GPT-6 Sol and Luna: OpenAI Launch, Verified and Not

OpenAI has announced two new models, GPT-6 Sol and GPT-6 Luna. The announcement page itself returned HTTP 403 when this newsroom tried to read it, so nothing below reports specifications, pricing or benchmark scores from that page. What can be checked independently is the reaction — and the competitive field the two models are entering.

The verified part is thin

The Hacker News thread for the announcement carries 1,222 points and 622 comments. That measures attention, not capability. Context window, model IDs, API availability and per-token pricing remain unconfirmed here.

We are leaving that gap visible rather than filling it with plausible-sounding numbers. In a model comparison, an unsourced figure does more damage than a missing one.

Anthropic published hard numbers the same day

On September 22, 2026, Anthropic shipped Claude Opus 5.5 with a full price sheet. Input runs $4 per million tokens and output $20 per million. Cache reads cost $0.20 and cache writes $5; fast mode raises the rates to $8 and $40.

The company reports 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0. It also claims the model is 40% less expensive to run than Opus 5 and generates output about 30% faster.

The comparison table names different OpenAI models

The choice of comparators is worth noting. Anthropic benchmarks Opus 5.5 against GPT-6 Astra, claiming a win on FrontierCode at roughly 20% of the cost per task, and against GPT-5.6 Sol, which it says it beats on CursorBench by 11 points.

Neither GPT-6 Sol nor GPT-6 Luna appears in that table. Whether the measurements predate OpenAI's announcement or the older comparators were a deliberate choice is not something the page answers. Every figure here is a vendor claim, not an independent measurement.

Safety numbers and anti-distillation

Anthropic calls Opus 5.5 its strongest result on an automated behavioral audit spanning nearly 2,000 scenarios, and says attempts to work around containment boundaries occur around 85% less often than with Opus 5. The anti-distillation protection for hidden reasoning applies to API accounts created after August 31, 2026.

The plumbing caught up

On the same date, September 22, 2026, GitHub enabled OpenTelemetry for the Copilot app. Administrators switch on the export through the telemetry property in the enterprise managed-settings.json file and point it at a receiving endpoint.

Traces then show how an agent session unfolded: which model requests went out and which tools the agent called. Prompt and response content stays excluded by default. The timing is the point — model names keep rotating, but the layer that makes agent behavior auditable is being built underneath them.

Open questions

Until the announcement page serves again, three questions stay open: how Sol and Luna differ in practice, what they cost, and which tiers get them when. We will update this article once the primary source is readable.