DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV
All guides
ComparisonLevel · Build#GPT-6#OpenAI#Cost

GPT-6 Astra, Sol and Luna: three tiers, not one default answer.

The current API IDs, 1.05M context windows and price conditions—plus the migration constraints that a model leaderboard will not tell you.

DCDariusz Czechowicz23 Sept 20263 min readSource briefing
GPT-6 Astra, Sol and Luna: three tiers, not one default answer.
Source-checked Sources & corrections
2 practical sections · checklist · sources

Source briefing. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.

For: Developers and teams putting AI into practical use

⚡ TL;DR — the 30-second version

  • Astra: $10/$50; Sol: $2/$10; Luna: $0.10/$0.50 per million tokens at standard short-prompt rates.
  • 1.05M context does not imply one flat price for long prompts.
  • Read endpoint-specific tool and reasoning restrictions before migrating.
  • Missing published scores are unknown—not zero and not estimates.

The GPT-6 family now includes Astra, Sol and Luna. OpenAI's API changelog dates Astra to 3 September and Sol/Luna to 22 September. These are different models; the ChatGPT user interface and the API also have their own availability and workspace controls.

Astra is positioned for the hardest end-to-end work: reasoning, coding, computer use, research and artifacts. Sol is positioned for complex coding and agentic workflows. Luna is positioned for focused, high-volume work. These are provider descriptions to test against your workload, not a DARC ranking.

“Start with the cheapest tier that clears your evals. Escalate with evidence, not with the name on the model card.”
— Editorial interpretation

The standard input/output rates are $10/$50 for Astra, $2/$10 for Sol and $0.10/$0.50 for Luna per million tokens. Each current model page lists a 1,050,000-token total context window and up to 128,000 output tokens. Prompts above 272K input tokens incur 2× input/cache rates and 1.5× output rates for the full request. Tool use, cache writes, regional processing and faster service tiers can add cost.

Astra does not support no-reasoning mode or custom temperature/top_p settings. Tool-enabled Astra integrations use Responses rather than Chat Completions. The Sol and Luna pages say that function calling through Chat Completions requires reasoning_effort=none; use Responses for the fuller tool workflow.

A sensible rollout compares three things on the same inputs: accepted-task rate, total cost per accepted task and review time. Do not transfer a score from GPT-5.6 Sol to GPT-6 Sol simply because the names resemble each other. If no comparable verified score is in the reviewed sources, this site's comparison shows an em dash.

Field guide / 01

Route by demonstrated task shape

A bounded classification or extraction task is a reasonable place to test a low-cost tier because you can validate its output directly. Complex tool-mediated work needs separate evaluation of recovery, context handling and side effects; fluent text is not a proxy for successful execution.

Keep routing thresholds outside the model where you can. Escalate on a failed deterministic check, missing evidence or an explicitly higher-risk class—not an uncalibrated statement that the model feels 95% confident.

WorkloadFirst experimentEscalation signal
ExtractionLuna with strict schema and evidenceMissing or inconsistent fields
Repository taskSol with fixed harness and testsRepeated failure or incomplete verification
Complex researchAstra on a bounded source taskUnsupported claims or unresolved contradictions

Field guide / 02

Test around the billing boundary

Use fixtures just below and above the 272K input threshold when your product routinely sends large contexts. Record the effective rates for the full request; an average prompt length can hide a costly tail.

The same caution applies to cache writes and fast processing. A comparison that uses one provider's cached price and another's uncached price is a workload assumption, not an objective ranking. Make that assumption visible.

Before you ship

0/5

Use this as a review aid. Checks reset when you leave this page.

Questions that come up

Is Luna cheaper for every completed task?

Its published token price is lower, but retries, output length, tools and human review determine completed-task cost. Measure both quality and cost on the same inputs.

Put it into practiceBuild a provider-routing budget

Sources & verification

Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.

  1. 01OpenAI API changelogOpenAI↗
  2. 02GPT-6 Astra API model and pricingOpenAI↗
  3. 03GPT-6 Sol API model and pricingOpenAI↗
  4. 04GPT-6 Luna API model and pricingOpenAI↗
  5. 05OpenAI — model migration and cache changesOpenAI↗
  6. 06OpenAI — current prompt-caching behaviorOpenAI↗
Editorial policy and correction log →
#GPT-6#OpenAI#Cost#Agents