent discount while fast mode receives a 200 percent premium. The same page notes a 10 percent regional processing uplift for models released on or after March 5 2026 when data residency is selected. Neither rate appears inside the headline standard short context column.

Source 1 lists supported tools for the Responses API that include computer use, hosted shell, apply patch, code interpreter and web search. Tool call fees sit outside the per token table and are billed separately at $2.50 per 1 000 calls for file search or $10.00 per 1 000 calls for web search. Modalities are restricted to text input output and image input only, with no audio or video support recorded for gpt-6-astra.

Rate limits in source 1 begin at tier 1 with 500 requests per minute and 500 000 tokens per minute and scale to tier 5 at 15 000 requests per minute and 40 000 000 tokens per minute. These caps are not reflected in the per token prices. Source 2 further records that fast mode is unavailable for GPT-6 Astra with EU data residency.

Container sessions for hosted shell or code interpreter add separate memory based charges from $0.03 for 1 GB to $1.92 for 64 GB per 20 minute block. Web search content tokens are billed at model rates in addition to the $10.00 per 1 000 calls base fee. Source 1 also records that fine-tuning is not supported for gpt-6-astra while the model accepts endpoints including v1/chat/completions, v1/responses and v1/realtime.

Workload cost variations

A workload that stays under 272 000 input tokens and uses standard processing pays the headline rates exactly. The same workload moved to batch processing pays half those rates while fast mode doubles them. Long context requests above the threshold pay the separate long context column: $20 input and $75 output for gpt-6-astra versus $8 input and $30 output for gpt-5.6-sol.

Cache writes at $12.50 for gpt-6-astra versus $5.00 for gpt-5.6-sol add further cost when prompts change frequently. Source 2 records no enterprise tier override or volume commitment discount beyond the published batch and flex columns. GPT-5.6 Sol promotional pricing remains available at least through November 21 2026 on the same page.

The documented rates would prove incorrect if OpenAI revised the long context multiplier or introduced a new cache write discount not present in the September 2026 pages.