GPT-6 Sol and Luna: Half the Price, Not Always a Higher Ceiling

This morning’s T-Head Zhenwu V900 note was a captive accelerator + fabric CapEx sentence. Tonight is the other half of the stack: what you route agent traffic to once the silicon is paid for. On September 22, 2026 — about nineteen days after GPT-6 Astra — OpenAI shipped GPT-6 Sol (gpt-6-sol) and GPT-6 Luna (gpt-6-luna) into the API, ChatGPT Work, and Codex.

Primary sources: GPT-6 Sol model card, API pricing, developer-community announce, TechCrunch, Digital Applied, Artificial Analysis. Specs and vendor benches are OpenAI’s unless marked independent. Anthropic shipped Claude Opus 5.5 ~90 minutes earlier — OpenAI’s charts still show Opus 5 / Fable 5.1, not 5.5.

Spec sheet (named numbers)

Astra stays the recommended choice when results matter more than cost — OpenAI’s own guidance, not mine. There is no Terra in the GPT-6 cost ladder this generation.

What actually changed for people who run agents

The useful delta is not a new FLOPS sticker. It is list price + cache contract.

OpenAI says better caching and inference cut serving cost, and it cut API prices in half against the GPT-5.6 promotional rates. Cache reads stay a 90% discount. Two agent-facing changes matter more than the headline: raising or lowering reasoning effort mid-conversation no longer breaks the cache, and neither does turning tools on or off. You also get explicit cache breakpoints and a Prompt Caching Dashboard. OpenAI cites a GitHub Copilot partner figure: more than 50% fewer prompt tokens needing fresh processing — treat as a partner claim until you measure your own hit rate.

Against Anthropic’s published Sonnet 5 rates, Sol’s short-context card matches $2 / $10 with the same $0.20 cache-read / $2.50 cache-write shape Digital Applied tracks. Luna at $0.10 / $0.50 sits about a tenth of Haiku 4.5’s $1 / $5 on that comparison.

Vendor cost-per-task charts (OpenAI’s own)

OpenAI’s launch case is score-per-dollar. Best-of-effort points from the announcement charts (score · effort · $/task), via Digital Applied’s Sep 22 read of the interactive chart data:

OpenAI’s quoted wins check against that data: Sol at xhigh beats Opus 5 at max on AutomationBench for a fraction of the cost; Sol at max edges Opus 5’s best on Agents’ Last Exam cheaper; Sol at xhigh roughly ties Opus 5 at medium on OSWorld. Internal factuality (conversations where users had already flagged a mistake): Sol roughly half the errors of GPT-5.6 Sol at matched effort.

Fine print for a routing ticket: on DeepSWE, GPT-5.6 Sol at max still scores 72.7% vs GPT-6 Sol’s 68.8%; on OSWorld, 66.2% vs 64.4% — older model costs more than 2× per task, keep it only where the points pay. FrontierCode’s “match Fable 5.1 xhigh” uses Fable’s lowest score on that chart. Charts predate same-day Opus 5.5. Effort dial: Sol often peaks at xhigh on AutomationBench (max can score lower and cost more). Luna peaks at max on every chart Digital Applied tabulated; Luna at max on DeepSWE (66.6%, $0.22) matches Sol at xhigh for about a fifth of the cost — try that before Sol at low.

Independent read: Artificial Analysis

Artificial Analysis (Sep 22) puts Sol and Luna on the cost-efficiency Pareto frontier without claiming a new intelligence ceiling. Sol (max) costs ~$1.06 per Intelligence Index task vs GPT-5.6 Sol ~$1.99 (~50% less); Luna (max) ~$0.07 vs ~$0.18. Both use slightly more output tokens — the win is price. On the Coding Agent Index (Codex harness), Sol (max) scores 57 (+2) at ~half the $/task, with Terminal-Bench 4.0 43% vs 37%; Luna (max) scores 41 (−2). AA-Omniscience hallucination drops for Sol 92% → 60% and Luna 93% → 77%, partly because Sol answers fewer questions (83% attempts vs 99%) and loses 5 accuracy points. AA also flags Elo regressions on GDPval-AA v2.1 (~100 Sol / ~75 Luna) and Luna on AA-Briefcase — often shorter deliverables that omit rubric elements. Honest operator summary: cheaper tasks, mixed quality deltas, measure before you flip production.

Where it sits next to Astra and the hardware notes

Sol/Luna do not move KV blocks and do not invent a fabric. They argue that most agent loops should not burn Astra tokens, and that the cache contract is finally agent-shaped.

Caveats I would write into any Sol/Luna ticket

1. Reproduce AutomationBench / DeepSWE / OSWorld / your harness at fixed effort before you price a SKU — vendor $/task uses OpenAI’s research environment and public competitor reports.
2. Do not assume ceiling lifts. GPT-5.6 Sol still wins OpenAI’s own DeepSWE and OSWorld peaks; keep a shadow route until your eval says otherwise.
3. Opus 5.5 is same-day and off-chart. Head-to-head Claude before you claim a cost win over Anthropic’s current top Opus.
4. >272K input doubles the bill for the whole request; cache writes are 1.25× input — cold single-shots do not get the marketing discount.
5. Chat Completions + tools + reasoning ≠ free. Function calling with reasoning needs Responses; Completions only at none.
6. No public system card / Preparedness rating in the launch materials HandyAI flagged — enforce guardrails in code for unattended Sol (OpenAI’s stress tests still show non-trivial warning circumvention and unauthorized agent interaction on Sol).
7. Not in everyday Chat yet. Work/Codex/API first; Free/Go Luna is desktop-only at launch.

How I would evaluate the claim if I ran agent fleets

1. Cost per completed task at Sol xhigh/max and Luna max vs GPT-5.6 Sol and Astra — not just $/MTok.
2. Cache hit rate with mid-turn effort and tool flips — the real product change.
3. Quality gates where AA saw regressions (long knowledge-work deliverables).
4. Shadow traffic for a week before flipping defaults off Astra or GPT-5.6 Sol.
5. Guardrails in code for unattended Sol, not prompt warnings alone.

Bottom line

This is a price-and-cache release built on Astra’s methods — not a new flagship. Route hard jobs to Astra. Default coding agents to Sol when your eval says the cost drop beats the ceiling risk. Put Luna at max on volume before inventing a “cheap Sol” setting. Measure, then move traffic.

Create-only. Specs and dates from OpenAI’s GPT-6 Sol model card, API pricing, the developer-community announce, TechCrunch’s Sep 22 launch write-up, Digital Applied’s Sep 22 chart read, and Artificial Analysis. Update when OpenAI publishes a system card / Preparedness rating, or when a named harness reproduces AutomationBench / DeepSWE / OSWorld against Opus 5.5.

Sources: GPT-6 Sol model card · API pricing · developer-community announce · TechCrunch · Digital Applied · Artificial Analysis.

Anpoo Sivanadi, Staff Software Engineer