ChatGPT Models Compared: Which One Does What Better

OpenAI's lineup on 27 October 2026 has four general text models: GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna. Their list prices differ by a factor of 100; their quality differs far less. For most API and agent work GPT-6.1 Sol is the sensible default, Terra handles screens and writing at volume, Luna handles bulk extraction, and Astra earns its price only on the hardest tasks. In ChatGPT Chat, the choice is mostly made for you.

This comparison is for the IT manager, CIO or digital lead at a Dutch or Flemish organisation of 250 to 5,000 employees who must decide which model to standardise on, in ChatGPT seats or in software you build. Crux Digits builds AI systems alongside your ICT partner. Every fact below comes from OpenAI's own pages, read on 27 October 2026, and every benchmark score is OpenAI's claim unless stated otherwise.

Which models are in OpenAI's lineup in October 2026?

Four text models, GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, each with a 1.05 million token context window (OpenAI models).

Astra arrived on 3 September 2026 as OpenAI's most capable model (OpenAI). GPT-6 Sol and GPT-6 Luna followed on 22 September (OpenAI), GPT-6.1 Sol upgraded Sol on 29 September (OpenAI), and GPT-6.1 Terra launched on 26 October 2026 (OpenAI Developer Community).

The four share a shape. Each accepts text and images, returns text, writes at most 128,000 tokens per answer, and supports web search, file search, MCP and computer use in the Responses API. Knowledge cutoffs differ: 30 April 2026 for Astra and 6.1 Sol, 18 May for Luna, 1 June for Terra (GPT-6.1 Terra model page, GPT-6 Luna model page). Reasoning effort runs from low to max on Astra and 6.1 Sol; Terra and Luna add none, for instant replies. Astra returns HTTP 400 if you send none (OpenAI reasoning guide).

Around them sit specialist models: GPT-Live 1 for full-duplex voice conversations (model page), GPT-Realtime-2.1 for voice agents that call tools (model page), and GPT-Image-2.5 Sunburst and Flare for images (OpenAI models). Video is gone: the Videos API and Sora 2 were shut down on 24 September 2026 (OpenAI deprecations).

Which model do employees actually get in ChatGPT?

In Chat, paid plans get GPT-6, which runs on GPT-6 Sol; the Pro option runs on Astra; Free and Go get GPT-5.6 Luna (OpenAI Help Center).

The thinking levels differ by plan. Plus gets Medium and High, but not Extra High or Pro. Pro, Business and Enterprise get every level, including GPT-6 Pro on Astra. A Business Standard seat includes 15 Pro messages a month and a Premium seat 50 a week, shared between GPT-6 Pro and GPT-5.6 Sol Pro (ChatGPT Business models and limits).

GPT-6.1 Sol and GPT-6.1 Terra are not available in Chat at all. They live in ChatGPT Work, the agent for multi-step deliverables, and in Codex, for software work (ChatGPT Work and Codex). There, GPT-6.1 Sol has a default context window of 500,000 tokens, half of what the API offers. OpenAI's Terra announcement lists Plus, Pro, Enterprise and Edu (OpenAI), while the Business help article already lists usage estimates for it, so check your own picker.

Those usage estimates are the real price list of a seat. Per five-hour window, a Business Standard seat allows roughly 5 to 45 local messages on Astra, 15 to 150 on 6.1 Sol and 35 to 285 on 6.1 Terra (ChatGPT Business models and limits). In ChatGPT you therefore standardise on a plan, a seat type and a permission set, not on a model.

What does each model do better, by OpenAI's own numbers?

Astra leads on the hardest coding, science and long-context work; 6.1 Sol comes within a few points; Terra trades depth for speed.

GPT-6 Astra: 57.9% on Terminal-Bench 4.0, 74.1% on DeepSWE v1.1 and 72.6% on the OSWorld 2.0 offline set. On OpenAI MRCR v2 with eight needles between 512,000 and 1 million tokens it scores 96.3%, against 73.8% for GPT-5.6 Sol (OpenAI). It is not best everywhere: in OpenAI's own table it scores 57.2% on Humanity's Last Exam with tools, against 65.0% for Claude Fable 5.1. Our GPT-6 Astra vs Claude Fable 5.1 analysis reads those footnotes in full.

GPT-6.1 Sol: OpenAI says it matches Astra on DeepSWE v1.1 at roughly a fifth of the cost and comes within 2.1 points of Astra on OSWorld 2.0 at about a seventh of the cost per task. On Terminal-Bench Science 0.1 it costs $5.47 per task against $23.80 for Astra, but Astra still scores highest at 68.1% and OpenAI advises it for the hardest science (OpenAI).

GPT-6.1 Terra: 15 points above GPT-5.6 Terra on Agents' Last Exam at medium effort, and on par with Astra there at a tenth of the price per token. It finishes tasks more than three times faster than Astra and 6.1 Sol. On Terminal-Bench 4.0 it scores 50.3% at max effort, and OpenAI itself says Sol and Astra remain the better choice for the most complex coding (OpenAI).

GPT-6 Luna: 66.6% on DeepSWE v1.1 at max effort, and on OSWorld 2.0, at max effort, it beats GPT-5.6 Sol at medium effort for a tenth of the cost (OpenAI).

Two caveats. OpenAI ran these evaluations in its research environment or API, which it says may differ from production ChatGPT. And a single score is not evidence for your workload, as we argue in AI agent model selection.

Which ChatGPT model fits which task?

Use the least expensive model that passes your own test set; for most office work that is Terra or 6.1 Sol, not Astra.

Drafting, email, minutes and proposals: GPT-6.1 Terra in Work, which OpenAI describes as built for clear, natural writing (GPT-6.1 Terra model page). In Chat, GPT-6 at Instant or Medium is enough.

Document analysis over long files, such as a tender, a bestek or an annual report: GPT-6.1 Sol. OpenAI reports it beats Claude Opus 5.5 on GDP.pdf, a test of questions over complex PDFs, at less than half the cost per task (OpenAI). Of the four, only Astra has an OpenAI long-context score published above 512,000 tokens.

Coding and refactoring: 6.1 Sol in Codex, which OpenAI recommends for repeated, long-running work; Astra for large migrations; Luna for summarisation, extraction and focused coding (Codex models).

Agents that use tools: 6.1 Sol. On AutomationBench, 47 tools across sales, finance and HR, OpenAI puts it 2.2 points above Opus 5.5 at medium effort for about a third of the cost (OpenAI).

Screen work, forms and data entry: Terra. OpenAI's own example is reading incoming invoices and entering them into accounting software (OpenAI). Read our computer use analysis before pointing it at an ERP.

Customer-facing chat, classification and extraction at volume: Luna, at effort none or low, after you have measured its error rate on your own questions.

Voice: GPT-Live 1 for natural conversation that hands reasoning to a backend model, at $0.05 per minute (model page), or GPT-Realtime-2.1 for a voice agent with tool calls and a 128,000 token context (model page).

Images: GPT-Image-2.5 Sunburst where editing precision matters (model page), Flare for everyday generation.

What does each model cost in the API?

Per million input and output tokens, standard: Astra $10 and $50, 6.1 Sol $2 and $10, Terra $1 and $5, Luna $0.10 and $0.50.

Cached input costs $1.00, $0.10, $0.05 and $0.01 respectively. Four modifiers change the bill. A prompt over 272,000 input tokens costs twice the input rate and 1.5 times the output rate for the whole request. Batch and Flex cost half. Fast mode costs double. Regional processing, which is how you keep inference in the EU, adds 10% for models released on or after 5 March 2026 (OpenAI pricing).

Reasoning tokens are hidden from the response but billed as output tokens and they occupy the context window (OpenAI reasoning guide). A higher effort setting therefore raises the bill even when the visible answer stays the same length.

If you still run GPT-5.6 Sol at $4 and $20, the move to 6.1 Sol halves the list price (OpenAI pricing).

What does the price gap mean at volume?

At 44,000 example requests a month, one workload costs about $4,400 on Astra, $880 on 6.1 Sol, $440 on Terra and $44 on Luna.

The assumptions, labelled as an example and not a client case: 2,000 incoming documents per working day, 22 working days, 6,000 input tokens and 800 output tokens per document including reasoning, standard processing, no caching, no prompt above 272,000 tokens. That is 264 million input tokens and 35.2 million output tokens a month.

Astra: 264 × $10 plus 35.2 × $50 is $4,400. GPT-6.1 Sol: $528 plus $352 is $880. GPT-6.1 Terra: $264 plus $176 is $440. GPT-6 Luna: $26.40 plus $17.60 is $44. EU regional processing adds 10%, so 6.1 Sol becomes $968. If overnight processing is acceptable, Batch halves it to $440.

Per-token price is not cost per correct answer. OpenAI's own Terminal-Bench Science figures put 6.1 Sol at about 4.4 times less expensive than Astra per task, not five, because the models spend different numbers of tokens. The decision rule: run 200 of your real cases on two adjacent models, count correct outcomes, and divide the monthly bill by that number.

What should a Dutch IT department configure before standardising?

Five things: Astra access, role-based model permissions, residency per model, how you pin model IDs, and a deprecation calendar.

Astra is off by default in eligible ChatGPT Enterprise and Edu workspaces; owners enable it per workspace or per role, and Early Model Access settings do not carry over. Admins can check a member's effective access with Model Test under Admin console, Models, Test. On token-based Enterprise agreements, even Instant usage is billable at your rates (ChatGPT Enterprise and Edu models and limits). In Business, the Higher intelligence setting can add reasoning, and use reasoning allowance, while Instant is selected (ChatGPT Business models and limits).

In the API, all four models support EU regional processing for Responses, Chat Completions and Batch, provided you have an approved retention control such as Zero Data Retention or Modified Abuse Monitoring. Fast mode is not available with EU data residency for Astra, and tracing on the Realtime endpoint is not EU residency compliant (OpenAI data controls). Training defaults and workspace residency are covered in our enterprise AI assistant comparison.

Pinning is weaker than it looks. Each GPT-6 family model page lists a single snapshot, the alias itself (GPT-6.1 Sol model page), while the image model has a dated one, gpt-image-2.5-sunburst-2026-09-08 (model page). Keep a regression test set and rerun it whenever OpenAI announces a change.

Put three shutdowns in your change calendar: gpt-5 and o3 on 11 December 2026, gpt-4o and gpt-4.1 on 14 January 2027 with gpt-6-luna as replacement, and gpt-5.1 and gpt-5.3-codex on 1 April 2027 (OpenAI deprecations).

Where Crux Digits fits

We do not sell ChatGPT seats and we do not administer your tenant; your ICT partner keeps that role. We run your own cases through two or three of these models, report cost per correct answer, and build the application around the model you choose, with the model ID in configuration so that a switch is a change request rather than a rebuild. You own the code. That work is described on our AI development agency page.

For the wider rollout question, see AI implementation in the Netherlands.

Last updated: 27 October 2026.

Frequently asked questions