Skip to content

ChatGPT

GPT 6.2 Sol vs Astra: 4x Cheaper, $10,800 Off on One Row

We gave GPT-6.2 Sol and GPT-6.2 Astra the same 40-row sales sheet, three runs per setting, the day after launch. Astra scored 10/10 every time. Sol cost about a quarter as much but dropped the same row in all six runs, against the prompt's rule.

By Duy Lam Le, written with AI9 min read

Why trust this: every result, number and screenshot comes from a real run on our own accounts, not from the AI. How we test

Đọc bản tiếng Việt →

Screenshot: median table of our GPT-6.2 Sol vs Astra test, Sol at $0.0085 a run against $0.0337 for Astra 6.2
Screenshot from the test run.

GPT 6.2 Sol costs a quarter of GPT-6.2 Astra per token and got our messy 40-row sales sheet right except for one row. Astra 6.2 scored 10/10 in six runs. Sol 6.2 scored 9/10 in six, each time dropping the same row on its own judgment, $10,800 off one branch total. Per run, Sol cost $0.0085 against $0.0337.

OpenAI released both models on 17 November, with Sol at $2/$10 per million tokens and Astra at $8/$40. On 18 November we ran them through the OpenAI API from a Python script, three runs per setting, on the same sheet we used for GPT-6.1 Terra in October (Claude Fable 5.2 ran through the Anthropic API for comparison). The sheet, the answer key and the scoring code are the ones we committed on 27 October, unchanged, so today's numbers sit next to the older ones. The raw answers are kept with this article's source files.

At a glance

Key results

  • GPT-6.2 Astra: 10/10 in all 6 runs (effort medium and high), median 8.8 s and $0.0337 per run on high at list price.
  • GPT-6.2 Sol: 9/10 in all 6 runs at $0.0070–$0.0085 per run, about 4 times cheaper than Astra 6.2, and the fastest to start answering (4.4 s on medium).
  • Sol's lost point was a judgment call, repeated in every run: it excluded a row priced "1.2k" because $1,200 did not match the $3,200 on the other filing cabinet rows. The prompt said to normalize shorthand and count it. Every other model counted it.
  • The previous Sol (GPT-6.1 Sol) scored 10/10 three times on the same sheet, for $0.0097 per run.
  • Astra 6.2 beat GPT-6 Astra on cost and time: $0.0337 vs $0.0556 per run, 8.8 s vs 12.4 s, same score.
  • Claims we checked from X: we did not see Astra 6.2 at "21x cheaper, 4x faster" than Claude Fable 5.2. On this task it was 2.4 times cheaper and 1.6 times faster. Another post reported Sol at 72 tokens a second against 200 for the previous Sol; through the API we measured about 98 against 105.

Hands-on test

Setup

  • Task: a CSV of 40 invented Q3 sales rows for an office furniture company with branches in Boston, Chicago and San Francisco, prices in US dollars. Every model got the same prompt: total revenue per branch, the top branch, and every abnormal row by number, returned as JSON.
  • Rules in the prompt: exclude rows that look like bad data; normalize rows that are only written differently (an abbreviation, another name for the same branch or price) and count them; count duplicates once.
  • Planted problems (6): a duplicated order (rows 13 and 39), the date 2026-09-31, a unit price written "1.2k", the branch written "SF", and a quantity of −2.
  • Scoring (code, not AI): 1 point per branch total correct to the dollar, 1 for the top branch, 1 per planted problem found, minus 1 for each clean row flagged. Maximum 10.
  • Models and settings: gpt-6.2-astra and gpt-6.2-sol at reasoning effort medium (OpenAI's suggested starting point for GPT-6.2) and high; gpt-6.1-sol and gpt-6-astra at high; claude-fable-5-2 at its API default. Three runs each, taken in turn so the time of day hit every model the same. 21 runs in total.
  • Timing: streaming on. "Total" is request sent to last token. "First text" is when the first word of the answer arrived, after the model finished thinking. Answer speed is answer tokens divided by the seconds the answer took to stream.
  • Cost: tokens × list price read on the vendor pages on the run day. The table uses full list price; OpenAI's automatic cache discount cut some repeat runs by up to about a third, and the 21 runs together were billed $0.61.

GPT-6.2 Astra vs Sol: the results

SetupScore, 3 runsMedian total timeFirst textAnswer speedCost per run (list)
GPT-6.2 Astra, medium10, 10, 109.6 s7.0 s98 tok/s$0.0275
GPT-6.2 Astra, high10, 10, 108.8 s6.2 s97 tok/s$0.0337
GPT-6.2 Sol, medium9, 9, 97.3 s4.4 s96 tok/s$0.0070
GPT-6.2 Sol, high9, 9, 98.0 s5.4 s98 tok/s$0.0085
GPT-6.1 Sol, high10, 10, 1011.3 s9.6 s105 tok/s$0.0097
GPT-6 Astra, high10, 10, 1012.4 s10.6 s90 tok/s$0.0556
Claude Fable 5.2, default10, 10, 1013.7 s11.2 sn/a$0.0820
Median table of the 7 setups: scores, time, first text, answer speed and list cost per run
Median of three runs per setup, printed by summary.py from the saved results. Sol 6.2 is the cheapest and quickest to answer; it is also the only model that lost a point.

Medium vs high made little difference on either GPT-6.2 model. Both used a median of about 225 reasoning tokens on medium and 375 on high, and total times on the two settings were within a second of each other. On Sol, high did not fix the lost point.

Terminal output of all 21 runs with time, tokens, reasoning tokens, cost and score
All 21 runs as they happened on 18 Nov. The red rows are GPT-6.2 Sol: 9/10 every time, on medium and on high.

Why Sol 6.2 lost the same point every time

Row 18 is a Chicago order for nine filing cabinets with the unit price written "1.2k". The other five filing cabinet rows are priced $3,200. Our prompt treats "1.2k" as shorthand: normalize it to $1,200 and count the row. That is what the answer key does.

Sol 6.2 read the price correctly and then dropped the row anyway. In all six runs it wrote a note like this one: "The listed unit price of 1.2k means $1,200, which differs from the $3,200 price on the other filing cabinet rows, so this row was excluded." Its Chicago total came out at $273,450 instead of $284,250. Astra 6.2 counted the row in all six runs. Claude Fable 5.2 noticed the same price gap and said so, then counted the row as the rule asked.

The six filing cabinet rows with row 18 highlighted, and the row 18 notes from Sol 6.2 and Astra 6.2
The rows that decide the Chicago total, from the sheet every model received, and what Sol 6.2 and Astra 6.2 each wrote about row 18 in run 1.

You can argue Sol has a point: a filing cabinet at a third of the usual price is worth a question. But the prompt gave a rule for shorthand prices, and Sol put its own judgment above it. What matters for anyone using it on real data is that the behavior was steady. Six runs out of six, on two effort levels, Sol did the same thing. In October, GPT-6.1 Terra on its default setting flipped on this row from run to run. Sol 6.2 did not flip; it decided.

GPT 6.2 Sol vs 6.1 Sol

The older Sol (GPT-6.1 Sol) counted row 18 and scored 10/10 three times. Sol 6.2 was faster in total (8.0 s vs 11.3 s on high) because it thought less: 376 reasoning tokens against 585. Its answer streamed slightly slower, about 98 tokens a second against 105. Both list at $2/$10 per million tokens, so the cost difference ($0.0085 vs $0.0097) comes only from the shorter thinking.

GPT-6.2 Astra vs GPT-6 Astra

Astra 6.2 lists at $8/$40 per million tokens, down from $10/$50 for GPT-6 Astra. On this sheet it also used fewer tokens (604 vs 873 output on high), so the cost per run fell 39%, from $0.0556 to $0.0337. It was faster in total too, 8.8 s against 12.4 s. Same score.

GPT-6.2 Astra vs Claude Fable 5.2

Fable 5.2 and Astra 6.2 have the same list price, $8/$40 per million tokens. Astra 6.2 still cost 2.4 times less per run here ($0.0337 vs $0.0820) because it wrote 604 output tokens to Fable's 1,722, thinking included, and the same prompt counted as 1,192 input tokens on OpenAI and 1,645 on Anthropic. With OpenAI's cache discount on repeat runs, the billed gap was about 3 times. Astra 6.2 finished in 8.8 s against 13.7 s. Both scored 10/10 three times. A claim on X put Astra at about 21 times cheaper and 4 times faster than Fable 5.2; on one short structured task, we measured 2.4 to 3 times cheaper and 1.6 times faster.

GPT-6.2 pricing

ModelInput per 1M tokensCached inputOutput per 1M tokens
GPT-6.2 Astra$8.00$0.40$40.00
GPT-6.2 Sol$2.00$0.10$10.00
GPT-6.1 Sol$2.00$0.10$10.00
GPT-6 Astra$10.00$1.00$50.00

Prices from OpenAI's model pages, read 18 Nov 2026. Requests over 272K input tokens cost more on all four. Output includes reasoning tokens, which is why the same model can cost more per task on a higher effort setting.

GPT-6.2 Sol release date and price

OpenAI's API changelog lists GPT-6.2 Astra and GPT-6.2 Sol under 17 November 2026. Both are in the Responses API and Chat Completions; tool calling needs the Responses API. Both reject reasoning effort "none"; OpenAI's migration guide says to use "low" if you used "none" or "minimal" before. Sol 6.2 keeps the GPT-6.1 Sol price.

What worked

What worked

  • JSON output parsed on all 21 runs; no retries for format.
  • Every setup found all six planted problems in every run and flagged no clean rows. Every lost point came from the same cell.
  • Reusing the sealed October task let us compare models across a month: same bytes in, same scoring code.

Where it fell short

What didn't work

  • One short task cannot rank models in general. OpenAI positions Sol 6.2 for coding and agents at scale; we did not test coding, tools or long context.
  • Answer speed is measured on short answers (about 150 to 250 answer tokens), so it is rough. The X posts may have measured inside Codex or ChatGPT, which we did not test. Anthropic does not report thinking tokens separately, so we could not split Fable's answer speed out.
  • We did not test Codex, so we have no number for the claim that Astra 6.2 runs slower than GPT-6 Astra inside Codex. Through the API, on this task, it was faster.

Our verdict

Use GPT-6.2 Astra when the rules have to be followed exactly, as in totals that go into a report. It scored 10/10 every time, and it is now cheaper and faster than GPT-6 Astra. Use GPT-6.2 Sol when cost and speed matter more and you can write down every edge case: at about a quarter of Astra's cost it got everything else right, but it overrode a rule we gave it in six runs out of six. If you move from GPT-6.1 Sol to 6.2 Sol for this kind of work, test your own edge cases first; on our sheet the older Sol was the more literal one. If you only need cheap and literal, GPT-6.1 Terra on high scored 10/10 on this sheet in October at $0.0062 a run (see our GPT-6.1 Terra spreadsheet test). For prices across vendors, see our AI model pricing and rate limits comparison.

This test is also written up in Vietnamese: GPT 6.2 Sol và Astra.

FAQ

Is GPT-6.2 Sol as good as Astra?

Not on our sales sheet. Astra 6.2 scored 10/10 in six runs; Sol 6.2 scored 9/10 in six, always dropping the same row that the prompt said to count. On everything else they matched.

How much does GPT-6.2 Sol cost compared with Astra?

About 4 times less per run on our task: $0.0085 vs $0.0337 on high effort at list price. The per-token price is a quarter of Astra's, and both used similar token counts.

Is GPT-6.2 faster than GPT-6.1?

In total time, yes on our task: Sol 6.2 took 8.0 s against 11.3 s for GPT-6.1 Sol, because it thought less. Its answer streamed slightly slower, about 98 vs 105 tokens a second.

How this was made: an AI (Claude) wrote the test script, ran the 21 API calls and wrote this page. Every number comes from the saved raw results; prices and dates were checked against the OpenAI and Anthropic pages linked in the sources.

ChatGPT

GPT Image 2.6 Sunburst vs Flare: Same Price, 2.3x Faster

We sent the same text-heavy banner prompt to both GPT Image 2.6 models over the API. Both spelled every line right, even in Vietnamese, at $0.034 an image; Flare returned in about 9 seconds, Sunburst in 21, and Sunburst added text we never asked for.

ChatGPT

ChatGPT Pro API Credits vs Claude Max: What $500 Buys

OpenAI now gives ChatGPT Pro and Business up to $500 a month in API credits, the same shape as Claude Max's. We ran one messy sales sheet on seven models both programs pay for: $200 buys about 32,700 GPT-6.1 Terra runs or 7,000 Claude Opus 5.5 runs.

ChatGPT

GPT-6 Astra vs Terra: Same Contract Answers, 10x the Price

We gave GPT-6 Astra, Claude Fable 5.1, GPT-6.1 Terra and Claude Opus 5.5 a 15,000-word contract with four amendments and ten trap questions. All four got every trap right in every run; cost per read ranged from $0.02 (Terra) to $0.38 (Fable 5.1).

Latest guides

All guides

ElevenLabs

ElevenLabs Free Plan: 10,000 Credits Is 10 or 21 Minutes

We spent ElevenLabs free credits through the API: 10,000 credits a month buys about 10 minutes of speech on Eleven v4 and 18 to 21 on Flash or v4 Turbo. The free tier API key runs every model but refuses cloning, 192 kbps and commercial use.

AI Model Pricing: Same $2 Price, Up to 1.9x the Bill

Claude, Gemini and GPT API prices read from each vendor's page on 14 Nov 2026, rate limits by tier, and our token test: the same paragraph billed 251 to 536 tokens at the same $2 price. Gemini 4 Argon doubles on 1 Dec.