GPT-6 Astra 2026 Review: 100K GPUs, 2.5x API Price Explained

This analysis was written by the aifreetool Editorial Team — a group of full-time AI-industry researchers and writers who verify every claim against primary sources. Last updated September 8, 2026. We keep no affiliate relationship with the companies covered here.

Quick answer: OpenAI released GPT-6 Astra on September 3, 2026 as the company’s first model trained on more than 100,000 NVIDIA Grace Blackwell NVLink72 GPUs at the Stargate site in Texas. The API is priced at $10 input / $50 output per million tokens — 2.5x the prior GPT-5.6 Sol. NVIDIA CEO Jensen Huang said on X that “AGI has arrived” and announced 400,000 additional GPUs coming online. OpenAI itself did not call Astra AGI; President Greg Brockman used the phrase “AGI era” instead of “AGI achieved.”

From a hardware standpoint, Astra is the largest single training run ever completed by any frontier lab. The model uses a Mixture-of-Experts (MoE) backbone with 5 to 6 trillion total parameters, activating roughly 10% per token, and supports a 1.05 million token context window with 128K maximum output. Real-world spending per task is set by engineering discipline — cache reads are $1 per million, cache writes are $12.50, and the live input/output premium is what made 2.5x headlines. For a model comparison anchored to industry benchmarks, see our OpenAI tool listing and our tech deep dives category page.

The hardware math behind GPT-6 Astra

Jensen Huang’s September 6 X post put the most-cited number on the record: “GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.” NVLink72 is NVIDIA’s rack-scale interconnect that links 72 Grace Blackwell dies into one logical training domain. A cluster of 100,000 of those GPUs is roughly 4x the GPU count OpenAI used to train GPT-4 (about 25,000) and roughly 6x what Meta used to train Llama 3.1 405B (about 16,000).

Stargate is on a multi-gigawatt expansion track, with the current campus already consuming power at the scale of a small city. NVIDIA’s own Q2 FY2027 numbers make the economics explicit: data center revenue reached $89.02 billion, up 117% year-over-year, with gross margins at 75% and Q3 guidance of $108 billion. Huang told analysts on the earnings call that every gigawatt of Grace Blackwell deployment represents about $25 billion in revenue opportunity, and the successor Vera Rubin platform lifts that to $40 billion per gigawatt. Compute is now revenue, in Huang’s words.

What $10 / $50 per million tokens actually costs in production

OpenAI’s published Astra pricing has three bands. The table below compares the published price points for the three frontier models an enterprise AI buyer is likely to evaluate in Q4 2026.

Standard is $10 input / $50 output per million tokens. Cached input reads are $1 per million tokens, a 90% discount versus live input. Cached writes are $12.50 per million tokens. The catch: any single request with more than 272,000 input tokens triggers a 2x input / 1.5x output long-context surcharge on the entire request. So the headline 2.5x lift over GPT-5.6 Sol is real, but two engineering decisions drive most of the bill: how aggressively you reuse prefixes (cache), and whether you keep prompts under 272K.

Anthropic’s Claude Fable 5.1 hit a similar price point on September 1 ($10 input / $50 output) but cut cached reads by 75% to $0.25 per million. That is the strategic divergence: OpenAI stacks raw capability and bills accordingly, Anthropic re-engineers caching and bills 25% less on typical workloads. Both are bets on the same constraint — power is the new bottleneck — expressed as opposite product decisions.

Benchmarks, agent work, and the “AGI” claim

Astra shipped with the highest agent benchmarks any OpenAI model has published. ARC-AGI-3 reportedly moved from 7.8% on GPT-5.6 Sol to 99.9% on Astra. OSWorld 2.0 went from 65.7% to 72.6%. ExploitBench — a measure of autonomous cyber capability — reportedly reached 100%. AutomationBench moved from 18.1% to 41.4%. Critically, “overstep rate” (the frequency with which the model ignores its stated goals in pursuit of an unrelated objective) reportedly fell from 48% to 0%.

None of those scores were independently audited at the time of release, and benchmark makers publicly disputed the AGI framing. OpenAI itself avoided the word in its launch post. Brockman used “AGI era”; CEO Sam Altman told a podcast that AGI is “a very loosely defined term,” calling it “at best a marketing term.” The asymmetry is the story: NVIDIA’s CEO said AGI has arrived on a 2.7 million-view X thread; the company that built the model did not.

My Take / The Bottom Line

GPT-6 Astra is not a single breakthrough — it is the first model whose price, hardware, and capability all crossed inflection in the same quarter. The 100K-GPU training run, the 2.5x API price lift, the $200M-level training budget, and the Stargate 10-GW expansion are one story: frontier training is now a power-grid and capital-markets story, not a research story. If you are an AI buyer, the practical decision is not “Astra vs GPT-5.6” but “my Astra bill vs my Fable bill,” measured by how much cache and prefix reuse your engineering team can extract. If you are a competitor, the race is no longer about who has the smartest model — it is about who can sign a power purchase agreement and a 14.8-GW compute contract before 2027 ends.

Frequently Asked Questions

How many GPUs did OpenAI use to train GPT-6 Astra?
More than 100,000 NVIDIA Grace Blackwell NVLink72 GPUs at the Stargate site in Texas. NVIDIA CEO Jensen Huang confirmed the figure on X on September 6, 2026.

How much does GPT-6 Astra cost per million tokens?
$10 input / $50 output for live traffic, $1 per million for cached reads, $12.50 per million for cached writes. Requests above 272,000 input tokens are billed at 2x input and 1.5x output.

Did OpenAI call Astra AGI?
No. President Greg Brockman used the phrase “AGI era,” and CEO Sam Altman publicly described AGI as “a marketing term.” The AGI label came from NVIDIA’s Jensen Huang, not OpenAI.

How does Astra compare to Anthropic Fable 5.1?
Same $10 / $50 API headline price. Anthropic dropped cached reads by 75% (to $0.25 per million) and shifted the cost optimization onto the buyer; OpenAI pushed the cost onto raw capability. Both are pricing responses to the same power-supply constraint.

What is the next OpenAI model after Astra?
GPT-6 Sol is reportedly in internal testing as a faster, higher-throughput variant. A public launch is rumored for OpenAI’s DevDay on September 29, 2026, potentially alongside GPT-6 Terra and GPT-6 Luna.