Skip to content

Gemini 4 Neon vs GPT-6.2 Sol and MAI Code 2: a 20x Cost Gap

Three coding models shipped in 48 hours. We ran Gemini 4 Neon on the same sales sheet as GPT-6.2 Sol: Neon scored 10/10 every time but cost about 20 times more per run, almost all of it thinking tokens. Prices, vendor benchmarks and what does not line up.

Đọc bản tiếng Việt →

Gemini 4 Neon scored 10/10 on our 40-row sales sheet in all three runs, but it cost about 20 times as much per run as GPT-6.2 Sol on the same task. Neon wrote roughly 18 times more output tokens, almost all of them thinking.

Three new coding models shipped in about 48 hours: OpenAI's GPT-6.2 Sol and Microsoft's MAI Code 2 on 17 November, Google's Gemini 4 Neon on 18 November. All three landed in GitHub Copilot the same week, and each vendor published a price per million tokens. Price per token is not price per job, so we ran Neon on the sheet we had used for GPT-6.2 Sol the day before.

Screenshot: median score, list cost, time and output tokens for Gemini 4 Neon, GPT-6.2 Sol and GPT-6.2 Astra on the same sales sheet

What launched, and when

ModelVendorReleasedAPI price per 1M tokens (input / output)Where you can use it
Gemini 4 NeonGoogle18 Nov 2026$3 / $15 ($0.15 cached input)Gemini API, Gemini app (Pro, Ultra), Antigravity, GitHub Copilot
GPT-6.2 SolOpenAI17 Nov 2026$2 / $10 ($0.10 cached input)OpenAI API, ChatGPT, Codex, GitHub Copilot
MAI Code 2Microsoft17 Nov 2026$0.50 / $2.50 promotional, $1 / $5 from 17 May 2027GitHub Copilot (paid plans); no public API yet

Google's post puts Neon between Carbon (its hardest-task model) and Hydrogen (fast everyday work), with a 1 million-token context window. OpenAI's model page lists GPT-6.2 Sol with a 1,050,000-token context window and notes that the API rejects reasoning effort none. Microsoft says MAI Code 2 is in GitHub Copilot now, that the lighter MAI Code 2 Flash is in limited preview in Microsoft Foundry, and that OpenRouter, Vercel and others come later.

Gemini 4 Neon price and benchmarks

Google lists Neon at $3 per million input tokens and $15 per million output tokens. On coding, Google's own figures are 65.2% on Terminal-Bench 4.1, 46.2% on FrontierCode 2.0 at the default high thinking level, and 58.6 on the Artificial Analysis Coding Agent Index, where Google says Neon ranks first. On SWE-sweep, Google reports 12.8% with the mini-SWE harness and 39.0% inside its own Antigravity tool.

Those are vendor numbers, run by Google or chosen by Google. Neon also has four thinking levels; Google says that at the default high level Neon costs 32% less per task than Carbon.

Gemini 4 Neon vs GPT-6.2 Sol: our same-task numbers

We gave Gemini 4 Neon the sealed sales-sheet task from our GPT-6.2 Sol vs Astra test: the same prompt, the same CSV, the same answer key and scoring script, three runs at the API's default thinking level, through the Gemini API on 19 November.

Model (thinking)Score in 3 runsMedian list cost per runMedian timeMedian output tokens
Gemini 4 Neon (default)10, 10, 10$0.168670.4 s10,924
GPT-6.2 Sol (high)9, 9, 9$0.00858.0 s612
GPT-6.2 Sol (medium)9, 9, 9$0.00707.3 s462
GPT-6.2 Astra (high)10, 10, 10$0.03378.8 s604
  • Neon got every planted problem right, including the row GPT-6.2 Sol dropped in all six of its runs.
  • Neon's per-token price is 1.5 times Sol's, but its cost per run was about 20 times higher. Thinking tokens are billed as output, and Neon thought for about 10,000 tokens per run on a task where Sol wrote about 600 tokens in total.
  • Neon took 70.4 seconds per run against 8.0 for Sol. On this sheet, GPT-6.2 Astra matched Neon's 10/10 for about a fifth of Neon's cost.

This is one task and three runs per model, at Neon's default thinking level. We did not test Neon's lower levels, which Google offers to cut token spend.

MAI Code 2 pricing and benchmarks

GitHub's price table lists MAI Code 2 at a promotional $0.50 per million input tokens and $2.50 per million output tokens until 17 May 2027, then $1 and $5. It is available on Copilot Pro, Pro+, Max, Business and Enterprise; MAI Code 2 Flash also reaches Copilot Free and Student.

Microsoft's headline claim: averaged over four coding benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), MAI Code 2 gives 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's at 5% of the cost. The fine print on the same page says all models ran at max reasoning effort and that MAI Code 2's cost is shown at its 50% promotional price.

We did not run MAI Code 2. On 19 November it had no public API we could call outside GitHub Copilot.

Why the vendor numbers do not line up

  • Each vendor picked its own set of benchmarks. FrontierCode appears in Google's and Microsoft's posts, but Google cites version 2.0 and Microsoft version 1.1. Terminal-Bench appears as version 4.1 for Google and 2.1 for Microsoft.
  • Microsoft compares against GPT-6.1 Sol, the previous Sol, not the GPT-6.2 Sol released the same day.
  • Cost claims depend on token use. Our sheet shows a 20x gap in cost per run between two models whose token prices differ by 1.5x.

Our verdict

For short structured tasks like our sheet, GPT-6.2 Sol is still the cheap option if you write down every edge case, and GPT-6.2 Astra gave Neon's accuracy for a fifth of the cost. Gemini 4 Neon earned its 10/10, but at the default thinking level it spends heavily on thinking, so check your own cost per task before switching, or try a lower thinking level first. MAI Code 2 has the lowest token price of the three until May 2027; for now you can only use it inside GitHub Copilot, and its benchmark claims compare it with the previous Sol.

For the per-token prices of other models, see our AI model pricing and rate limits table. Our Gemini 4 Hydrogen page covers the cheaper model in Neon's family. This piece is also in Vietnamese: Gemini 4 Neon, GPT-6.2 Sol, MAI Code 2.

FAQ

How much does Gemini 4 Neon cost?

$3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, per Google's launch post. On our sales sheet that came to a median $0.1686 per run at the default thinking level.

Is Gemini 4 Neon better than GPT-6.2 Sol?

On our one sales-sheet task, Neon was more accurate (10/10 vs 9/10) but about 20 times more expensive and 9 times slower per run. Google's benchmark figures do not include a direct comparison with GPT-6.2 Sol.

How much does MAI Code 2 cost?

$0.50 per million input tokens and $2.50 per million output tokens in GitHub Copilot until 17 May 2027, then $1 and $5.

Is MAI Code 2 available through an API?

Not publicly on 19 November. Microsoft lists GitHub Copilot now and OpenRouter, Vercel, Hermes, OpenCode and Kilo as coming soon.

How this was made: an AI (Claude) gathered the vendor pages, wrote the script that ran Gemini 4 Neon three times through the Gemini API, and wrote this page. Every score and cost comes from the saved raw results; prices, dates and vendor claims were checked word for word against the Google, OpenAI, Microsoft and GitHub pages in the sources.

Sources

  1. Google: Introducing Gemini 4 Neon (18 Nov 2026)

    Vendor page · Primary source

  2. OpenAI API docs: GPT-6.2 Sol model page

    Vendor document

  3. Microsoft AI: MAI Code 2 and MAI Code 2 Flash

    Vendor page

  4. GitHub Docs: Models and pricing for GitHub Copilot

    Vendor document

  5. GitHub Changelog: MAI Code 2 and MAI Code 2 Flash in GitHub Copilot (17 Nov 2026)

    Vendor page

  6. RanThis: GPT-6.2 Sol vs Astra on the same sales sheet (18 Nov 2026)

    Other source