GPT-6 Astra API Pricing: $10 and $50, Double GPT-5.6 Sol

GPT-6 Astra costs $10.00 per million input tokens and $50.00 per million output tokens. That is double GPT-5.6 Sol’s input rate and two thirds more on output. OpenAI shipped it on 3 September 2026 as gpt-6-astra, with a 1.1M context window, a 128K output ceiling and an April 2026 knowledge cutoff.

Model ID:      gpt-6-astra
Standard:      $10.00 in / $50.00 out per 1M
Cached input:  $1.00 per 1M
Fast mode:     ~$20 / ~$100, up to 2.5x standard speed
Context:       1.1M in / 128K out
Cutoff:        April 2026
Variants:      Astra, Astra Pro (no Sol / Terra / Luna split this generation)
Access:        OpenAI API, ChatGPT paid plans, Amazon Bedrock
Gateways:      Ofox as openai/gpt-6-astra (listed 2026-09-05)

Last updated 5 September 2026. GPT-6 Astra is now live in the Ofox catalog; the access section below reflects that.

How Much Does GPT-6 Astra Cost?

$10 in, $50 out, and the cache read is where it differs most from its neighbours.

Fast mode is the unusual line. OpenAI describes it as up to 2.5x standard speed at 2x standard price. That is a latency lever rather than a capability one: the same model, answering sooner, for double. Whether that trade is worth taking depends entirely on whether a human is waiting on the response.

GPT-6 Astra vs GPT-5.6 Sol: The Price Went Up

This is a generation that costs more, not less.

Both rates are the live Ofox catalog on 5 September 2026. Every row doubled or close to it.

That direction is worth pausing on, because it is not the recent norm. Wan 3.0 halved its predecessor’s rate and Gemini 3.8 Flash held price flat against its previous generation. Astra is priced as a step up in class rather than a refresh, and the upgrade question has to be answered on output quality rather than on cost. The Astra against GPT-5.6 Sol comparison works through when the extra pays for itself, and why the Sol / Terra / Luna tier names did not carry over.

What It Actually Costs per Task

$1.67 per task at max effort, $0.96 at high, on Artificial Analysis’s measurement.

The gap between those two rows is the practical pricing decision on this model. Max writes 2.6x the output tokens and costs 2.1x as much across the index, and buys one index point. On a workload narrower than AA’s nine-evaluation mix, that point may not exist at all.

Start on high. Measure whether max changes your outputs before paying double for it. Reasoning tokens bill at the output rate whether or not you read them, so an effort setting is a price multiplier with a quality hypothesis attached.

Where GPT-6 Astra Ranks

#8 of 202 models on the Artificial Analysis Intelligence Index, at 61. Strong, and specifically not first. Every model above it is a Claude or Muse Spark variant:

Two readings of that table, both true:

• At 61, Astra is not the cheapest way to get 61. Muse Spark 1.3 at xhigh reaches the same score for $0.55 per task, roughly a third of Astra’s cost.
• At $10 / $50, Astra is not the strongest model at that price. Claude Fable 5.1 charges the same per-token rate and scores five points higher. The full comparison covers where each one still wins, including the workloads where OpenAI’s own benchmarks favour Astra.

That 61 is the number to keep in view, because it did not move from GPT-5.6 Sol. Our full review works through what a flat index means next to a 2.5x price rise, why the headline ARC-AGI-3 score has two very different values depending on the harness, and what OpenAI’s own system card says about the model’s reasoning becoming harder to monitor.

The Benchmarks OpenAI Published

The vendor’s own numbers lean agentic rather than general:

Read the qualifiers with the numbers: ARC-AGI-3’s 99.9% is with an adapter harness, and Humanity’s Last Exam at 57.2% is with tools. Those are legitimate configurations and they are also not the same test as the unqualified version. Vendor benchmarks on vendor-chosen tasks answer “what is this model good at”, and a third-party composite answers “how does it compare”; they are different questions.

How to Access GPT-6 Astra

Four routes, and one of them is off by default.

• OpenAI API. Model ID gpt-6-astra.
• ChatGPT. Plus, Pro, Business and Enterprise, rolling out over the days after the 3 September launch. Astra Pro is the higher tier on Pro, Business and Enterprise plans.
• Amazon Bedrock. Supported from launch.
• Enterprise workspaces. Off by default; an admin enables it per workspace. If your organisation has Enterprise and the model is missing, that is the first thing to check rather than a rollout delay.

from openai import OpenAI

client = OpenAI(api_key="YOUR_OPENAI_API_KEY")

r = client.chat.completions.create(
    model="gpt-6-astra",
    messages=[{"role": "user", "content": "Refactor this function to remove the nested loop."}],
)
print(r.usage)

Read usage on your own prompts before forecasting. At $50 per million output tokens with reasoning billed at that rate, the token count matters more than the rate card.

Through Ofox

GPT-6 Astra is live in the Ofox catalog as of 5 September 2026, as openai/gpt-6-astra, at the same $10.00 / $50.00 OpenAI charges directly.

Cache write on Astra is $12.50. The model reports a 1,050,000-token context and a 128K completion ceiling, on /v1/chat/completions and /v1/responses, with gpt-6-astra and gpt-6-astra-2026-09-03 as aliases.

curl -X POST https://api.ofox.ai/v1/chat/completions \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-6-astra",
    "messages": [{"role": "user", "content": "..."}],
    "reasoning": {"effort": "high"}
  }'

Read the cache row before you commit. Fable 5.1 matches Astra’s per-token price exactly while reading cache at a quarter of Astra’s rate, which compounds on any agent loop that replays a long prefix. GPT-5.6 Sol remains at half the input price. GET https://api.ofox.ai/v1/models stays the authority on what is callable.

Who Should Pay $50 per Million

• Computer use and terminal automation, where OSWorld 2.0 at 72.6% and Terminal-Bench 4.0 at 57.9% are the specific claims.
• Security tooling, where ExploitBench at 100% and SRE-Bench at 88% on one attempt have no direct third-party equivalent.
• Latency-critical interactive products that can justify Fast mode at $20 / $100.

And who should not: anything that runs at volume on general reasoning, where the same index says $0.55 per task buys the same score. The cost difference at scale is not marginal, and the tier-by-tier breakdown of OpenAI’s previous generation is still the cheaper starting point for most production workloads.

Sources

• https://artificialanalysis.ai/models/gpt-6-astra
• https://artificialanalysis.ai/models/gpt-6-astra-high
• https://llm-stats.com/models/gpt-6-astra
• https://ofox.ai/models/openai/gpt-6-astra
• https://ofox.ai/models/openai/gpt-5.6-sol
• https://ofox.ai/models/anthropic/claude-fable-5.1

GPT-6 Astra benchmarks are OpenAI’s launch figures from 3 September 2026. Index scores and per-task costs were read from Artificial Analysis on 4 September 2026. Ofox rates for GPT-6 Astra and the comparison models were read from the live /v1/models endpoint on 5 September 2026.