OpenAI says GPT-5.6 Sol helped post-train GPT-5.6 Luna

• OpenAI framed GPT-5.6 as a three-model family, but the line that set off the timeline was Sol helping post-train Luna, as captured in the livestream clip and repeated in the Sol and Luna reaction.
• The strongest artifact was a redacted prompt that asked the model to wire training configs, choose a GPU count, launch a run, and let it step, shown in the prompt screenshot and highlighted by the 100k GPU post.
• The useful reading is training-ops autonomy inside OpenAI’s existing stack: nrehiew_ described config edits and experiment launches, while scaling01's breakdown argued humans still supplied the research idea and infrastructure.
• The same release moved into product surfaces fast, with GPT-5.6 variants appearing in Codex diffs in the Codex snapshot and ChatGPT Work landing in the release-note screenshots.

The redacted prompt included branch checkout, cherry-picking, training config changes, a launch script entrypoint, and a warning not to unblock with unsafe operations. The Simon Willison roundup put all three models at a February 16, 2026 knowledge cutoff, a 1M token context window, and 128K max output, while the ParseBench run found no document-understanding gain over GPT-5.5. The release-note capture bundled Work, Sites, Plugin Directory, a new desktop app, group chat retirement, and Atlas shutdown into the same July 9 drop.

OpenAI announced GPT-5.6 Sol, Terra, and Luna for public launch on Thursday, with preview access expanding globally. sama's post reduced the launch to three words: “GPT-5.6 sol launches thursday.”

The model ladder in the launch roundup was simple:

• Luna: $1 input, $6 output per 1M tokens.
• Terra: $2.50 input, $15 output per 1M tokens.
• Sol: $5 input, $30 output per 1M tokens.
• Shared specs: February 16, 2026 knowledge cutoff, 1M token context, 128K max output.

That pricing made the Sol story sharper because the claim was not just “better coding model.” It was a frontier model cheap enough to run inside long agent loops and, according to the launch chatter, inside OpenAI’s own training workflow.

The prompt artifact looked like a training-ops handoff. Its visible instructions asked the model to:

1. Check whether the local branch had usable training configs.
2. Wire those configs into the target project.
3. Make a Python entrypoint support the target training mode.
4. Choose a GPU count under a stated cap.
5. Modify scripts/launch_train.py.
6. Use OpenAI compute to launch the run and confirm it stepped.
7. Create a branch, then cherry-pick missing master changes if needed.
8. Use judgment and avoid unsafe unblocking.

eliebakouch's post focused on the screenshot’s “~100000 GPUs” reproduction, which turned the demo into instant recursive-self-improvement bait. A follow-up from eliebakouch kept the narrower version alive: handling failure, adjusting GPU config, and making choices was still cool, even without doing the science.

The technical dispute settled around where the autonomy lived. nrehiew_ read the demo as Sol going from high-level ideas to editing configs and launching experiments, rather than owning Luna’s end-to-end post-training.

In scaling01's longer breakdown, the likely loop was:

• A researcher has an idea, such as improving sycophancy, honesty, or intent recognition.
• The model implements a grader, a multi-agent game, or a reward-model workflow.
• OpenAI’s existing RL infrastructure executes the run.
• The model babysits failures, analyzes results, and suggests reward-shaping changes.

aidan_mclau's reply said it was routine for 5.6 to do an entire RL run end to end, and a follow-up reply answered “a lot” when asked how often. scaling01's caveat drew the boundary: the claims showed useful internal acceleration, while literal autonomous end-to-end research remained out of reach.

The important phrase after the Luna line was “automated researcher.” deredleritt3r quoted the follow-up: work that previously took a team of senior OpenAI researchers now made the automated researcher feel “pretty close.”

The experiment-throughput chart claimed OpenAI had doubled experiments per researcher since the start of 2026. That is the concrete metric behind the hype: more launched experiments per researcher, not a standalone AI lab in a box.

scaling01's later reply called the initial claim overhyped after hearing the livestream statement in context. The useful part is boring and technical: a model that can operate the research harness changes throughput before it changes who invents the research program.

The launch tables showed strong agentic-coding gains, plus one obvious Fable-shaped hole:

• Terminal-Bench 2.1: GPT-5.6 Sol Ultra scored 91.9% versus GPT-5.5 at 85.6%, +6.3 points, in the benchmark table.
• Terminal-Bench 2.1: GPT-5.6 Sol scored 88.8% versus GPT-5.5 at 85.6%, +3.2 points, in the same table.
• Agents’ Last Exam: GPT-5.6 Sol scored 52.7% versus GPT-5.5 at 46.9%, +5.8 points, and versus Claude Fable 5 at 40.5%, +12.2 points.
• SWE-Bench Pro: GPT-5.6 Sol scored 64.6% versus GPT-5.5 at 59.4%, +5.2 points, while Claude Mythos 5 scored 80.3%, +15.7 points over Sol.
• KernelGen 1P: the kernel chart said Sol improved over GPT-5.5 by +31.8 points on first-party Jalapeño chip kernel tasks.

Document understanding did not move with the same pattern. The ParseBench run reported “no change” between GPT-5.6 Sol and GPT-5.5 across tables, text, charts, layout, and bounding boxes, with an average score of 62.1 for Sol versus 64.4 for GPT-5.5, -2.3 points. The linked ParseBench leaderboard framed that as a 70-plus-model comparison rather than a single anecdote.

The product rollout put the same model family behind broader work agents. The release-note screenshots captured these July 9 changes:

• ChatGPT Work: an agent for longer tasks across files, plugins, connected apps, documents, spreadsheets, presentations, reports, and Sites.
• Scheduled Tasks: once, recurring, trigger-based, or change-monitoring runs.
• Enterprise and Edu: a two-week preview where Work starts off by default, then turns on automatically unless admins opt out.
• Desktop app: Chat, Work, and Codex in one macOS and Windows app, with local files and desktop apps available to Work by permission.
• Codex desktop: inline diffs, pull-request review in the side panel, faster Computer Use, and multi-repository projects.
• Sites: public URL publishing for Business and Enterprise, with public publishing unavailable in the EEA, Switzerland, and the UK at launch.
• Plugin Directory: a replacement for the App Directory, packaging skills, apps, and app templates across ChatGPT and Codex.
• Retirements: group chats stopped accepting new creation or joins on July 9, and Atlas was scheduled to stop working on August 9.

The Codex snapshot showed gpt-5.6-sol becoming the new default, with Terra and Luna added as balanced and fast agentic coding models. The VS Code rollout said Sol, Terra, and Luna were already rolling out in VS Code.