Releases

GPT-6.2 Astra and Sol

MiniRouter and GPT-6.2 Astra and Sol by OpenAI, separated by a silver fold of light on black.

OpenAI's new Astra and Sol for demanding projects and everyday work. Compare the models on a task you can verify, then choose the effort and cost that fit.

Astra for demanding projects. Sol for everyday coding and writing.

Example brief

When a payment times out, should we retry?

Current policyRetry every failure three times.

  1. Duplicate-charge and timeout failure cases
  2. An idempotent retry policy
  3. Acceptance tests that catch both
Choose a model
GPT-6.2 Sol · cURL
curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H 'Content-Type: application/json' \
  --data '{
  "model": "openai/gpt-6.2-sol",
  "reasoning_effort": "low",
  "max_completion_tokens": 4096,
  "messages": [
    {
      "role": "user",
      "content": "Review this payment retry policy: retry every failed payment three times. Identify duplicate-charge and timeout failure cases. Propose an idempotent retry policy and acceptance tests. Return a short engineering decision memo."
    }
  ]
}'

Set MINIROUTER_KEY to your API key.

Better work, fewer tokens.

On GDPval v2, Astra scores higher with 22% fewer output tokens than GPT-6 Astra.

GDPval v2 improvement

Percentage points over each predecessor

GPT-6.2 Astra vs GPT-6 Astra+6.2
GPT-6.2 Sol vs GPT-6.1 Sol+5.3

Astra output tokens

Indexed, GPT-6 Astra = 100

−22% output tokens

Source: OpenAI
Data and methodology

Each model is compared with its own predecessor. The token index uses OpenAI's reported 22% reduction on the same evaluation. Effort settings were not specified.

ComparisonChange
GPT-6.2 Astravs GPT-6 Astra+6.2 pp
GPT-6.2 Solvs GPT-6.1 Sol+5.3 pp
Astra output tokensvs GPT-6 Astra−22%

Near-Astra coding at a lower cost.

On FrontierCode 2.0, Sol comes within half a point of Astra at roughly one-fifth of the task cost.

FrontierCode 2.0 · Sol relative to Astra

≤0.5points behind

Task cost

Relative, Astra = 1

Source: OpenAI · both models at max effort.

Interfaces, judged by people.

In AA-3D, people explore anonymous browser experiences and choose their favorite.

AA-3D · interactive experience preference

Elo · higher is better · axis starts at 1,200

  • GPT-6.2 Astra1,318
  • Claude Fable 5.21,271
  • GPT-6.2 Sol1,259
Source: OpenAI
Chart data

OpenAI-reported Elo ratings. Effort settings and uncertainty intervals were not supplied.

ModelElo
GPT-6.2 Astra1318
Claude Fable 5.21271
GPT-6.2 Sol1259

Try GPT-6.2.

Get an API key, choose a model and copy the request above.

Context
1.05M tokens
Output ceiling
128K tokens
Reasoning effort
low, medium, high, xhigh or max
Sampling
Omit temperature and top_p
Balance required. A small output cap can still require a hold for 128K output tokens. Final charges follow usage.
Specifications: Astra API guide · Sol API guide.

Astra pricing and setup →
Sol pricing and setup →

Keep exploring

Compare published results in the model rankings and coding benchmarks, or read more model releases.

For MiniRouter product updates, visit the changelog.

Sources

  1. Introducing GPT-6.2 Astra and Sol

    Published Nov 17, 2026

    Read at OpenAI ↗
  2. GPT-6.2 Astra API model guide

    Published Nov 17, 2026

    Read at OpenAI ↗
  3. GPT-6.2 Sol API model guide

    Published Nov 17, 2026

    Read at OpenAI ↗
  4. Vercel AI Gateway model catalog

    Published Nov 17, 2026

    Read at Vercel ↗

Editorial reference revised 2026-11-17. Catalog status comes from the generated routing snapshot.

Every model, one key.