Models & labs

GPT-6.2 Astra or Gemini 4 Neon: Neon ahead on the new Agent Arena Index, at a 24% lower cost per task

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on November 17, and Google answered the next day with Gemini 4 Neon. On the Agent Arena Index checked on November 21, Neon ranks 3rd with 1,333 points, ahead of GPT-6.2 Astra (6th, 1,291), at $2.52 per task versus $3.33. At list prices, Neon costs about 2.7 times less than Astra per token, and the gap widens beyond 272,000 tokens of context.

Abstract generated illustration for the Models & labs section

Within two days, OpenAI and Google each released a new model for professional work and coding. OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, November 17. Google answered the next day with Gemini 4 Neon, positioned between Gemini 4 Carbon and Gemini 4 Hydrogen. Each vendor publishes its own tests. We compared the two on what can be measured from the outside: list price and the Arena’s new agent ranking.

Price: Neon 2.7 times cheaper than Astra per token

Per million tokens GPT-6.2 Astra Gemini 4 Neon GPT-6.2 Sol
Input $8 $3 $2
Cached input $0.40 $0.15 $0.10
Output $40 $15 $10
Context 1.05 million (922,000 input) 1 million 922,000 input

OpenAI’s API model pages add an important rule: beyond 272,000 input tokens, input costs double and output 1.5 times as much, for the whole request. Google’s price list shows a single rate for Neon. Two examples at list price (our calculation):

Request GPT-6.2 Astra Gemini 4 Neon GPT-6.2 Sol
100,000 input tokens, 10,000 output $1.20 $0.45 $0.30
400,000 input tokens, 10,000 output $7.00 $1.35 $1.75

On a large codebase or a big set of documents, Neon thus comes out five times cheaper than Astra, and even cheaper than Sol. These calculations do not include reasoning tokens, billed as output tokens, whose volume varies by model.

The ranking: Neon ahead, Sol the cheapest

The Arena, formerly LMArena, launched the Agent Arena Index this week. It scores models on real agent sessions (web search, running commands, writing files) and publishes their median cost per task over 14 days. Here is the ranking checked on Saturday morning, November 21:

Rank Model (setting) Index Cost per task Time per task
1 Claude Fable 5.2 (Max) 1,448 $9.11 19.2 min
3 Gemini 4 Neon (High) 1,333 $2.52 8.2 min
6 GPT-6.2 Astra (Max) 1,291 $3.33 11.7 min
11 GPT-6.2 Sol (Max) 1,241 $1.01 8 min

Neon beats Astra by 42 points, at a 24% lower cost per task and with tasks 30% shorter (our calculation). Note that Neon is measured at its High setting and Astra at its maximum setting. Sol costs 60% less than Neon per task, for 92 fewer points. Claude Fable 5.2 remains far ahead of everyone, but at a cost per task more than three times Neon’s. The same day, the Arena’s “New Release Rankings” put Neon 3rd in text and GPT-6.2 Astra 5th in WebDev, the website-building category.

The ranking moves: on November 20, the Arena announced Neon 2nd with 1,336 points. It has slipped to 3rd, just behind Claude Opus 5.5 (Max), at 1,336.

What the vendors say

OpenAI stresses efficiency: on Agents’ Last Exam, GPT-6.2 Astra reportedly gains 5.2 points over GPT-6 Astra at less than a third of the cost per task, and its prices are 20% lower than GPT-6 Astra’s. Google claims a score of 61 on the Artificial Analysis index (3rd place) and 70.9% on the Vals Index (2nd). These figures come from the vendors; they are not measured at the same settings and cannot be compared directly.

Our take

For a developer running agents, Neon is today the best measurable trade-off between score and cost. It is ahead of Astra on the Agent Arena Index, it costs less per task, and its price does not double with long contexts. Astra keeps arguments this ranking does not capture, notably the professional document creation highlighted by OpenAI, and a 5th place in WebDev. For everyday tasks, GPT-6.2 Sol is the budget choice: about one dollar per task. These rankings rely on votes and change from day to day. It is best to test on your own tasks before switching your default model.

Sources

  1. OpenAI — Introducing GPT-6.2 Astra and Sol Primary source
  2. OpenAI Developer Forum — GPT-6.2 Astra and Sol are now available (November 17, 2026) Primary source
  3. OpenAI API — GPT-6.2 Astra model page (pricing) Primary source
  4. OpenAI API — GPT-6.2 Sol model page (pricing) Primary source
  5. Google — Introducing Gemini 4 Neon (November 18, 2026) Primary source
  6. Google AI for Developers — Gemini API pricing Primary source
  7. Arena — Leaderboard (checked November 21, 2026) Primary source
  8. Arena (X) — This Week in the Arena (November 20, 2026) Primary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.