OpenAI’s Recommended Migration Path Costs Twice What You Need to Spend

If you are deciding which OpenAI model to use on a client project this month, the documentation can give you the wrong answer. It is wrong by a factor of two on price.

OpenAI’s deprecations page still sends the October 23 and December 11 shutdowns to GPT-5.6 Sol, Terra and Luna. GPT-5.6 Sol costs $4.00 per million input tokens and $20.00 output. GPT-6.1 Sol, released on September 29, costs $2.00 and $10.00. It also scores five points higher on the main independent index.

The short version: the recommended path costs twice as much as the better option. Nothing on OpenAI’s site is untrue. The pages were written at different times, and the rows for this month’s deadlines were never updated. Land on GPT-6.1 Sol for most work and GPT-6 Luna for volume. Use GPT-6 Astra only where your own tests show a gap.

Prices and pages in this guide were checked on October 4, 2026. The lineup changed twice in September, so check the source before you quote anything.

Which OpenAI model to use: the lineup as of October 4

There are seven current models across two generations. None of them is deprecated.

Prices are per million tokens, taken from each model’s page on OpenAI’s developer site. The index column is the Artificial Analysis Intelligence Index at max effort. It is the only independent measure of the GPT-6 tiers. We looked at the launch results in our coverage of the Sol and Luna release.

Read the table before going on. GPT-6.1 Sol costs the same per token as GPT-6 Sol. Its cached input is half the price, at $0.10. It scores four points higher. It sits one point behind Astra at one-fifth of Astra’s price. GPT-6 Luna is half the price of GPT-5.6 Luna at the same score.

GPT-6 Sol is not deprecated and has no shutdown date. Its model page now says "See GPT-6.1 Sol for the newer Sol model." It is gone from OpenAI’s pricing page and from the featured models list. Treat it as a model you can keep but should not start on.

There is still no GPT-6 Terra, and OpenAI has said nothing about filling that slot. A GPT-6.1 Astra was due in October, but OpenAI canceled that release. We covered the decision in OpenAI pauses GPT-6.1 Astra. GPT-6 Astra remains the top tier.

Three OpenAI pages, three different answers

Nobody is hiding anything here. Three documents give three answers, and each is right in its own frame.

The deprecations page answers "what is the nearest match for the thing you are losing." For the October 23 and December 11 shutdowns, it still names GPT-5.6 Sol, Terra and Luna. Those rows were written in April and June, before GPT-6 Sol and Luna existed.

The page has moved a little since September. A new entry dated October 1 sends three older models to gpt-6-sol and gpt-6-luna. Those models shut down on April 1, 2027. So OpenAI now points some migrations at GPT-6. That entry names the older GPT-6 Sol, not GPT-6.1 Sol. This month’s rows are unchanged.

The current-model guide answers "what should you build on now." It now covers GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. It tells teams on GPT-5.6 and earlier to move to the GPT-6 family.

The pricing page holds the fact that decides it. GPT-6.1 Sol does the GPT-5.6 Sol job for half the money.

What the deadlines actually force

Two dates, and neither is negotiable.

October 23, 2026. Seventeen model identifiers stop working. They include gpt-4-0613, gpt-4-turbo, gpt-4o-2024-05-13, gpt-4.1-nano, gpt-3.5-turbo-0125, gpt-image-1, o1, o1-pro, o3-mini and o4-mini. Six fine-tuned model rows go too. We covered what breaks, and the fine-tuning trap, in the October shutoff piece.

December 11, 2026. The GPT-5 snapshot line goes. That covers gpt-5-2025-08-07, its mini, nano and pro variants, plus o3 and o3-pro.

Neither date moved when GPT-6 Sol, GPT-6 Luna or GPT-6.1 Sol launched. Only the best place to land changed.

If you migrate before October 23, the choice is simple. You can move once, to a current model. Or you can move twice because you followed a mapping written months ago.

Picking by workload, not by tier name

The tier names describe price, not fit. Here is how we would decide.

High-volume, low-judgment work. Classification, extraction, tagging, routing and first-pass summaries belong on Luna. This is where the cost gap is largest. One warning from the launch data: GPT-6 Luna scored two points below GPT-5.6 Luna on Artificial Analysis’s Coding Agent Index. It also dropped on real-world work measures while matching on general intelligence. If GPT-5.6 Luna already works for you, test before you switch. If you are starting fresh, start on GPT-6 Luna.

Everyday reasoning work. Drafting, analysis and code a person will review go to GPT-6.1 Sol. It is half the price of GPT-5.6 Sol and five points better. This is the default, and this month’s deprecation rows will not name it.

Work where a wrong answer ships. This means customer-facing output with no review, or anything touching money, eligibility or safety. Until September 29 this was clearly Astra’s job. Now GPT-6.1 Sol sits one point behind Astra at a fifth of the price. Test GPT-6.1 Sol at higher effort first. Pay for Astra only where your own tests show a gap that matters.

Work you can run overnight. Batch and Flex processing cost 50% of standard on every model here. On GPT-6 Luna that is $0.05 input and $0.25 output. Archive tagging and bulk metadata work get cheap enough that budget stops being the limit. For a small publisher with ten years of content, this may be the most useful line in this guide.

One workload, priced four ways

Price tables do not make decisions. A concrete job does.

A small publisher has 8,000 archive articles of about 1,500 words, roughly 2,000 tokens each. The task is a short summary and five topic tags per article. Each needs about 300 output tokens. That is 16 million input tokens and 2.4 million output tokens in total.

The same job costs two hundred times more at the top of the range.

Astra is not bad value. The point is that model choice on a task like this beats almost any other optimization. A careless choice becomes a line item somebody questions. A considered one is a line item nobody notices.

Two honest caveats. Archive tagging is close to the best case for a cheap model. The work repeats, the schema is fixed, and a human will skim the results. A task that needs judgment would not survive the drop to Luna. These are also list prices for one clean pass. Retries, failed parses and evaluation runs are real costs, and none of them is in the math.

The costs that are not in the price table

Three of these have caught people out.

The 272,000-token cliff. Requests above 272,000 input tokens are billed at 2x input and 1.5x output across the entire request. That includes the tokens under the line. A 280,000-token call on GPT-6.1 Sol is billed at $4.00 and $15.00 per million on every token. Chunking below the threshold halves the input cost. It is a change to a retrieval step, not a rewrite.

Cache writes. Writing to the cache costs $2.50 per million on both Sol models and $0.125 on Luna. Reads are cheap: 90% off on GPT-6 Sol and 95% off on GPT-6.1 Sol. Cost models often count the saving and skip the write charge. With high prompt churn and low reuse, caching can cost more than it saves.

Regional processing and EU residency. Regional processing adds 10%. EU data residency works only with standard processing, so you cannot have both. For a client with a European data rule, confirm this before you quote.

There is a quieter cost that never shows on an invoice. Every model in this table has a 1,050,000-token context window. The research says filling it is a mistake. We set out the evidence in when more tokens make AI output worse. Performance dropped with input length on every model tested. A focused prompt beat a full context that held the same answer. A large window is a capability, not an instruction.

Reasoning effort is a price dial

The tier is only half the cost decision. The other half is a parameter.

GPT-6 models take a reasoning effort of none, low, medium, high, xhigh or max. The default is medium. GPT-6 Sol and GPT-6 Luna support none. GPT-6 Astra does not, and neither does GPT-6.1 Sol. A GPT-6 Sol job that runs at none is not a drop-in swap to GPT-6.1 Sol. Set low and test, or keep that job on GPT-6 Sol or Luna. OpenAI’s GPT-6 guide also says to review its migration guidance before switching from gpt-6-sol.

Higher effort means more reasoning tokens. Reasoning tokens bill as output, the expensive side of the meter. Output runs five to six times input across this lineup. That is why the input and output split matters more than the headline price. Turning effort up is a spend decision made in a parameter. On Astra, at $50 per million output tokens, it is an expensive one.

Most benchmark figures OpenAI publishes are at xhigh or max. They describe the expensive setup, not the default you get if you leave the parameter alone.

There is also a quality case against maxing it. Research shows that past a point, longer reasoning makes output worse on some kinds of task. Our token usage and quality piece covers that evidence. Set effort per task, not globally. The maximum is neither the cheapest setting nor reliably the most accurate.

Fast mode and Ultrafast

Speed is a separate purchase from quality, and OpenAI now sells it in two steps.

Fast mode bills at 2x standard where it is offered. It is not available with EU data residency.

Ultrafast arrived at DevDay on September 29. It is live only for GPT-6 Astra, through the ultrafast service tier in the Responses API. OpenAI’s pricing page lists it at $60.00 input and $300.00 output per million tokens. That is six times standard. The GPT-6.1 Sol announcement says a Sol Ultrafast is on the way. It was not live as of October 4. Our DevDay 2026 recap covers which plans include it.

For most client work, neither tier is worth it. Pay for speed only when a slow response blocks paid work.

Where the rest of the market sits

A model selection guide that pretends only one vendor exists would not survive a real project. So, one short section.

On the same index, Claude Opus 5.5 scores 58. Claude Fable 5.1 scores 53, level with GPT-6 Astra. GPT-6.1 Sol scores 52. If you need maximum capability, the top of the OpenAI range is not the top of the market today. If you need low cost per unit of acceptable work, GPT-6 Luna on Batch has no direct match.

We are not running a full cross-vendor comparison here. That needs both sides established first, and we would rather do it properly.

What to tell a client about access

Two notes come up in the first client conversation.

GPT-6.1 Sol is live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. OpenAI says it is "not yet available in Chat." Staff with Plus seats may not find it in the ordinary chat screen. That is a rollout question, not a licensing one. Check before you promise anything.

On the API side, GPT-6.1 Sol and GPT-6 Luna are not gated by spend tier. Both start at Tier 1 with 500 requests and 500,000 tokens per minute. That is enough for a small client integration without a procurement conversation. At Tier 5, Luna reaches 30,000 requests and 180 million tokens per minute. GPT-6.1 Sol reaches 15,000 and 40 million.

We covered what GPT-6 Astra actually shipped in September, including the gap between announcement and launch day. The same caution applies here. A price on a pricing page and access in a client’s account are different questions.

What we would pick

Stated plainly, because a guide that recommends nothing is not a guide.

For most agency and small-publisher work, GPT-6.1 Sol is the default. It is half the price of the GPT-5.6 Sol target, five points better on the independent index, and current.

Move GPT-6 Sol workloads to GPT-6.1 Sol after a test. The per-token price is the same and cached input is cheaper. Check anything that runs at none effort first.

Drop to GPT-6 Luna for volume work. Run it through Batch wherever latency allows. Test before you move an existing GPT-5.6 Luna workload.

Treat Astra as the exception. With GPT-6.1 Sol one point behind at a fifth of the price, Astra needs a measured reason. In our experience, that is less of a project than people assume when choosing a model. It is more than they assume when writing the invoice.

Do the October 23 migration now. Land on GPT-6.1 Sol or GPT-6 Luna, not the GPT-5.6 models the page names. If you already moved to GPT-5.6 this year, the December 11 deadline does not touch you. The price gap still does.

One habit is worth adopting while the lineup moves this fast. Put the model identifier in a configuration value, not in code. Write prices into your own cost model rather than reading them off a page you will not revisit. At this pace, hardcoded model names become technical debt within a quarter.

And check the pricing page before you quote anything. OpenAI changed this lineup twice in September, and there is no sign it is done.

Frequently Asked Questions