@openai/gpt-6-sol and @openai/gpt-6-luna are live on Brainbase

GPT-6 Sol and GPT-6 Luna are now available on Brainbase as @openai/gpt-6-sol and @openai/gpt-6-luna. Use them with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model.

OpenAI released both on September 22, three weeks after GPT-6 Astra, as the mid-tier and low-cost members of the GPT-6 family. Both share Astra's 1.05M-token context window and 128K output limit, take text and images, and offer effort levels from none to max, with medium the default. Sol's knowledge cutoff is April 20, 2026; Luna's is May 18. There is no GPT-6 Terra.

#Two models, two jobs

Sol is the one most agents will run on. OpenAI positions it for the work developers and knowledge workers repeat all day: building features, reviewing code, debugging, analysing data, and multi-step agents that carry context through tool calls. Luna is for work with a clear shape and a lot of repetitions: routing, classification, extraction, summarisation, short answers. At its price it is meant to run thousands of times a day on outputs nobody reads.

The release is a cost argument first. OpenAI charted every benchmark as score against cost per task at each effort level, and the claim throughout is not a new peak but the same score for a fraction of the money. The other thread is reliability: on OpenAI's internal factuality evaluation, built from real conversations where users flagged mistakes, Sol makes about half as many errors as GPT-5.6 Sol, and on its coding-deception test Sol's rate fell from 10.4% to 1.3%.

#Sol and Luna benchmarks

On OpenAI's published set, Sol at max effort sits within 1.1 points of Claude Fable 5 on DeepSWE at about 80% lower cost per task, beats Claude Opus 5 on Agents' Last Exam and AutomationBench, and matches Opus 5 on OSWorld. Luna at max effort lands within a few points of Sol on DeepSWE and Agents' Last Exam for a tenth of the cost. Note that Sol's best scores on DeepSWE and OSWorld are below GPT-5.6 Sol's; the gain there is price, not peak. Select a metric below to compare the reported results.

All scores are each model's best across effort levels as charted by OpenAI; Sol and Luna at max, except AutomationBench at xhigh for Sol. Independent testing from Artificial Analysis tells the same story with less flattering framing: Sol scores 48 on the Intelligence Index against 47 for GPT-5.6 Sol and 57 against 55 on the Coding Agent Index, at $1.06 per index task instead of $1.99. Luna holds at 37 on intelligence, slips two points on coding, and costs $0.07 per task instead of $0.18. Sol's hallucination rate on AA-Omniscience fell from 92% to 60%, partly because it now declines more questions. On GDPval-AA, Sol scores 1487, below GPT-5.6 Sol's 1588, so it is not an upgrade for long-form professional work.

#Sol and Luna pricing and context

Both rate cards are half of their GPT-5.6 predecessors, and OpenAI has confirmed the prices are permanent rather than introductory. Cached input is 90% off, and GPT-6 caching survives changes to effort and tool availability, which matters for agents that adjust settings mid-task. Prompts over 272K input tokens are billed at 2× input and 1.5× output for the whole request.

These are model-token prices before sandbox runtime, tool calls, and other infrastructure usage. Two cautions from OpenAI's own charts: Luna at low effort is not usable for agentic work (2.4% on DeepSWE), so its real floor is medium; and more effort does not always help, since Sol peaks at xhigh on AutomationBench and Luna scores worse at high than at medium on Agents' Last Exam. Test the effort setting on your own workload.

#Run Sol and Luna on Brainbase

Select the model in the agent spec and leave the rest of your stack unchanged. This example runs Sol inside the Codex harness on a managed sandbox; swap the model id for @openai/gpt-6-luna for the high-volume version.

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-sol"
    },
    "input": "Audit this repository, fix the highest-impact reliability issue, and run the tests."
  }'

#Where Sol and Luna fit

Sol is the sensible default for recurring coding and agent work at Sonnet-class pricing: it covers most of the ground of GPT-6 Astra on tool-driven loops for a fifth of the price, and it is half the uncached price of Claude Opus 5.5, released the same morning. Where Astra or Opus 5.5 lead clearly, on computer use and the hardest repository work, the question is whether the gap is worth the multiple for your task.

Luna changes what is worth automating at all. At a tenth of a cent per thousand input tokens, extraction, routing and summarisation that used to be batched can run on every event. The combination that works in routing policies is Luna for the volume, Sol for the judgment, and a frontier model for the few steps that need it. Brainbase keeps the harness, sandbox, event stream and observability layer constant so the model is the only variable.