TL;DR: OpenAI launched GPT-6 Astra on September 3, 2026, priced at $10 per million input tokens and $50 per million output tokens through the API, 2.5 times the rate of its predecessor GPT-5.6 Sol; it is also the first OpenAI model rated Critical for cybersecurity under the company's own Preparedness Framework.
After eighteen months of price war across AI models, the top end is heading back up. For an SME wiring AI into its processes, this is not lab news: it is management information, with direct consequences on the monthly bill.
Key takeaways
- OpenAI launched GPT-6 Astra on September 3, 2026: a 1.05-million-token context window, 128,000 output tokens, and a rollout across Plus, Pro, Business and Enterprise plans, the API and AWS.
- GPT-6 Astra's API price is $10 per million input tokens and $50 per million output tokens, against $4 and $20 for GPT-5.6 Sol, a 150% increase.
- On the same debugging test run by ComputingForGeeks, one request costs about $0.19 with GPT-6 Astra versus $0.05 with GPT-5.6 Sol, for a more rigorous answer at four times the price.
- GPT-6 Astra reaches the Critical cybersecurity threshold of OpenAI's Preparedness Framework: full capability is initially reserved for a defensive security program, and the public version is restricted.
The facts: what OpenAI shipped on September 3
OpenAI announced GPT-6 Astra on September 3, 2026, with no livestream, and API access opening on September 5. The model offers a 1.05-million-token context window, up to 128,000 output tokens, and five reasoning-effort levels adjustable per request. The New Stack sums up the stated ambition in its headline: OpenAI is welcoming readers to the "AGI era".
The measured gains are real on agent tasks. According to ComputingForGeeks' comparison, GPT-6 Astra scores 57.9% on Terminal-Bench 4.0 against 37.3% for GPT-5.6 Sol, 72.6% on OSWorld 2.0 (computer use) against 65.7%, and 97.6% on the research-grade mathematics of FrontierMath Tier 4. The measured hallucination rate drops to 4.2%, from 12.2% for Sol.
The most spectacular number, however, deserves a magnifying glass. OpenAI reports 99.9% on ARC-AGI-3, a reasoning test built on unseen games; but that score comes from OpenAI's own evaluation harness. On the standard harness, comparable across models, Explainx's analysis records 62.7%. The same caution applies to general intelligence: the Artificial Analysis index moves from 60.9 to 61.2, a marginal gain, and Explainx notes that GPT-6 Astra still trails Claude Fable 5.1 on that index. The progress is concentrated exactly where OpenAI wants to sell: agents that work for a long time, alone, on a computer.
Why is the price going up while the market was going down?
Because the market has split in two. At the bottom, the price war goes on: Chinese models and mid-tier offerings keep pushing the cost of routine tasks toward zero, a movement I covered in the article on the Qwen and DeepSeek price war. At the top, the labs have understood that long agentic tasks, the ones that replace hours of skilled work, can carry a premium. GPT-6 Astra owns that positioning: no mini or nano tier at launch, a surcharge beyond 272,000 tokens of context, and a fast mode billed at double the rate.
The table below sums up the published API price list, in dollars per million tokens.
| Item | GPT-6 Astra price | Reference point |
|---|---|---|
| Standard input | $10 | 2.5x GPT-5.6 Sol ($4) |
| Standard output | $50 | 2.5x GPT-5.6 Sol ($20) |
| Cached input reads | $1 | 90% off standard input |
| Batch processing | $5 and $25 | Half price, for deferred workloads |
| Long context (beyond 272,000 tokens) | $20 and $75 | Roughly double the standard rate |
| Fast mode | Double rate | Up to 2x the processing speed |
The practical consequence shows up in unit tests of real usage. ComputingForGeeks submitted the same FastAPI debugging task to both models at the same effort level: GPT-6 Astra returns a more complete answer, with checks Sol omitted, but the request costs about $0.19 against $0.05. Multiplied by thousands of calls a month, the gap becomes a budget line.
A model rated Critical for cybersecurity
The launch's other headline is as much regulatory as technical. GPT-6 Astra is the first OpenAI model to reach the Critical cybersecurity threshold of its Preparedness Framework, the internal system that assesses dangerous capabilities before release: the model can discover unknown security flaws and exploit them without step-by-step guidance, and it scores 100% on the ExploitBench test suite. That is precisely the threshold that led OpenAI to suspend the model's release in early August, an episode I analyzed in the article on the Astra suspension.
OpenAI's answer is a staged rollout: full capability goes first to Daybreak, its defensive security program, while the public version is restricted on offensive uses. The published robustness numbers point the right way: the success rate of indirect prompt injections falls to 8.5%, against 27% for GPT-5.6 Sol. One regression, though, deserves attention in regulated sectors: the model no longer returns its full reasoning but paraphrased summaries, which complicates auditing and compliance logging. For a company that must prove why an AI-assisted decision was made, that is a concrete step backwards.
What does this change for an SME?
The main change fits in one sentence: the AI bill becomes an architecture decision, no longer a fixed cost line. As long as every model converged downward in price, picking "the best one" cost almost nothing. With a flagship at $50 per million output tokens, putting the wrong model on the wrong task costs money every single month.
Three measured levers stand out from the published tests. Reasoning effort first: on the same request, ComputingForGeeks measures from $0.03 at minimal effort to $0.30 at maximum effort, a 10x factor adjustable per call. Prompt caching next: on a working document resent with every call, the tested bill drops from $0.28 to $0.02, a 91.9% saving. Batch processing last, at half price for anything that does not need an immediate answer.
The fourth lever remains the most important one: reserve the frontier model for tasks that justify it. Most SME workloads (email triage, data extraction, summaries, standard replies) do not need a model rated Critical for cybersecurity. My daily financial dashboard, featured on the projects page, runs on Claude Haiku 4.5, a small model: to summarize a portfolio every morning, it does the job for a few cents. That is exactly the kind of trade-off OpenAI's new price list now makes mandatory.
GPT-6 Astra does not make AI more expensive; it makes discernment more profitable. Companies that know which task deserves which model just gained a cost advantage over those that plug the flagship in everywhere. The question to ask is not "should we move to GPT-6?" but "which of my tasks are worth $50 per million tokens?".
Frequently asked questions
How much does GPT-6 Astra cost through the API?
GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, 2.5 times the price of GPT-5.6 Sol. Cached input reads drop to $1, batch processing runs at half price, and fast mode doubles the standard rate.
Should an SME move its workloads to GPT-6 Astra?
Not by default. GPT-6 Astra improves mostly on long agentic tasks; on general intelligence the gain is marginal. For sorting, extraction or standard replies, a smaller model does the same job for a fraction of the price.
Why is GPT-6 Astra rated Critical for cybersecurity?
Under OpenAI's Preparedness Framework, the model can discover and exploit security flaws without step-by-step guidance. Full capability is initially reserved for a defensive security program; the public version is restricted.