GPT-6.2 Sol and Astra review roundup: benchmarks, hands-on tests and API price
GPT-6.2 Astra and Sol launched on November 17, 2026. What Artificial Analysis, Vals AI, Bilibili creators and the tech press found: second on general intelligence, flat on coding, cheap per task. Plus API prices and which one to pick.
In short
- GPT-6.2 comes in two tiers: Astra, OpenAI's top model, and Sol, the everyday one. List prices are $8 in / $40 out per 1M tokens for Astra, 20% below its predecessor, and an unchanged $2 / $10 for Sol.
- Independent tests put Astra second on general intelligence, behind only Claude Fable 5.2. On coding agents Astra is level with its predecessor and Sol is 3 points lower.
- Opinion is split. A Bilibili creator's private question bank gave Astra 93 points at nearly four times Fable 5.2's speed; several well-known developers say it stops mid-task, writes worse and is hard to work with.
- Within its launch week gpt-6.2-sol and gpt-6.2-astra were 3rd and 6th by usage on MoeRouter. Start with Sol at high reasoning effort for daily coding and move to Astra for the hard problems.
On this page
What GPT-6.2 is#
OpenAI released two models on November 17, 2026[1]:
- GPT-6.2 Astra, the top tier. It replaces GPT-6 Astra and is aimed at the hardest engineering, research and long-running agent work.
- GPT-6.2 Sol, the everyday tier. It replaces GPT-6.1 Sol and is meant to be enough for most work.
| GPT-6.2 Astra | GPT-6.2 Sol | |
|---|---|---|
| List price per 1M tokens (input / cached input / output) | $8 / $0.40 / $40 | $2 / $0.10 / $10 |
| Against the previous model | Input and output 20% below GPT-6 Astra ($10 / $50), cached input 60% below | Same as GPT-6.1 Sol |
| Model name on MoeRouter | gpt-6.2-astra |
gpt-6.2-sol |
| Endpoints on MoeRouter | Chat Completions, Responses, Messages | Chat Completions, Responses, Messages |
List prices are as The Decoder and DataCamp report them[1][2].
What it costs on MoeRouter#
| Model | Group | Input | Output | Cache read |
|---|---|---|---|---|
gpt-6.2-sol | default | CN¥0.36≈ US$0.054 | CN¥1.8≈ US$0.27 | CN¥0.018≈ US$0.0027 |
gpt-6.2-astra | default | CN¥1.44≈ US$0.216 | CN¥7.2≈ US$1.08 | CN¥0.072≈ US$0.0108 |
Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().
Both models are in the default group. That group sends each request down the cheapest route that is healthy, which is why the reference price is below list; when the cheap route is down, that request settles at the price of the route that served it, and the usage log shows what was charged. See dynamic routing and billing.
The two previous models are still offered, for comparison:
| Model | Group | Input | Output | Cache read |
|---|---|---|---|---|
gpt-6-astra | default | CN¥1.8≈ US$0.27 | CN¥9≈ US$1.35 | CN¥0.18≈ US$0.027 |
gpt-6.1-sol | default | CN¥0.324≈ US$0.0486 | CN¥1.62≈ US$0.243 | CN¥0.0162≈ US$0.00243 |
Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().
Benchmarks: second overall, standing still on coding#
It matters who ran the test.
Measured by third parties#
- General intelligence. On the Artificial Analysis Intelligence Index, Claude Fable 5.2 scores 65, GPT-6.2 Astra 62, Gemini 4 Carbon 60, Claude Opus 5.5 59 and GPT-6.2 Sol 57. Astra gained 5 points on GPT-6 Astra, Sol 2 on GPT-6.1 Sol[1].
- Coding agents. On the same firm's Coding Agent Index, run in the Codex harness at maximum effort, Astra is level with GPT-6 Astra and Sol is 3 points below GPT-6.1 Sol, because of a 14-point drop on SWE-Atlas-QnA-AA, where it leaves key facts out of its answers more often than its predecessor did[1].
- Coding in a terminal. Vals AI ranks Astra 5th and Sol 7th on Terminal-Bench 4. On the Vals Index Astra is 7th and Sol 9th, with Fable 5.2 first; but Sol costs $3.18 per task, the cheapest in the top ten, against $10.68 for Astra[1].
- Bito's coding benchmark. At high effort: Claude Opus 5.5 85.0%, Claude Fable 5.2 83.1%, GPT-6.2 Astra 77.5%, GPT-6.2 Sol 76.5%. Sol trails Astra by a point and a session costs $2.22 against $8.19, so Bito recommends Sol, with a warning that at its default low effort Sol drops to 69.9%[3].
- Hallucination. On Artificial Analysis's AA-Omniscience test Astra's rate is 59% and Sol's 55%. The questions are unusually hard, so the figures do not reflect everyday use[1].
Reported by OpenAI#
- Agents' Last Exam. Astra is 5.2 points above GPT-6 Astra at under a third of the cost[1]. In the table DataCamp compiled, Astra scores 78.3% at $2.89 per task and Fable 5.2 72.8% at $60.05[2].
- FrontierCode 2.0. Astra 44.2%, Sol 43.8%, at $5.97 and $1.29 per task. In the same table Fable 5.2 scores 49.1% and Opus 5.5 at maximum effort 45.5%, both above GPT-6.2[2].
- GDPval v2, an OpenAI internal evaluation. Astra is 6.2 points above its predecessor with 22% fewer output tokens[1]; in the table DataCamp compiled, Sol scores 46.9% and GPT-6.1 Sol 41.5%[2].
DataCamp's author notes that every score in that table is vendor-reported and that OpenAI gave no Fable 5.2 results on its three security benchmarks. His summary: GPT-6.2 is a price cut that OpenAI has turned into a capability release[2].
Hands-on: Bilibili and the Chinese tech press#
程序员阿江-Relakkes (Bilibili) had pre-release access and ran his private ModelDial question bank at maximum effort. In his video description he reports 93 points for Astra, second on his board, one point behind Claude Fable 5.2's 94 and four to five ahead of Opus 5.5 in third, in 94 minutes where Fable 5.2 took six hours. Three tasks on a ten-thousand-line legacy ERP codebase (partial receiving, return reconciliation, three-warehouse recalculation) all passed at the first attempt[6]. He tested again after launch: Astra passed each of the three in the first round and Sol passed all three too; on front-end work he writes that the taste has improved but he still prefers what the Claude models produce[7].
源叽森 (Bilibili) compared Sol with Astra and describes Sol as visibly improved, better at finishing tasks, steady and free of obvious mistakes[8].
ifanr published two hands-on pieces. In one, Astra built a small FPS level from a single prompt in 17 minutes at Ultra effort, under a heading that calls its 3D ability beyond doubt; the writing test ended with the reviewer saying they would not use it to write, and with this comparison: GPT-6.2 is "like the stereotype of a science student: it does not know how to communicate and cannot work with us well" (our translation)[5]. The other piece argues that the bad reviews after launch were amplified, and relays a developer's cost comparison: on the same prompt, Sol finished in about 31 minutes for $2.07 and Claude Opus 5.5 in about 2 hours 46 minutes for $31.85[4].
DataCamp's author ran two small tests of his own. Building a single-file dashboard from data on 173 college majors cost $1.79 on GPT-6.2 Astra and $2.93 on GPT-6 Astra; both worked first time, but he found the older model's way of aggregating more useful to the reader. Rebuilding a five-slide deck from a screenshot of a template looked the same from all three models; only Sol matched the font sizes exactly, for $0.07[2].
The criticism: why opinion split on day one#
The complaints are about how it is to work with, not about how clever it is.
- Long tasks stall or go in circles. The developer Theo was the harshest: on a rewrite of a TypeScript project in Rust, Astra ran for days without landing an improvement and at one point made the compiler 56% slower; in long tasks it stops to ask trivial questions. He also calls it an excellent reviewer that finds what other models miss, and says it built him a reverse proxy that saves $25,000 a month[4][5].
- Writing got worse. Dan Shipper's team at Every tested for weeks and found the writing weaker than its predecessor's and too terse, and that it often stops early on real projects[4]. ifanr's own writing test reached the same conclusion[5].
- Little progress on real research tasks. Epoch AI tested Astra before launch: its capabilities index ties Fable 5.2, but on real-world research tasks it trails Fable 5.2[1].
These are the experiences of a handful of testers in the first two days; test on your own tasks before drawing conclusions.
Which one to use#
Good for
- Daily coding, bug fixes and tests that a test suite can check: Sol at high effort
- Code review, 3D and game prototypes, problems that need a lot of reasoning: Astra
- High-volume agent pipelines where cost matters
Less suited to
- Long-form writing and explaining what was done
- Multi-day tasks with nobody watching
- Front ends where visual polish matters most
| If you | Then |
|---|---|
| Use GPT-6.1 Sol today | Do not switch everything yet: 6.2 Sol is 3 points lower on the independent coding-agent index. Compare on your own tasks first |
| Use GPT-6 Astra today | Move to GPT-6.2 Astra: no score is lower and the price is |
| Want the highest coding success rate | Read the Claude Fable 5.2 review, or use Claude Opus 5.5 |
| Are on a budget | Sol at high effort is among the cheapest per task in its class |
How to call it#
In Codex, set the model to gpt-6.2-sol or gpt-6.2-astra. Both can continue a session from the previous turn's reasoning:
model_provider = "moerouter"
model = "gpt-6.2-sol"
[model_providers.moerouter]
name = "MoeRouter"
base_url = "https://moerouter.com/v1"
wire_api = "responses"
env_key = "MOEROUTER_API_KEY"
requires_openai_auth = false
With the OpenAI SDK, the base URL is https://moerouter.com/v1:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://moerouter.com/v1",
api_key=os.environ["MOEROUTER_API_KEY"],
)
reply = client.chat.completions.create(
model="gpt-6.2-sol",
messages=[{"role": "user", "content": "Rewrite this function as async and add unit tests"}],
)
print(reply.choices[0].message.content)
Both models also answer on the Anthropic Messages endpoint, so they can be set as the model in Claude Code. If you have no key yet, the quick start creates one. The docs are in Chinese for now.
Frequently asked questions#
What is the difference between GPT-6.2 Sol and Astra?#
Astra is the top tier at four times Sol's list price, and suits hard problems, code review and 3D work. Sol is the everyday tier: on OpenAI's own FrontierCode 2.0 figures it is 0.4 points behind Astra at about a fifth of the cost per task. Sol is enough for most daily coding.
How much does the GPT-6.2 API cost?#
List prices per 1M tokens: Astra $8 input, $0.40 cached input, $40 output; Sol $2, $0.10 and $10. MoeRouter's reference prices are in the table above; accounts are charged in Chinese yuan.
Is GPT-6.2 better than Claude Fable 5.2?#
It depends on the task. Fable 5.2 scores 65 and Astra 62 on the Artificial Analysis Intelligence Index, and 83.1% against 77.5% on Bito's coding benchmark. Astra uses far fewer tokens for the same work: on a Bilibili creator's private question bank it was one point behind Fable 5.2 in under a third of the time.
GPT-6.2 or Claude for coding?#
For success rate and readable code, the independent benchmarks and several developers' experience favour Claude Opus 5.5 and Fable 5.2. For cost and speed, with tests to catch mistakes, GPT-6.2 Sol is the better deal. A middle way is to let Claude write and GPT-6.2 Astra review: even its harshest critic says its code review finds what other models miss.
Should I upgrade from GPT-6.1 Sol to GPT-6.2 Sol?#
The price is the same and the general intelligence index is 2 points higher, but the Artificial Analysis coding-agent index is 3 points lower. If you mostly write code with it, compare on your own tasks before switching. As of November 18, 2026, GPT-6.1 Sol was still the most used model on MoeRouter.
Can I use GPT-6.2 in Codex?#
Yes. Set model in ~/.codex/config.toml to gpt-6.2-sol or gpt-6.2-astra. On MoeRouter both support the Responses endpoint and accept the encrypted reasoning Codex sends back to continue a long session.
Why is the price on MoeRouter different from the list price?#
MoeRouter's default group picks the cheapest of several routes for each request. The model square shows the reference price of the cheapest route now, in Chinese yuan. Prices move with the routes, and each request is charged at the route that served it.
Sources
The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.
- OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains
- GPT-6.2 Astra and Sol: Features, Benchmarks, Tests & Pricing
- GPT-6.2 Sol and Astra benchmark: how they compare to Claude Opus 5.5, Fable 5.2 and GPT-6.1 Sol for coding
- GPT-6.2 Astra 刚发布即遭大量差评,实测后我觉得它有点冤 (in Chinese)
- 深度实测 GPT-6.2:强是真强,难用也是真难用 (in Chinese)
- GPT6.2 Astra全网首测:93分硬刚Fable 5.2,屎山题库直接给我秒了? (video, in Chinese)
- GPT6.2 Astra 实测,又把我的屎山代码题库给刷爆了?! (video, in Chinese)
- 【源说AI】gpt 6.2 sol VS astra 对比如何? (video, in Chinese)
Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.