@openai/gpt-6-astra is live on Brainbase

GPT-6 Astra is now available on Brainbase as @openai/gpt-6-astra. Use it with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model.

OpenAI released GPT-6 Astra on September 3 as the first model in the GPT-6 generation. It is a text-and-image model with a context window of just over a million tokens, a 128K output limit, effort levels from low to max, and a knowledge cutoff of April 30, 2026. It is also the first OpenAI model rated Critical for cybersecurity under the company's Preparedness Framework, which shapes how it is being rolled out.

#Why GPT-6 Astra matters for agents

Astra is built around one idea: finish long tasks. OpenAI reports it completes a representative multi-hour task in about 40 minutes where GPT-5.6 Sol took 75, and that it does so with a fraction of the tokens — Artificial Analysis measured roughly a quarter of the output tokens of the median frontier model, and about a third of Sol's on coding work at max effort. Fewer tokens per step and fewer steps per task compound over a long trajectory.

The second change is how it operates software. Astra is trained to work through screens and controls rather than requiring an API integration: filling forms, updating a CRM, editing a spreadsheet, driving engineering tools. That is the capability behind its OSWorld 2.0 result, and it matters most for the long tail of internal systems that were never going to get a clean API.

#GPT-6 Astra benchmarks

On OpenAI's published set, Astra takes the lead on the three benchmarks that measure sustained autonomy — Terminal-Bench 4.0 over Claude Fable 5.1, DeepSWE v1.1 and OSWorld 2.0 over Claude Opus 5 — and posts step changes over GPT-5.6 Sol on FrontierMath Tier 4, ARC-AGI-3, and long-context recall in the 512K–1M band. It trails Fable 5.1 on Humanity's Last Exam with tools. Select a metric below to compare the reported results.

OpenAI chose a different comparison model for each row, so read the chart as a set of claims rather than a league table. Independent testing from Artificial Analysis fills in the table: Astra scores 61 on the Intelligence Index at max effort — level with GPT-5.6 Sol, behind Muse Spark 1.3 at 62 and Claude Fable 5.1 at 66 — and 67 on the Coding Agent Index, matching Opus 5 and Fable 5 and three points behind Fable 5.1. The benchmark-level picture is mixed: an 80-point gain on AA-Briefcase and six points on Humanity's Last Exam, against an 80-Elo regression on GDPval-AA v2 and small declines on τ³-Banking, SciCode, and long-context reasoning.

#GPT-6 Astra pricing and context

Astra is priced at 2.5× GPT-5.6 Sol on both input and output, which puts it level with Claude Fable 5.1 per token. Cached input is discounted 90%, batch runs at half price, and a Fast mode doubles the standard rate.

These are model-token prices before sandbox runtime, tool calls, and other infrastructure usage. Note that Astra's cache read is four times Fable 5.1's $0.25, which matters for agents with long, stable prefixes.

#The 2.5× question

The obvious objection is that Astra scores about the same as Sol on general intelligence and costs two and a half times as much per token. Artificial Analysis measured it at 75% more per Intelligence Index task than Sol at max effort, and $0.96 per task overall — the token efficiency offsets part of the price increase, not all of it. On the intelligence-versus-cost frontier, Astra sits behind its own predecessor.

The counter-case is coding and long runs. On the Coding Agent Index, Astra reaches Fable 5's score at less than half the cost, because it uses a third of the tokens. And on a multi-hour agentic task, a model that fails fewer intermediate steps is cheaper than one that has to retry them, regardless of the per-token rate. That is the workload Astra is priced for. For classification, extraction, and anything schema-validated, it is not the right tool, and Gemini 3.8 Flash or Muse Spark 1.3 will do the job for a fraction of the cost.

#Rollout and the Critical rating

Astra's Critical cybersecurity rating reflects results like a perfect score on ExploitBench, and it gates the rollout. Access opened first to enterprise customers in OpenAI's Daybreak programme, with paid ChatGPT plans and the API following; in organisations that have it, the model is off by default until an administrator enables it. OpenAI also notes that Astra's written reasoning is harder to monitor than Sol's because it takes fewer visible steps, which is worth knowing if your governance process reads traces.

On Brainbase, Astra is available today through the API, and your existing trace and observability tooling applies to it unchanged.

#Run GPT-6 Astra on Brainbase

Select the model in the agent spec and leave the rest of your stack unchanged. This example runs Astra inside the Codex harness on a managed sandbox:

curl https://api.brainbaselabs.com/v2/threads \
  -H "Authorization: Bearer $BRAINBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": {
      "harness": "codex",
      "model": "@openai/gpt-6-astra"
    },
    "input": "Upgrade the service to the new SDK major version, fix every breaking change, and get CI green."
  }'

#Where GPT-6 Astra fits

Astra is the model to reach for when the task is long, the environment is a real machine rather than an API, and the cost of a failed run is higher than the cost of the tokens: multi-hour migrations, computer-use workflows over internal tools, and research that has to hold a million tokens of context without losing the thread. It is not a general upgrade from GPT-5.6 Sol, and it is not competitive with Fable 5.1 on knowledge work.

Route your longest-running agents to it first and compare cost per completed task, not cost per token, against your production traces. Brainbase keeps the harness, sandbox, event stream, and observability layer constant so the model is the only variable.