GPT-5.4 vs GPT-5.4 mini: Routing Math After GPT-6
GPT-5.4 mini costs 70% less per token than GPT-5.4. October 2026 prices, worked cost math, and the same routing choice with GPT-6 and GPT-5.6 models.
OpenAI released GPT-5.4 on March 5, 2026 and GPT-5.4 mini on March 17, 2026, according to its API changelog. Mini costs $0.75 per million input tokens and $4.50 per million output tokens. GPT-5.4 costs $2.50 and $15.00. On both sides that is 30% of the price (0.75 / 2.50 = 0.30 and 4.50 / 15.00 = 0.30), so every request mini can handle costs 70% less.
GPT-5.4 is no longer the model OpenAI points you to. Its models page now says to start with GPT-6 Astra, "our flagship model for complex reasoning and coding," and to choose GPT-6.1 Sol to balance intelligence and cost or GPT-6 Luna for cost-sensitive, high-volume workloads. Plenty of production code still calls gpt-5.4 and gpt-5.4-mini, though, and neither has a shutdown date on OpenAI's deprecations page.
This post covers the GPT-5.4 versus mini decision first, then the same decision with OpenAI's current models. All prices are OpenAI's standard rates per million tokens for prompts up to 272K input tokens, checked on its pricing page and model pages on October 12, 2026.
GPT-5.4 and GPT-5.4 mini side by side
| GPT-5.4 | GPT-5.4 mini | |
|---|---|---|
| How OpenAI's models page describes it | "A more affordable model for coding and professional work" | "Our strongest mini model yet for coding, computer use, and subagents" |
| Released | March 5, 2026 | March 17, 2026 |
| Input | $2.50 | $0.75 |
| Cached input | $0.25 | $0.075 |
| Output | $15.00 | $4.50 |
| Batch input / output | $1.25 / $7.50 | $0.375 / $2.25 |
| Context window | 1,050,000 tokens | 400,000 tokens (272,000 max input) |
| Max output | 128,000 tokens | 128,000 tokens |
| Default reasoning effort | none | none |
Two details matter for routing. First, mini accepts at most 272,000 input tokens, while GPT-5.4 takes up to 1,050,000. A request above 272K input tokens cannot go to mini at all, and on GPT-5.4 such a prompt is priced at 2x input and 1.5x output for the full session. Second, both models default to reasoning effort none, so unless your code sets an effort, neither one bills reasoning tokens. That changes with the newer models, as the next sections show.
OpenAI's model pages list no benchmark scores for either model, so this post does not quote any. Your own traffic is the better test anyway.
Where mini is the wrong call
Mini is a cheaper model for a defined set of jobs, not a drop-in replacement for every call. Keep GPT-5.4 when:
- the prompt is longer than 272,000 input tokens, which mini does not accept;
- the task is long, multi-step reasoning, such as planning several steps ahead or reconciling contradictions across many sources;
- the output is the product a customer reads word for word, and a weaker answer costs more than the saving;
- your own tests show a quality gap on that kind of request.
Mini is worth testing first for classification, intent detection, short summaries, structured extraction (JSON from unstructured text), tagging, reranking retrieved chunks, code completion on well-defined functions, and agent steps where the set of possible actions is small.
GPT-5.4 nano ($0.20 input, $1.25 output) was the third option for the simplest tasks. OpenAI deprecated it on October 1, 2026. It shuts down on April 1, 2027, and OpenAI recommends gpt-6-luna as the replacement, so do not build new routing around nano.
The cost math: GPT-5.4 vs mini
Take an app that sends 100,000 requests a month, each with 2,000 input tokens and 500 output tokens. That is 200 million input tokens (100,000 x 2,000) and 50 million output tokens (100,000 x 500).
- GPT-5.4: 200 x $2.50 = $500 for input, plus 50 x $15.00 = $750 for output, so $1,250 a month.
- GPT-5.4 mini: 200 x $0.75 = $150 for input, plus 50 x $4.50 = $225 for output, so $375 a month.
The difference is $1,250 - $375 = $875 a month, or $875 x 12 = $10,500 a year. That is $875 / $1,250 = 70% of the bill.
You do not have to move everything. If half of those requests pass your quality test on mini, and they are the same size as the rest, you save half the difference: $875 x 0.5 = $437.50 a month, and the bill drops to $1,250 - $437.50 = $812.50. That is arithmetic on an assumed share, not a forecast. The share that passes depends on your traffic.
Prompt caching does not change the ratio. Cached input on mini is $0.075 against $0.25 on GPT-5.4, which is also 30% (0.075 / 0.25 = 0.30).
The same decision with today's OpenAI models
GPT-6 Astra came out on September 3, 2026, GPT-6 Luna on September 22 and GPT-6.1 Sol on September 29. The GPT-5.6 family (Sol, Terra and Luna) came out on July 9, and OpenAI's model pages map it onto the older tiers: GPT-5.6 Sol "roughly corresponds to the unsuffixed model tier used in earlier GPT-5 families," Terra to the mini tier, and Luna to the nano tier.
| Model | How OpenAI's models page describes it | Input | Cached input | Output | Default reasoning effort |
|---|---|---|---|---|---|
| GPT-6 Astra | "Our most capable model for the most demanding work" | $10.00 | $1.00 | $50.00 | Not stated; supports low to max, not none |
| GPT-6.1 Sol | "Near-Astra performance for complex work at a lower cost" | $2.00 | $0.10 | $10.00 | medium (none and minimal not supported) |
| GPT-6 Luna | "Our most efficient model for focused, high-volume tasks" | $0.10 | $0.01 | $0.50 | medium (none supported) |
| GPT-5.6 Sol | "GPT-5.6 flagship model for complex professional work" | $4.00 (promotional) | $0.40 | $20.00 (promotional) | medium |
| GPT-5.6 Terra | "GPT-5.6 model that balances intelligence and cost" | $2.00 | $0.20 | $12.00 | medium |
| GPT-5.6 Luna | "GPT-5.6 model optimized for cost-sensitive workloads" | $0.20 | $0.02 | $1.20 | medium |
| GPT-5.4 | "A more affordable model for coding and professional work" | $2.50 | $0.25 | $15.00 | none |
| GPT-5.4 mini | "Our strongest mini model yet for coding, computer use, and subagents" | $0.75 | $0.075 | $4.50 | none |
GPT-5.6 Sol's price is promotional. OpenAI lowered it on August 21, 2026 (20% lower input, 33% lower output) and says the promotional pricing is available at least through November 21, 2026. The GPT-6 and GPT-5.6 models also charge for cache writes, at 1.25 times the input price. OpenAI's pricing page lists no cache-write price for GPT-5.4 or mini.
Reasoning effort changes the math
GPT-5.4 and mini default to effort none. GPT-6.1 Sol, GPT-6 Luna and all three GPT-5.6 models default to medium, according to their model pages and OpenAI's reasoning guide. GPT-6 Astra's model page lists low, medium, high, xhigh and max but does not name a default, so set it explicitly. The guide is clear about billing: reasoning tokens are not visible through the API, but they "are billed as output tokens." It also says a request can use "anywhere from a few hundred to tens of thousands of reasoning tokens," and the real count is in the response's usage object under output_tokens_details.
So if you swap gpt-5.4 for a newer model name and change nothing else, you go from zero reasoning tokens to medium effort, and you pay for that thinking at the output rate.
Here is the same workload as above: 100,000 requests a month, 200 million input tokens and 50 million visible output tokens. The last column assumes 1,000 reasoning tokens per request, which adds 100,000 x 1,000 = 100 million output tokens (150 million in total). That figure is an illustration only. Measure your own.
| Model | No reasoning tokens | With 1,000 reasoning tokens per request |
|---|---|---|
| GPT-6 Astra | 200 x $10 + 50 x $50 = $2,000 + $2,500 = $4,500 | $2,000 + 150 x $50 = $2,000 + $7,500 = $9,500 |
| GPT-6.1 Sol | 200 x $2 + 50 x $10 = $400 + $500 = $900 | $400 + 150 x $10 = $400 + $1,500 = $1,900 |
| GPT-6 Luna | 200 x $0.10 + 50 x $0.50 = $20 + $25 = $45 | $20 + 150 x $0.50 = $20 + $75 = $95 |
| GPT-5.6 Sol | 200 x $4 + 50 x $20 = $800 + $1,000 = $1,800 | $800 + 150 x $20 = $800 + $3,000 = $3,800 |
| GPT-5.6 Terra | 200 x $2 + 50 x $12 = $400 + $600 = $1,000 | $400 + 150 x $12 = $400 + $1,800 = $2,200 |
| GPT-5.6 Luna | 200 x $0.20 + 50 x $1.20 = $40 + $60 = $100 | $40 + 150 x $1.20 = $40 + $180 = $220 |
| GPT-5.4 | $500 + $750 = $1,250 | $1,250 at its default (none) |
| GPT-5.4 mini | $150 + $225 = $375 | $375 at its default (none) |
Three things stand out.
- A lower price per token is not always a lower bill. With no reasoning tokens, GPT-6.1 Sol ($900) costs less than GPT-5.4 ($1,250). At its medium default with 1,000 reasoning tokens per request, it costs $1,900, which is more than GPT-5.4 at its default. GPT-6.1 Sol does not accept
none;lowis its lowest setting. - GPT-5.6 Terra is not the cheaper step down from mini. OpenAI calls it the mini-tier model of the GPT-5.6 family, yet it costs more per token than GPT-5.4 mini ($2.00 against $0.75 input, $12.00 against $4.50 output). On this workload that is $1,000 against $375 with no reasoning tokens.
- GPT-6 Luna is. Per token it costs $0.10 / $0.75 = 13% of mini's input price and $0.50 / $4.50 = 11% of its output price. On this workload it is $45 against $375 with no reasoning tokens ($45 / $375 = 12%, so 88% less), and $95 even with 1,000 reasoning tokens per request. Luna accepts
none, so you can compare it with mini at the same setting. In Chat Completions, Luna supports function calling only withreasoning_effortset tonone.
Astra is the hardest one to price from a table. OpenAI's GPT-6 guide says that in several evaluations Astra "achieved stronger results using substantially fewer output tokens," and that its estimated API cost per task was lower than earlier models despite its higher per-token price. That is OpenAI's claim about its own evaluations. Whether it holds for your requests is something the token counts in your usage logs will tell you.
Which pairs to test
- GPT-6 Astra to GPT-6.1 Sol. OpenAI itself says to "Compare it with Astra on your tasks to assess the tradeoff between quality and cost." Per token, 6.1 Sol is 20% of Astra's price on both input ($2 / $10) and output ($10 / $50), so 80% less, and its cached input is $0.10 against $1.00.
- GPT-6.1 Sol to GPT-6 Luna. This is the GPT-6 version of the flagship-to-mini question. Luna is 5% of 6.1 Sol's price per token ($0.10 / $2.00 and $0.50 / $10.00), so 95% less.
- GPT-5.4 to GPT-5.4 mini. Still 70% less per token if you stay on the 5.4 models.
- GPT-5.4 mini to GPT-6 Luna. Per token, Luna costs about 87% less on input (1 - 0.10 / 0.75) and about 89% less on output (1 - 0.50 / 4.50).
In every pair, keep the reasoning effort the same on both sides, or you end up measuring reasoning tokens instead of models. OpenAI's model selection guide gives starting points that pair a model with an effort level, from Luna at low effort for "simple data extraction" up to Astra at extra high for "demanding analysis and complex deliverables with exacting requirements."
What we saw on our own traffic
Our first-party data on GPT-5.4 mini is small. The requests it covers were sent to Claude Sonnet 5, not to GPT-5.4, so it shows how mini did against Sonnet 5 on those requests. It is not a GPT-5.4 versus mini test.
- In one app the PromptUnit team runs itself (checks up to October 12, 2026), GPT-5.4 mini was judged as good in 23 of 23 checks on the app's video-prompt step, at about 73% lower cost.
- In 24 side-by-side comparisons on one production account (October 5 to 7, 2026), Claude Sonnet 5 to GPT-5.4 mini matched 8 of 10.
- Not every step moves. On the same team app's web-research step, Claude Haiku 4.5 had served for five days with no errors and answers that passed the automatic format checks, yet it was judged worse than Sonnet 5 in 2 of 3 checks, so PromptUnit stopped moving that step and kept it on Sonnet 5.
In each case the judge was Claude Sonnet 5, the model the app asked for, so the judge is not independent, and the samples are small. Treat this as early data, not as a quality guarantee for your traffic.
What to do next
Pull a week of requests from your logs and group them by feature or step. For each group, send 20 real requests to both the model you use now and the cheaper candidate, at the same reasoning effort, and compare the answers with whatever quality check your team already uses: an eval pass rate, a human rating, or a match against expected output. Record input, output and reasoning tokens for both, because the cheaper model per token is not always cheaper per request. Any group where the cheaper model meets your bar is a candidate to move.
If you would rather not run that bake-off by hand, PromptUnit checks each request and uses a cheaper model, such as GPT-5.4 mini, when one is just as good for that request. Otherwise the request goes to the model you asked for. Setup is two settings: point your app's API base URL at PromptUnit and add your PromptUnit key. Your own OpenAI key stays.
Its optional side-by-side checks are off by default and run on your own provider keys within a daily budget you set. After your app has its answer, cheaper models answer a sample of real requests, and the model you asked for judges whether each cheaper answer is as good. A step of your app moves to a cheaper model on its own only after at least 20 judged checks with 9 in 10 as good and at least 10% lower cost, and it moves back if fewer than 7 in 10 stay as good. The dashboard shows each request with the model asked for, the model that answered, its cost and its saving.
It is free until it has saved you $5 (individual accounts) or $100 (business accounts with a work email), then it takes 20% of the savings. If it saves nothing, you pay nothing.
See Also
- LLM Routing: How It Works, Strategies and Real Data
- OpenAI API Cost Too High? Six Fixes, Ordered by Effort
- The Hidden Cost of Defaulting to Your Flagship Model
Try It Free
See where your AI budget goes. PromptUnit shows every request with the model you asked for, the model that answered, what it cost and what it saved.
Start free. It costs nothing until it has saved you money.