GPT Models: GPT-6 Astra, Sol, Terra and Luna

OpenAI's current GPT line is GPT-6: Astra ($10 in, $50 out per million tokens), Sol 6.1 ($2 and $10), Terra 6.1 ($1 and $5) and Luna ($0.10 and $0.50). All four share a 1,050,000-token context window, 128,000 max output, text and image input and text output. They are closed-weight. OpenAI's open weights are the separate gpt-oss models. Checked 31 October 2026.

The lineup at a glance

OpenAI's models page says to start with Astra, "our flagship model for complex reasoning and coding," choose Sol 6.1 "to balance intelligence and cost," or Luna for "cost-sensitive, high-volume workloads." Terra sits between Sol and Luna.

Every row comes from the four model pages on developers.openai.com, fetched 31 October 2026. The models page also still lists GPT-6 Sol ($2 and $10, cached $0.20), the GPT-5.6 line and the whole GPT-5.x range, so the naming has layers. Sol 6.1 is the newer of the two Sol models.

Maker, licence and open weights

OpenAI makes GPT. The GPT-6 family has no published weights and no open licence, so there is nothing to download. You reach it through the OpenAI API, ChatGPT and Codex, and through resellers that carry their own price lists.

OpenAI does ship open weights, but under a different name. The models page lists gpt-oss-120b as the "most powerful open-weight model, fits into an H100 GPU" and gpt-oss-20b as a medium-sized model for low latency. Hugging Face tags both Apache-2.0 with a 131,072-token context. Those are not GPT-6 and are not in the same quality class, so we give them their own page: gpt-oss. For the wider choice of downloadable models, see open source LLMs.

Modalities, context and output

All latest OpenAI models, per the models page, support text and image input, text output, multilingual use and vision. Sol and Terra list audio and video input as unsupported. Image generation, speech and realtime voice are separate models: GPT-Image-2.5, GPT-Realtime-2.1 and GPT-Live 1 are examples on the catalogue.

The 1,050,000-token window includes an input cap of 922,000, which leaves room for 128,000 output tokens. One catch on Sol 6.1: its model page says it uses a default context window of 500K tokens in Codex and ChatGPT Work, and usage drawn from a subscription has no long-context surcharge there.

Tool use needs the Responses API on Sol, Terra and Luna. On Luna and the older Sol, Chat Completions supports function calling only when reasoning effort is none. Astra supports both Chat Completions and Responses, and Batch.

Prices, checked today

Source: each model page and the API pricing page, 31 October 2026, USD per million tokens, standard processing, prompts up to 272K input tokens.

Modifiers that change a bill:

• A prompt over 272K input tokens is priced at 2x on input and cache rates and 1.5x on output, for the entire request. Astra becomes $20 and $75 on such a prompt.
• Cached input is 5 percent of the input rate on Sol and Terra and 10 percent on Luna and Astra.
• A cache write costs 1.25x the input rate at 30-minute retention. Terra's page also lists a 12-hour retention at $2.50.
• Batch and Flex are 50 percent of standard. Fast mode is 2x.
• Regional processing carries a 10 percent premium where available.
• Reasoning tokens are billed as output, so effort is a cost dial.

For the full list including GPT-5.x and Pro models, read OpenAI API pricing. Our cross-vendor view is LLM API prices.

Where it is served

OpenAI's own API is the primary home, and the one our price table uses. We did not find the GPT-6 models on any host in our directory, so we link none. If you want OpenAI-made weights on a host you control, the options are the gpt-oss models, which appear on Hugging Face Inference and AWS Bedrock among others.

Which coding agents use GPT

OpenAI Codex CLI is OpenAI's own agent. It comes with ChatGPT plans or runs on an API key, and it lists GPT-6 Astra, Sol 6.1, Terra 6.1 and Luna with different local message allowances per five hours. Those figures and the credit rates are on Codex pricing.

Third-party agents carry GPT too. GitHub Copilot's docs list GPT models from 5.3-Codex upward. Cursor has model pages for GPT-5.6 Sol, Terra and Luna. Amp names GPT-5.6 in its manual. OpenCode lists OpenAI as a provider.

Published benchmarks

OpenAI's model pages that we fetched carry no benchmark table. The only GPT-6 scores we found are in Anthropic's Opus 5.5 launch table, where the GPT-6 Astra column is labelled "as reported by OpenAI" for some rows and from Zapier's public leaderboard for others. That makes them second-hand, so treat them as vendor-reported and compare cautiously.

Source: Anthropic's Claude Opus 5.5 announcement, vendor-reported by Anthropic with GPT figures attributed to OpenAI or public leaderboards. Setups differ: Anthropic states that Opus 5.5 ran at max or xhigh effort and Astra at high on Terminal-Bench. Because Anthropic made the table, it is the party with most to gain from it.

Local hardware needs

None for GPT-6. For gpt-oss-120b, OpenAI says it fits on a single H100 GPU, and the 20b version is the lower-latency option. See gpt-oss for the model card details.

The Codex branch

OpenAI has also published models branded for its coding agent: GPT-5-Codex, 5.1-Codex, 5.1-Codex-Max, 5.1-Codex Mini, 5.2-Codex and 5.3-Codex. By our check on 31 October 2026, 5.3-Codex is marked deprecated and the others shut down. The GPT-6 line is not split into Codex variants, which is why GPT Codex models is mostly a history page.

Reasoning effort, in plain terms

Every GPT-6 model thinks before it answers, and the reasoning.effort setting controls how long. Astra and Sol 6.1 accept low, medium, high, xhigh and max. Terra 6.1 and Luna add none, which skips reasoning altogether. Medium is the default on Sol, Terra and Luna.

This matters for the bill more than the rate card does. Reasoning tokens are billed as output, so a long think on Astra at $50 per million can cost more than the visible answer. A sensible habit is to start at low or none on a task, run your own checks, and raise effort only where the answers fall short. Because OpenAI prices output at five times input across the line, a prompt that produces a short answer after heavy reasoning can still cost more than one that produces a long answer after light reasoning.

Moving from GPT-5.x

The pricing page keeps the GPT-5.x families next to GPT-6, so a move is a table lookup first and an evaluation second.

Our own percentages, computed from the listed rates. One scheduling note: GPT-5.5 retires from Codex with ChatGPT sign-in on 14 October 2026, per OpenAI, and API-key use is not affected, so anything still pointing at it in a script should be checked. Price is only part of a migration. Prompt sensitivity and tool behaviour change between generations, so rerun your evaluation set before you move traffic.

Data handling and regions

OpenAI keeps a guide titled "how OpenAI uses your data" and a data residency feature. Luna's page says EU data residency is available across Standard, Fast, Flex and Batch processing, and regional processing carries a 10 percent premium where available. Terra's coverage differs: it does not support UAE inference residency. We did not review the retention terms in full.

For regulated work, the practical rule is to read the page for the exact model you call, because residency support differs by model.

Choosing within the family

Cost per request helps. Our arithmetic from the listed rates, for 50,000 input and 2,000 output tokens with no caching and before reasoning tokens: Astra $0.60, Sol 6.1 $0.12, Terra 6.1 $0.06, Luna $0.006.

How it compares with the other big families

Claude's prices line up almost step for step: Opus 5.5 at $4 and $20 sits between Astra and Sol, Sonnet 5.5 at $2 and $10 matches Sol 6.1, and Haiku 5.5 at $0.10 and $0.50 matches Luna. See Claude models. Google's current line is on Gemini models. For a chat-product comparison, ChatGPT vs Claude is the place.

Matching prices do not mean matching behaviour. The tokenizers differ, reasoning-token accounting differs, and long-prompt rules differ: OpenAI doubles input above 272K while Claude's Opus and Sonnet charge the same rate across the full 1M window.

Our take

The GPT-6 line is easy to price and hard to name. Four current models, plus GPT-6 Sol and a full GPT-5.6 tier, mean you need the model page open to be sure which snapshot you are calling. The clear points are that each model has the same 1.05M window and 128K output, and that Sol 6.1 is OpenAI's pick for balance.

For a team already on OpenAI, we would start coding agents on Sol 6.1, move individual tasks up to Astra only when evaluations show a gap, and use Luna for the cheap steps. The 272K surcharge is the main thing to design around: keep prompts under it unless the task needs it. We have not run these models ourselves. This is a read of the published price list and documentation.

Checklist

• Pin the model ID, not a dated nickname, and note the knowledge cutoff.
• Set reasoning effort on purpose, since reasoning tokens bill as output.
• Keep prompts under 272K input tokens where possible.
• Cache static prefixes and check the retention you pay for.
• Use Responses API for tool calling on Sol, Terra and Luna.
• Recheck the pricing page before committing; this table is dated 31 October 2026.

Questions we hear

Is Sol 6.1 better than GPT-6 Sol? OpenAI lists both and prices them the same at $2 and $10. We found no side-by-side score from OpenAI.

Does GPT-6 take audio? Not on Sol and Terra, which list audio and video as unsupported. Voice is handled by separate realtime models.

Can I fine-tune it? Astra's endpoint table lists fine-tuning as not supported.