Roar AI
DiscordGet started
Resources

GPT-6.2 Astra, GPT-6.2 Sol and Gemini 4 Neon: Pricing and API Tests

GPT-6.2 Astra, GPT-6.2 Sol and Gemini 4 Neon are live on the Roar AI API. Their pricing and benchmarks, plus what a call cost and how long it took in our tests.

In short
  • GPT-6.2 Astra, GPT-6.2 Sol and Gemini 4 Neon are available on the Roar AI API today as gpt-6.2-astra, gpt-6.2-sol and gemini-4-neon.
  • GPT-6.2 Astra costs less than the GPT-6 Astra it replaces: $8 per million input tokens and $40 per million output, against $10 and $50. GPT-6.2 Sol is priced the same as GPT-6.1 Sol.
  • In our tests, Neon's thinking made a five-word reply cost seven times or more what Astra charged for the same reply, though Astra's rates are higher. Compare what a finished task costs, not the rate.

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November, and Google released Gemini 4 Neon a day later. All three are live on Roar AI from today: send the model name with your existing key. Before announcing them, we sent each one the same requests through our API. Below are the numbers people ask about first, then what we measured.

What are the three models?

GPT-6.2 Astra is OpenAI's new flagship, which OpenAI calls its most capable model for the hardest projects, from start to finish. GPT-6.2 Sol is its everyday model, "from coding to powering agents at scale", in OpenAI's words. Gemini 4 Neon is the Gemini 4 model Google describes as balancing depth and efficiency, between the faster Hydrogen and the heavier Carbon. Google suggests it in place of Gemini 3.1 Pro.

GPT-6.2 AstraGPT-6.2 SolGemini 4 Neon
Model IDgpt-6.2-astragpt-6.2-solgemini-4-neon
Released17 November 202617 November 202618 November 2026
Context window1,050,000 tokens1,050,000 tokens1,048,576 tokens
Longest answer128,000 tokens128,000 tokens65,536 tokens
InputText, imagesText, imagesText, images, PDFs
ReasoningAlways on, adjustableAlways on, adjustableAlways on, adjustable

None of the three can be told not to think. The full specs and the current rates are on each model's page: GPT-6.2 Astra, GPT-6.2 Sol and Gemini 4 Neon.

Benchmarks

OpenAI's launch post was not available to us, so for the GPT-6.2 models we report independent results only. Two testers have run all three: Artificial Analysis, whose Intelligence Index combines tests of reasoning, knowledge, coding and agent work, and Vals AI, whose Vals Index covers finance, legal, tax and coding tasks.

Artificial Analysis Intelligence IndexVals Index
GPT-6.2 Astra6267.7%
Gemini 4 Neon6170.9%
Claude Opus 5.55970.2%
GPT-6.2 Sol5766.1%
GPT-6 Astra57—
Claude Sonnet 5.5—70.9%
Gemini 4 Argon—69.2%

Artificial Analysis scores are at each model's highest effort, which for Neon is a level our API does not offer yet (see the end of this post). On the Vals Index, Neon and Sonnet 5.5 are a fraction of a point apart. A dash means the tester did not publish a comparable figure.

The two indexes disagree, and the reason is useful. Astra's lead is in reasoning and science: Vals ranked it first of all models on its Science Index and on Terminal-Bench-Science. On finance, legal and coding work it falls behind the Claude models and both Gemini 4s. Neon is the reverse, ranking first on Vals' finance and legal indexes.

GPT-6.2 Sol is not a clear upgrade for coding agents. Artificial Analysis found its coding-agent score three points below GPT-6.1 Sol's, mostly because it more often left key facts out of its answers about code. On everything else it gained, and it now scores level with the older GPT-6 Astra.

Google published a table comparing Neon with Claude Opus 5.5 and GPT-6.2 Sol. These are Google's figures:

BenchmarkNeonClaude Opus 5.5GPT-6.2 Sol
GraphWalks, 256K to 1M tokens95.2%66.8%80.1%
AutomationBench56.3%42.5%41.3%
HLE-Diamond71.3%67.9%64.0%
Terminal-bench 4.165.2%65.8%61.5%
Chartography73.5%66.3%75.1%

The first row is the clearest gap: finding and connecting facts across a very long input. On agentic coding, Google's own figures put Neon level with Opus rather than ahead.

Pricing

Our rates, per million tokens:

InputOutputCached input
GPT-6.2 Astra$8$40$0.40
GPT-6 Astra$10$50$1
GPT-6.2 Sol$2$10$0.10
GPT-6.1 Sol$2$10$0.10
Gemini 4 Neon$3$15$0.15
Gemini 4 Argon$2$10$0.10
Claude Opus 5.5$4$20$0.20
Claude Sonnet 5.5$2$10$0.20

Batch requests cost half on all three new models. Both GPT-6.2 models charge more for a request over 272,000 input tokens: double for input and one and a half times for output. Neon also has flex and priority tiers, at half and 1.8 times the standard rate. The pricing page always shows the rates in force.

What we found testing them

Neon thinks even when you ask it not to think much. We sent each model "Say hi in five words" at low, medium and high effort. The GPT-6.2 models answered in 10 or 11 output tokens at almost every setting; only Astra on high thought a little, using 59. Neon used 225 to 299 output tokens every time, nearly all of it thinking, and low saved little: three runs at low averaged 246 tokens, against 299 at high. Google bills thinking as output, so a hello on Neon cost about 0.4 cents, against 0.05 cents on Astra at low or medium effort and about 0.01 cents on Sol.

The same goes for a harder question. We asked each model for the sum of the first 1,000 prime numbers, at medium effort. All three answered 3,682,913, which is correct. Astra used 19 output tokens and Sol 31; Neon used 1,274. That made Neon's answer cost 1.9 cents, about twenty times Astra's and fifty times Sol's.

The GPT-6.2 models were quick at low effort. Every Astra and Sol request returned in 2 to 3 seconds. Neon took 5 to 21 seconds for the hello and 14 seconds for the primes. Independent tests at the highest settings are much slower for all three: Artificial Analysis measured over three minutes before Astra's first answer word at max effort, about 80 seconds for Sol and about 40 seconds for Neon at its default. Use the lowest effort that gets the answer you need.

These are small samples: a handful of requests per model, not a benchmark. The pattern was consistent across them, but measure it on your own prompts.

Which one to use

  • GPT-6.2 Sol for everyday work and agents at volume. It is concise, fast at low effort, and level with GPT-6 Astra on Artificial Analysis's index. For a coding agent that already works on GPT-6.1 Sol, test before you switch.
  • GPT-6.2 Astra for hard science and maths, where the answer can wait. It replaces GPT-6 Astra at a lower rate, so there is no reason to start new work on the older model.
  • Gemini 4 Neon for finance, legal and long-document work, where it leads the independent tests. Budget for its thinking, and keep it out of live chat.

We compared each one with the model it replaces and with its closest rival: GPT-6.2 Astra vs GPT-6 Astra, GPT-6.2 Astra vs Claude Opus 5.5, GPT-6.2 Sol vs GPT-6.1 Sol, GPT-6.2 Sol vs Claude Sonnet 5.5, Gemini 4 Neon vs Claude Opus 5.5 and Gemini 4 Neon vs Gemini 4 Argon.

Calling them

Each is one model name on the endpoint you already use:

{
  "model": "gemini-4-neon",
  "messages": [
    { "role": "user", "content": "Summarise the risks in this contract." }
  ],
  "reasoning_effort": "low"
}

reasoning_effort takes low, medium or high on all three, and Neon defaults to high when you leave it out. Neon ignores sampling settings such as temperature, as every Gemini since 3.6 Flash does.

Limits through our API

Neon accepts a fourth effort level, xhigh, and Google offers answers of up to a million tokens and video and audio input for it. None of these are available through our API yet: a request for xhigh is refused with a message saying so, and Neon's answers are capped at 65,536 tokens. The GPT-6.2 models read text and images, and answer in text.

Start building
in minutes

Get a key, call any model, ship your app. Pay in your local currency.