Launches

GPT-6 Astra Lands at $10/$50 With Benchmark Claims to Check

A two-stage rollout, frontier-tier pricing, and a benchmark chart with at least one number that should make you suspicious.

Priscilla Hayes

Senior AI Correspondent

Published 3 min read
Abstract steel structure with futuristic geometric patterns, creating an intricate industrial design.
Jump to 7 sections

OpenAI shipped GPT-6 Astra on September 3, 2026, opening it first to approved organizations before widening access the following day. It sits at the top of the GPT-6 line, above the Sol and Luna tiers that arrived over the summer, and it is priced accordingly.

The headline number for anyone building on it is $10 per million input tokens and $50 per million output. That is five times what Sol costs on output and a hundred times what Luna costs. Astra is not a drop-in upgrade for a production workload; it is a tier you route to deliberately.

Close-up of a digital market analysis display showing Bitcoin and cryptocurrency price trends.

What OpenAI says is different

The pitch is reasoning depth and agentic reliability rather than raw speed. OpenAI's published figures put Astra at the top of several evaluation suites, including FrontierMath's hardest tier and the ARC-AGI-3 abstraction benchmark, alongside a specialized security-testing variant.

Two of those claimed scores deserve a flag. A near-perfect result on an adversarial benchmark usually means the benchmark has been saturated, contaminated, or scoped narrowly enough that the number stops being informative. It does not automatically mean the model is weak — it means the number is not doing the work the marketing wants it to do.

Model Drop has not independently reproduced any of these figures. We are reporting what the vendor published, which is a different thing from confirming it, and we would treat every launch-day benchmark chart the same way. The methodology caveats in our guide to reading a model launch announcement apply directly here.

What the ARC Prize results actually show

The ARC-AGI-3 figure is the clearest example of a number that depends on how it was produced. ARC Prize published its own results for Astra on the semi-private set under two setups. With its Standard harness, a minimal provider-neutral interface, the best score was 62.7% at max reasoning effort, at a cost of $26,098. With the Provider Adapter harness, the best score was 99.9% at high effort, at a cost of $18,817.

The gap comes from context handling. The Standard harness has the model carry forward only the notes it chooses to keep. The Provider Adapter uses OpenAI's own context-management features, preserving opaque reasoning state between requests and using compaction for longer conversations. ARC Prize reported that, across the 167 game-reasoning pairs both harnesses solved, the adapter runs were about 3.66 times faster and used 49% fewer total tokens.

Both numbers are real measurements, but they answer different questions. One asks how well the model does in a neutral setup. The other asks how well it does when the vendor's own tooling carries its state. A buyer who will not use that tooling should weight the lower figure.

The pricing tier is the real story

Astra's cost structure tells you who it is for. At $50 per million output tokens, a single agent run that emits 40,000 tokens costs $2 before you count the input side. Multiply that across a support queue or a nightly batch job and the arithmetic stops working quickly.

Woman working on cybersecurity programming with laptops and multiple screens
The pricing tier is the real story
TierInput / 1MOutput / 1MSensible use
GPT-6 Astra$10$50Hard reasoning, low volume
GPT-6 Sol$2$10General production work
GPT-6 Luna$0.10$0.50High-volume, simple tasks

Those figures are OpenAI's list prices as of the launch window; check the OpenAI platform pricing documentation before you budget against them, because tiers and promotional rates move.

The three-tier structure itself is now the industry default. Anthropic runs the same shape — Anthropic's published Claude pricing lists Fable 5.1 at $10/$50, Opus 5.5 at $4/$20, and Sonnet 5 at $2/$10 — and Google splits Pro and Flash the same way. Every serious lab has concluded that one model at one price does not fit the market.

What the pricing page adds

The list price has a second tier that matters for large prompts. OpenAI's current API pricing page sets the long-context boundary at 272K input tokens. Above it, Astra input rises from $10 to $20 per million tokens and output from $50 to $75. Cached input is $1.00 per million, or $2.00 above the boundary.

The eesel pricing guide lists a 1,050,000-token context window and says Batch and Flex run at half the standard rate, which would be $5 input and $25 output. It also reports a Fast mode at twice the standard price. eesel sells a competing support product, so confirm those service-tier figures against OpenAI's page before budgeting.

What this changes if you are shipping

Very little this week, and that is the honest answer. A frontier release at the top price tier is not something you migrate to on launch day. The sensible sequence is to benchmark it on your own evaluation set, price the delta against whatever you run now, and route only the traffic that genuinely needs the extra capability.

A woman using a laptop navigating a contemporary data center with mirrored servers.

The workloads where a jump like this tends to pay for itself are narrow: multi-step agents that fail expensively, code changes against large unfamiliar repositories, and anything where a wrong answer costs more than $2 to clean up. Everything else belongs on a cheaper tier (see Astra vs Fable 5.1).

Independent evaluation will take a few weeks to arrive. Community leaderboards such as LMArena's public model rankings collect head-to-head human preference data over time, and third-party coding benchmarks like SWE-bench will post verified numbers on their own schedule. Those are worth more than any launch-day chart.

What independent scoring says so far

Early third-party numbers are more modest than the launch framing. A MindStudio write-up of the launch, which cites Artificial Analysis, reports an Intelligence Index score of 61 for Astra, tied with GPT-5.6 Sol and behind Claude Fable 5.1 at 66. On the Coding Agent Index it puts Astra at 67 against 70 for Fable 5.1.

The same write-up says Astra uses about a third as many output tokens as Sol on coding agent evaluations, which makes per-task cost roughly equal to Sol there even though the per-token rate is 2.5 times higher. That is a secondary report, so treat it as a lead to verify on your own workload rather than a settled result.

At Model Drop we will revisit Astra once there is independent data to compare against the vendor's own. Until then, the defensible summary is that OpenAI has a new top tier, it costs $10 and $50, and the capability claims are unconfirmed.

What to watch next

Three things. Whether the claimed reasoning gains hold up on tasks that were not in anyone's benchmark suite. Whether OpenAI cuts Astra's price within two quarters, as has happened with every previous top tier. And whether the specialized variants turn into a permanent product line or quietly disappear.

The release also lands inside an unusually dense launch window — four major labs shipped within days of each other, which we cover in our coverage of Anthropic's Fable 5.1 release two days earlier.

By Greg Halston, Staff Writer at Model Drop. Reported September 4, 2026; benchmark claims are the vendor's and have not been independently verified by Model Drop.

Model Drop covers AI launches — new models, platforms, features, and tools — for the people who have to decide what to actually ship on.

How much does GPT-6 Astra cost?
OpenAI listed Astra at $10 per million input tokens and $50 per million output at launch, with cheaper Sol and Luna tiers beneath it. Output dominates most real bills, so a single agent run emitting 40,000 tokens costs about $2 before input is counted. Verify current rates on the platform pricing page.
Are the GPT-6 Astra benchmark scores verified?
Not independently, and Model Drop has not reproduced them. The figures are OpenAI's own, measured with OpenAI's scaffolding. A claimed near-perfect result on an adversarial benchmark usually indicates saturation or narrow scoping rather than a capability leap, so treat launch-day charts as marketing until third parties publish.
Should I switch my production workload to GPT-6 Astra?
Not on launch day. Benchmark it against your own evaluation set, price the cost delta versus what you run now, and route only traffic that genuinely needs the extra capability. At $50 per million output tokens, most production work belongs on a cheaper tier without any quality difference.
What kind of workload justifies frontier-tier pricing?
Narrow ones: multi-step agents that fail expensively, code changes across large unfamiliar repositories, and any task where a wrong answer costs more to clean up than the inference costs to run. If a mistake is cheap to catch and fix, a mid-tier model almost always makes more sense.

Written by

Priscilla Hayes

Senior AI Correspondent

8+ years experience

Priya has covered model releases since the first wave of chatbot launches and has never met a benchmark leaderboard she didn't immediately try to break. Before Model Drop, she ran technical comms for a model-eval startup, which means she reads system cards for fun and gets suspicious of any release that ships without one. She keeps a running spreadsheet of every 'frontier' claim made in a press release and how long it survived contact with actual users.

Credentials

  • 8 years covering AI model releases and infrastructure
  • Former technical communications lead, an independent model-evaluation lab
  • Built and maintains a public tracker of context-window and pricing claims across 40+ model families