Back to Blog

GPT-6.2 Coding Benchmark: Sol Solved 97% of 175 Tasks at 60% of Claude Sonnet 5.5's Cost, and Astra Scored the Same at 3.6× the Price

2026-11-17·7 min read·CodeRouter Team

Disclosure: we build CodeRouter, a router for coding agents. We called both GPT-6.2 models through OpenRouter, which sent about half the requests to OpenAI and half to Azure, at the same price. The other models' numbers are the ones already on our benchmark page, measured on November 3, and in our Qwen4-Max benchmark from November 16.

GPT-6.2 Sol and GPT-6.2 Astra went live on OpenRouter on November 17 (UTC). Sol costs $2 per million input tokens and $10 per million output, the same as Claude Sonnet 5.5; Astra costs $8 and $40. These are OpenAI's list prices, and OpenRouter charges the same. Both have a 1.05M-token context. We ran both within two hours of the listing, on the same tasks as every model on our benchmark page. Both are also available by name in CodeRouter's Direct mode since November 17; setup is below the results.

GPT-6.2 Sol solved 97.1% of the tasks with one retry, 92.0% on the first try, for $0.88. That's 60% of what Claude Sonnet 5.5 cost, and Sonnet solved 100%. GPT-6.2 Astra solved exactly as many, 97.1%, for $3.17: 3.6× Sol's cost.

The results

Same 175 single-file Python tasks (Exercism plus Aider's hard set). The model sees the instructions and the stub file and returns the whole file; we run the tests, and on a failure it gets one retry with the failing output.

Model First try With one retry Cost of all 175 tasks Median time per task
Claude Opus 5.5 98.3% 100% $2.61 4.8 s
Claude Sonnet 5.5 93.7% 100% $1.48 4.0 s
GPT-6.2 Astra 92.6% 97.1% $3.17 3.3 s
GPT-6.2 Sol 92.0% 97.1% $0.88 2.7 s
Qwen4-Max 83.4% 96.0% $1.47 11.3 s
Kimi K2.6 86.9% 94.9% $0.57 9.4 s
Claude Haiku 5.5 72.0% 93.7% $0.15 5.5 s

Costs are list prices applied to the tokens each request reported, retries included. For both GPT-6.2 models that equals what OpenRouter charged, request by request.

Both GPT-6.2 models missed the same 5 tasks: dot-dsl and paasio in both task sets, plus tree-building from Aider's hard set. Opus 5.5 and Sonnet 5.5 solved all of them. Neither GPT-6.2 model came near our 32,768-token output limit (the longest response was 1,046 tokens), and neither returned an error.

Try GPT-6.2 in your coding agent

CodeRouter serves both models by name since November 17: gpt-6.2-sol and gpt-6.2-astra. Alibaba's Qwen4-Max has been there since November 16 as qwen4-max (setup and price). They're billed at list price with no markup, plus a 5.5% fee when you buy credits; exact prices, including cache and long-prompt rates, are on the models page. Smart routing doesn't use any of the three: you pick the model.

  1. Sign in and create an API key. New accounts get a 7-day free trial: $1 of usage with no card, $5 once you add a card (it isn't charged; one trial per card). The trial covers GPT-6.2 Sol and Qwen4-Max. GPT-6.2 Astra is in the Opus-class group, like Claude Opus, so it isn't on the free trial.
  2. In the dashboard, switch the account to Direct mode (trial and pay-as-you-go accounts; Team keys always use smart routing). In smart routing, the default, a model name in the request is treated as auto. In Direct mode the model you name answers, with no routing and no fallback. More: Direct mode.
  3. Point your tool at CodeRouter.

Claude Code. Put the key in ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY, and set every model variable, or some calls will still ask for a Claude model:

export ANTHROPIC_BASE_URL="https://www.coderouter.io"
export ANTHROPIC_AUTH_TOKEN="cr_YOUR_KEY_HERE"
export ANTHROPIC_MODEL="gpt-6.2-sol"
export ANTHROPIC_DEFAULT_OPUS_MODEL="gpt-6.2-sol"
export ANTHROPIC_DEFAULT_SONNET_MODEL="gpt-6.2-sol"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="gpt-6.2-sol"
export CLAUDE_CODE_SUBAGENT_MODEL="gpt-6.2-sol"
claude

Cursor, Cline or any OpenAI-compatible tool. Base URL https://www.coderouter.io/api/v1, your CodeRouter key, model gpt-6.2-sol (or gpt-6.2-astra). With the OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://www.coderouter.io/api/v1",
    api_key="cr_YOUR_KEY_HERE",
)
resp = client.chat.completions.create(
    model="gpt-6.2-sol",
    messages=[{"role": "user", "content": "Refactor this function"}],
)
print(resp.choices[0].message.content)

CodeRouter sends GPT-6.2 requests to OpenAI's Responses API, so tools work with reasoning on; reasoning effort and other details are in the integration notes. We haven't run a Claude Code session on GPT-6.2 ourselves yet; it's the setup we tested with DeepSeek, GLM-5.3 and Kimi K2.6 and with Qwen4-Max.

Sol against the models we already use

Astra: the same score as Sol, at 3.6× the cost

Astra's list price is 4× Sol's. It wrote fewer tokens on these tasks (51,887 output tokens against Sol's 59,282) and reasoned less, so its run cost 3.6× as much rather than 4×. That's $3.17, more than Claude Opus 5.5's $2.61, and Opus solved 100%.

Astra solved one more task than Sol on the first try (162 vs 161) and the same 170 with a retry. On this suite, the higher price bought no extra solved tasks.

Why Sol is cheap here: short answers, little reasoning

Sol has Sonnet 5.5's list price, so its lower cost comes from billing fewer tokens. It wrote 59,282 output tokens across all 175 tasks, and 20.4% of them were reasoning. On 105 of its 175 first answers OpenRouter reported no reasoning tokens at all, and no first answer reasoned for more than 690 tokens. Astra reasoned even less: none on 120 first answers, 11.5% of its output.

Qwen4-Max, the day before, was the opposite: 271,582 output tokens on the same tasks, 72% of them reasoning. Its list price is lower than Sol's, and its run still cost $1.47 against Sol's $0.88.

The short answers also make Sol fast: a median of 2.7 seconds per task, against 4 to 5.5 for the Claude models.

All of this is at the default reasoning effort; we didn't set one. A higher effort may change both the cost and the score, and we haven't measured it.

What it means for CodeRouter's routing

Neither GPT-6.2 model is in smart routing. On this suite, Sol comes close to Sonnet 5.5 for less money and time. But single-file tasks aren't long agent sessions, and whether Sol earns a routing slot needs agent-session tests.

Limits

Per-task data: coderouter-gpt62-175-2026-11.json. The other models' data is on the benchmark page and in the Qwen4-Max benchmark.

To try GPT-6.2 Sol on your own code, start a free trial and name gpt-6.2-sol in Direct mode.

Try phase-aware routing on your own workload

One base URL for Claude Code, Cursor or Aider. Free trial, no credit card.

Start free trial

Get weekly AI cost optimization tips

Join developers cutting their agent bills