Table of contents
- 1. The question: 2.5x the unit price, one third the tokens
- 2. Who actually said that low is enough?
- 3. The measured numbers, effort by effort
- 4. Why the totals come out level
- 5. Speed: a quarter of the wait
- 6. The day the premise changes: Sol's promotional price
- 7. For coding, the gap narrows
- 8. Measure it on your own work
- FAQ
A newer model is smarter but more expensive, or so the thinking usually goes. With GPT-6 Astra and GPT-5.6 Sol, though, comparing unit prices settles nothing, because Astra finishes the same job on fewer tokens.
The short answer first. In measurements by the independent benchmarking firm Artificial Analysis, running Astra at its weakest setting, low, costs almost exactly what Sol costs at high for one task ($0.82 against $0.81), and it scores higher (46 against 42). It spends one third the output tokens, and the wait before the reply starts is a quarter as long.
That "almost exactly" rests on a premise with an expiry date, though. Sol is on a promotional price right now, and if the announcement holds it ends in late November 2026, after which the price may go back up. Working from Artificial Analysis measurements and OpenAI's official information as of September 18, 2026, this article lays out cost, tokens and speed at each reasoning effort, what the numbers look like once the premise changes, and how the difference plays out in coding.
Astra (low) and Sol (high) measured with the same ruler
Artificial Analysis Intelligence Index, per task (as of September 18, 2026)
Cost
$0.82 vs $0.81
A tie. The unit price is 2.5x, yet the totals are level
Output tokens
4,000 vs 13,000
Astra reaches the answer on about one third
Wait before the reply starts
2.7 s vs 11.9 s
Less time spent deliberating. This is what you feel
Intelligence Index
46 vs 42
Astra on its weakest setting beats Sol on a strong one
Source: Artificial Analysis, "GPT-6 Astra (low) vs GPT-5.6 Sol (high)". Costs are computed at Sol's promotional price ($4/$20)
1. The question: 2.5x the unit price, one third the tokens
Stop at the unit price and Astra is plainly the expensive model.
| Model | Input (per 1M tokens) | Cached input | Output (per 1M tokens) |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-5.6 Sol (promotional price, 3 months from August 21, 2026) | $4.00 | $0.40 | $20.00 |
| GPT-5.6 Sol (price before the promotion) | $5.00 | $0.50 | $30.00 |
Sources: OpenAI developer documentation, "GPT-6 Astra", the price-cut announcement on the OpenAI developer community (August 21, 2026)
Against the promotional price, Astra's unit price is 2.5x on input and 2.5x on output as well. Judge on that alone and the matter ends at "too expensive to bother with".
What you are billed, though, is unit price multiplied by the tokens actually spent. A reasoning model thinks before it answers, and that thinking is billed as output tokens too, so a model that reaches the right answer on less thinking can be cheaper in total even at a higher unit price. That is the question this article sets out to answer.
2. Who actually said that low is enough?
It started with a claim: Astra at low is better than Sol at high. Let us check where that came from.
It traces back to a post an OpenAI staff member (Tibo) made on X (formerly Twitter) on September 6, 2026. On the OpenAI developer community, an OpenAI-side poster restates it as well, writing that OpenAI's advice is to treat GPT-6-Astra at reasoning low as the replacement for GPT-5.6-Sol at high effort.
🟡 That said, this is guidance given in a post, not something guaranteed as a specification. OpenAI's developer documentation contains neither a comparison between one effort and another nor any claim about token efficiency (we checked). What it does state is that there are five levels of reasoning effort (low / medium / high / xhigh / max), plus pricing and limits. Take it as a rule of thumb and confirm it on your own work.
Which means that backing up "low is enough" takes an independent measurement.
3. The measured numbers, effort by effort
Artificial Analysis runs its Intelligence Index, a bundle of several benchmarks, at each level of reasoning effort, and publishes the cost per task and the output token count along with the score. Here are the figures as of September 18, 2026.
| Model and effort | Intelligence Index | Cost per task | Output tokens | Time to first token | Generation speed |
|---|---|---|---|---|---|
| Astra low | 46 | $0.82 | 4,000 | 2.74 seconds | 49 tok/s |
| Astra medium | 50 | $1.54 | 10,000 | 4.46 seconds | 46 tok/s |
| Astra high | 51 | $1.72 | 12,000 | 47.76 seconds | 46 tok/s |
| Astra xhigh | 53 | $2.31 | 17,000 | 154.86 seconds | 49 tok/s |
| Astra max | 53 | $3.26 | 27,000 | 300.45 seconds | 52.6 tok/s |
| Sol medium | 39 | $0.50 | 8,000 | 4.74 seconds | 60 tok/s |
| Sol high | 42 | $0.81 | 13,000 | 11.91 seconds | 60 tok/s |
| Sol xhigh | 44 | $1.18 | 20,000 | 40.15 seconds | 66 tok/s |
| Sol max | 47 | $1.99 | — | 129.63 seconds | 61.7 tok/s |
Sources: the Artificial Analysis model comparison pages (low vs high, medium vs high, high vs high, xhigh vs xhigh, low vs medium) and the Sol model page. Costs are computed at Sol's promotional price. The output token figures are the displayed values rounded to thousands, and they cover the answer and the reasoning together. Sol max's output token figure is omitted because we could not pin its value down on the chart
Three things come out of this.
- Astra at low wins out against Sol high and Sol xhigh alike. It costs about the same as Sol high (42 points, $0.81) while scoring 4 points higher, and it scores 2 points above Sol xhigh (44 points, $1.18) while costing 30 percent less. It is not, however, cheaper than Sol medium (39 points, $0.50), so that comparison comes down to whether 7 more points is worth $0.32
- Astra gets expensive fast as the effort goes up. $1.54 at medium (about 1.9x Sol high), $3.26 at max (about 4x). The score climbs only 7 points, from 46 to 53
- The only Sol setting that scores above Astra's low is max. For those 47 points it costs $1.99, which is 2.4x as much
4. Why the totals come out level
The reason a 2.5x unit price lands on the same total is that it spends one third the tokens. Plain multiplication is enough to check it.
| Breakdown | Astra low | Sol high |
|---|---|---|
| Output tokens (per task) | 4,000 | 13,000 |
| Output unit price (per 1M tokens) | $50.00 | $20.00 |
| Cost of the output portion (from the displayed figures) | about $0.20 | about $0.26 |
| Total per task (input and caching included) | $0.82 | $0.81 |
Sources: the cost of the output portion is our own calculation from the token counts above and the official unit prices (4,000 x $50 / 1,000,000 = $0.20). The token counts are rounded to thousands, so this calculation is an approximation too. Artificial Analysis also publishes a chart that splits cost by token type (answer, reasoning, caching, input), but the values we could read off it varied, so the comparison here is made on the totals and the displayed token figures. The totals are the figures Artificial Analysis publishes
Look at output alone and it flips: even at 2.5x the unit price, Astra pays less ($0.20 against $0.26). The totals end up level because Astra's higher unit price then bites on the remaining input and caching.
From there comes a practical rule of thumb. The shorter the input and the more the work is dominated by thinking, the better Astra at low does; the more context you make it read every single time, the worse. We cover how context behaves in What Is AI Context? The "Reads but Doesn't Read" Reality of the 1M-Token Era.
5. Speed: a quarter of the wait
If cost is a tie, speed comes next. Sol generates faster per token (60 against 49 tok/s), and yet the time until the answer is fully back is shorter with Astra low. The reason is the thinking time (the wait for the first token) and the difference in how many tokens get emitted.
| Model and effort | Time to first token | Generation time (measured) | Total |
|---|---|---|---|
| Astra low | 2.74 seconds | 91.69 seconds | about 94 seconds |
| Astra medium | 4.46 seconds | 209.95 seconds | about 214 seconds |
| Sol high | 11.91 seconds | 219.90 seconds | about 232 seconds |
| Sol xhigh | 40.15 seconds | 296.86 seconds | about 337 seconds |
| Astra max | 300.45 seconds | 522.00 seconds | about 822 seconds |
Source: Artificial Analysis measurements (the time to the first token, and the generation time per task). The totals are our own addition. The generation-time metric is defined so that it excludes the wait for the first token, which is why the two can simply be added. Note, though, that the wait time and the speed come from a separate probe that emits 500 tokens while the generation time is the index's per-task average, so the two measurements have different origins
Astra low is done in about 1/2.5 of the time Sol high needs. Turn it up to max instead and you will wait five minutes for the first character to appear (300.45 seconds). For interactive use, that gap in waiting is felt more keenly than the cost.
6. The day the premise changes: Sol's promotional price
The "near parity" up to this point holds only at Sol's promotional price. On August 21, 2026, OpenAI announced that it was "dropping API and credit pricing of GPT-5.6 Sol by over 20% for the next 3 months" (input from $5 to $4, cached input from $0.50 to $0.40, output from $30 to $20; it applies to ChatGPT Work and Codex credits as well). If the announcement holds, that makes it a limited-time price running until late November 2026.
🟡 Neither a statement naming the end date nor the price that follows it has come from OpenAI. In the same thread, a user writes "at least through November 21, 2026" and says the Terra and Luna cuts (Terra from $2.50 to $2.00 and from $15 to $12, Luna from $1.00 to $0.20 and from $6.00 to $1.20, effective July 30, 2026) are permanent, but that is a user's post, not OpenAI's announcement.
Assume it returns to the original $5/$30, and input and caching go up 1.25x while output goes up 1.5x. Of one Sol high task ($0.81), about $0.26 is the output portion, so with the remainder at 1.25x and the output at 1.5x the total works out to about $1.08. Take the two extremes, everything weighted to input or everything to output, and it stays inside a $1.01 to $1.22 band. Either way, Astra low ($0.82) comes out 20 to 30 percent cheaper.
Right now (promotion running)
Cost is a tie
Astra low $0.82 / Sol high $0.81. On score and speed, Astra low has the edge.
After the promotion ends (if it reverts)
Astra low is 20 to 30 percent cheaper
Sol high rises to about $1.08 ($1.01 to $1.22 if you take the extremes). Astra low stays at $0.82.
The call
Switching now is unlikely to backfire
Prices are moving in Astra's favor. Even so, measure on your own work before you decide.
7. For coding, the gap narrows
Everything so far concerns the Intelligence Index, the overall score. On the index that measures coding agent work, the gap in price narrows.
On the overall index, at the maximum effort on both sides, Astra ($3.26) is about 60 percent more expensive than Sol ($1.99). On the coding agent index, by contrast, Astra comes to $7.09 per task, only about 15 percent more than Sol. The scores are 62 against 55, putting Astra 7 points ahead, and the tokens are one third. Coding, in other words, is relatively favorable ground for Astra.
That said, 🟡 the coding index does not publish costs at the lower efforts. The very question this article is about, whether low is enough, is unverified for coding. Take the rule of thumb the overall index does support, "a bargain if a weak effort is enough, overpriced once you turn it all the way up", and measure it on your own agent.
Source: Artificial Analysis, "Benchmarking GPT-6 Astra" (Coding Agent Index, compared at max effort. Published September 9, 2026)
Two more conditions make Astra's bill jump, by the way. Above 272,000 tokens of input, input and caching double and output goes up 1.5x. In the other direction, Batch and Flex are half price, and Fast is double. If your pattern is handing over a long document whole every time, that surcharge is what will hit you.
8. Measure it on your own work
Every number above was measured on an assortment of benchmarks. Your work is made of something else. Decide only after measuring on tasks that represent what you actually do. Three steps.
- Run the same request through both settings: set Astra to
lowand Sol tohigh, and send the same inputs 5 to 10 times each. With a single run, one request that happened to think for a long time drags the result around - Read the numbers in the response, not the invoice: record the token counts the response carries (input, output, reasoning). Multiply by the unit price and you have the cost per request. The basics of API pricing are collected in What Is an AI API? A Beginner's Guide to Pricing, Tokens and Model Choice
- Count the retries as well: if the cheap model does not finish in one pass and you send it twice, the real cost is double. Compare cost as "the total until the job is done", not "the price of one call"
FAQ
Q1. Is Astra's low the same strength as Sol's low?
No. The effort names are per-model settings, not a shared scale. In the measurements, Astra low (46 points) comes out above Sol's second-highest setting, xhigh (44 points). Compare on score and cost, not on the name.
Q2. Which effort should I start from?
Starting at low and working up is the safe choice. Astra's cost climbs steeply with effort: about 1.9x Sol high at medium, about 4x at max. The score moves only 7 points, from 46 to 53, so the further you turn it up, the less it pays off.
Q3. Should I switch once the promotional price ends?
Prices are moving in Astra's favor. If the original $5/$30 returns, Sol high lands at $1.01 to $1.22 and Astra low ($0.82) comes out 20 to 30 percent cheaper. That said, 🟡 the price after the promotion has not been published.
Q4. If it uses one third the tokens, does that help on plans with usage caps?
On metered API billing, yes. But nothing published says that a subscription's usage cap is consumed in proportion to token count. Check how your plan treats it separately from the pricing question.
Q5. As of when are these numbers?
The measurements in this article are what Artificial Analysis was publishing as of September 18, 2026. Prices and per-effort results both get updated, so check the source pages for the current values before you decide.
Sources
- Artificial Analysis — GPT-6 Astra (low) vs GPT-5.6 Sol (high) (cost, tokens, speed, wait time)
- Same site — Astra (medium) vs Sol (high), Astra (high) vs Sol (high), Astra (xhigh) vs Sol (xhigh), Astra (low) vs Sol (medium)
- Artificial Analysis — GPT-5.6 Sol (cost and speed at max)
- Artificial Analysis, "Benchmarking GPT-6 Astra" (Coding Agent Index, how cost lines up across efforts)
- OpenAI developer documentation — GPT-6 Astra (unit prices, the five reasoning-effort levels, the 272K surcharge, Batch/Flex/Fast, context)
- OpenAI developer community — the 20% price cut for GPT-5.6 Sol (August 21, 2026) (the promotional price and its deadline, the permanent cuts for Terra and Luna)
- OpenAI developer community — Astra low/medium as a replacement for Sol high (the recommendation, presented as a staff member's remark)
Related articles
- GPT-6 Astra Release: The Complete Guide — what changed, and which plans can use it
- GPT-5.6 Release: The Complete Guide (Luna/Terra/Sol) — where the three models sit
- Claude Opus 5 Released — the other option on the table
- What Is an AI API? — the basics of pricing and tokens
- What Is AI Context? — the cost that bites harder the longer the input
