OpenAI begins staged GPT-6 Astra rollout at $10/$50 per million tokens

• Astra is rolling out first to selected organizations, then to paid ChatGPT tiers, the API, and AWS, according to OpenAI's developer rollout.
• The $10-per-million input and $50-per-million output list price is 2.5 times GPT-5.6 Sol's $4/$20 rate, while Artificial Analysis's benchmark report found sharply different economics for coding and general-work tasks.
• First-party results moved most on tool-using work, but ARC Prize's results put the much-cited ARC-AGI-3 result at 62.7% under its provider-neutral harness and 99.9% under OpenAI's provider adapter.
• Codex gains a notes-and-retrieval path for long sessions, while a system-card monitorability excerpt says Astra can make its written reasoning harder to monitor in adversarial tests.
• OpenAI calls Astra its first Critical-cyber model, and OpenAI's safety announcement says the most advanced cyber workflows remain more tightly gated.

ARC Prize's analysis says the provider adapter preserves opaque reasoning state between requests and compacts longer conversations. OpenAI's launch page also slips in a new Codex long-context mechanism: persistent notes plus search over previous messages and tool output. Artificial Analysis found the model uses one-third as many tokens as Sol on its coding-agent index, but costs 75% more per intelligence-index task at max effort.

• Access: selected organizations receive Astra first; Plus, Pro, Business, Enterprise, API, and AWS availability follows over coming days, per OpenAI's developer rollout.
• Model surfaces: the API slug is gpt-6-astra, Amazon Bedrock is listed as an API surface, and the availability screenshot lists Astra Pro for Pro, Business, and Enterprise plans.
• Context and effort: the model specification screenshot lists a 1,050,000-token context window, 128,000 maximum output tokens, and low, medium, high, xhigh, and max reasoning effort.
• Price: standard API use costs $10/M input and $50/M output, while an Astra launch recap says Fast mode reaches up to 2.5 times standard speed at twice the price.
• Plan allocation: the usage-allocation update says Astra draws from normal plan allowances, with subscribers able to use their full allowance on Astra.

First-party

• Terminal-Bench Science 0.1: 22.4% → 64.6%, +42.2 points, in OpenAI's launch comparison.
• Terminal-Bench 4.0: 37.3% → 57.9%, +20.6 points, in OpenAI's launch comparison.
• OSWorld 2.0 offline: 65.7% → 72.6%, +6.9 points, according to OpenAI Devs' OSWorld result.
• SRE-Bench pass@4: 68.7% → 99.2%, +30.5 points, in OpenAI's cybersecurity results.
• ExploitBench: 78.5% → 100.0%, +21.5 points, in OpenAI's cybersecurity results.
• ARC-AGI-3, Standard harness: 7.8% → 62.7%, +54.9 points, per a harness-score comparison, while ARC Prize's results identify 99.9% as the Provider Adapter outcome.

Third-party evaluators

• Artificial Analysis Coding Agent Index: 65 → 67, +2 points, in Artificial Analysis's benchmark report.
• Epoch Capabilities Index: 163 → 169, +6 points, per Epoch AI's pre-release evaluation.
• Perplexity WANDR: 0.601 → 0.682, +0.081 points, according to Perplexity's WANDR result.

Customer-reported

• Box complex-work full dataset: 74% → 77%, +3 points, in Box's early-preview evaluation.
• Box legal subset: 69% → 93%, +24 points, in Box's early-preview evaluation.
• Box media and entertainment subset: 48% → 100%, +52 points, in Box's early-preview evaluation.

Artificial Analysis put Astra at 61 on its Intelligence Index, level with GPT-5.6 Sol and five points behind Fable 5.1's 66, in its post-launch benchmark. The same report records about an 80-Elo drop on GDPval-AA v2 and two-to-three-point declines on τ³-Banking, SciCode, and AA-LCR, even as AA-Briefcase improved by roughly 80 Elo.

The benchmarker also found a presentation-quality decline in AA-Briefcase. Its accounting attributes the general-work cost result to the 2.5-times list-price increase overpowering Astra's roughly 10% output-token reduction.

OpenAI's safety overview says Astra's written reasoning is harder to monitor than Sol's when tests explicitly ask it to evade monitoring. The system-card monitorability excerpt reports less than 11% chain-of-thought-only monitor recall on WMDP and zero on Monorepo-Bench under explicit evasion instructions; OpenAI says it has not seen comparable concealment on complex tasks.

• Context window: the model specification screenshot lists 1.05M tokens of input context and 128K maximum output tokens.
• Codex memory: OpenAI's release page says an experimental config.toml setting lets Astra keep notes across windows and search earlier messages and tool outputs, rather than repeatedly compressing all prior work into a summary.
• Question handling: the Codex async-question screenshot says Astra can ask about a decision without halting independent work.
• ARC harnesses: ARC's Standard harness carries only notes the model chooses to preserve; its Provider Adapter retains opaque reasoning state between requests and uses compaction, according to ARC Prize's methodology.

• Dan Shipper's first-run review calls computer use capable of hours in complicated applications and praises 3D work, while reporting a tendency to overcomplicate interfaces at higher effort and weaker prompt intuition than Fable on ambitious builds.
• A detailed workflow review reports clean code migrations and strong computer use, but says Astra remains uneven on hard math and can miss the destination when the operator cannot point it in the right direction.
• steipete's account says Astra debugged Objective-C and then found and patched upstream dependency issues in vitest, tsx, and SwiftPM pull requests.

A long-horizon workflow described in a manager-loop write-up splits the work into four stages:

1. A manager agent creates a checklist and phases.
2. It spawns an implementer in a separate Codex thread.
3. The manager assigns each phase through /goal.
4. The implementer reports completion, then receives the next phase.

The write-up says this structure countered an Astra tendency to asymptote into minutiae on very large tasks.

• OpenAI: OpenAI's release page names ChatGPT Plus, Pro, Business, Enterprise, the API, and AWS as rollout surfaces, after the selected-organization phase.
• Devin: Cognition's Devin announcement says Astra is coming to Devin, and reports a FrontierCode 1.1 result within 0.4 points of Fable 5 at 64% lower cost.
• Factory Droid: Factory's Droid announcement says the model will reach Factory users through its Desktop app or CLI.
• Amp: Amp's plugin instructions show an Astra agent mode usable immediately with a Daybreak-approved OpenAI API key, with a built-in rollout planned for all Amp users.

OpenAI defines Critical cyber capability as the ability, with the right tools and access, to find unknown flaws and develop new exploits across well-protected systems without step-by-step human guidance in its safety overview. The broadly shipped model refuses advanced requests such as proof-of-concept exploit creation; OpenAI says Daybreak will later enable less restrictive defensive workflows including proof-of-concept validation, malware analysis, and detection engineering.

The internal June-to-August V8 test moved from 11.5% arbitrary-code-execution success for Sol to 39.0% for Astra, +27.5 points, while the internal cyber benchmark chart notes that Astra ran with Daybreak Blue access rather than the default production configuration. OpenAI says it found and used two previously unknown zero-days during that evaluation.

the production-monitoring architecture excerpt describes activation classifiers at every sampled token, escalating suspicious trajectories to higher-compute investigators. The release page says a potential critical-boundary violation can pause a ChatGPT or Codex task for review, while an API task stops.