Thinking Variant

Not yet available.

Some models support an extended internal reasoning phase — a "thinking" step where the model works through the problem before producing its final answer. Append :thinking to an eligible model slug to activate this mode. The model's reasoning trace is returned alongside the answer, and accuracy on complex tasks improves significantly.

Usage#

curl https://opendunes.com/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENDUNES_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-opus-4.8:thinking",
    "messages": [{ "role": "user", "content": "Prove that there are infinitely many prime numbers." }]
  }'

Response shape#

When :thinking is active, the response includes an additional thinking field in the message object alongside the standard content:

{
  "id": "chatcmpl-...",
  "model": "anthropic/claude-opus-4.8:thinking",
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "There are infinitely many prime numbers. Here is the proof...",
        "thinking": "Let me think about Euclid's proof. Suppose there are finitely many primes p1, p2, ..., pn. Consider N = (p1 × p2 × ... × pn) + 1..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 24,
    "completion_tokens": 312,
    "thinking_tokens": 580
  }
}

The thinking field contains the model's internal reasoning steps in plain text. The usage object includes a thinking_tokens count — these are billed separately (see Pricing below).

Streaming with thinking#

When "stream": true is set, thinking tokens stream in a separate thinking delta field before the content deltas begin:

data: {"choices":[{"delta":{"thinking":"Let me consider..."},"index":0}]}
data: {"choices":[{"delta":{"content":"There are infinitely..."},"index":0}]}
data: [DONE]

When to use thinking#

The :thinking variant is most valuable for:

• Multi-step mathematical proofs
• Complex logical reasoning or deduction
• Code debugging where the root cause requires tracing through many steps
• Answers where you want to audit the model's reasoning, not just its conclusion

For simple queries, thinking adds latency and cost without much benefit. Use the standard variant for conversational or lookup tasks.

Pricing (DA)#

Thinking tokens are billed at the model's thinking token rate, which may differ from the standard output token rate. Check the model catalog at GET /api/models/{slug} for the exact DA per million thinking tokens. A typical thinking trace is 300–1000 tokens per request.

Using with the OpenAI SDK#

The thinking field in the response message is an extension — the OpenAI SDK will ignore it unless you access the raw response. Use .model_dump() in Python or inspect the raw JSON in Node.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://opendunes.com/api/v1",
    api_key=os.environ["OPENDUNES_API_KEY"],
)

resp = client.chat.completions.create(
    model="anthropic/claude-opus-4.8:thinking",
    messages=[{"role": "user", "content": "Why is the sky blue?"}],
)
msg = resp.choices[0].message
print("Answer:", msg.content)
# Access thinking via the raw dict
raw = resp.model_dump()
print("Thinking:", raw["choices"][0]["message"].get("thinking"))

Note. Not all models support the :thinking variant. Check the catalog — models with capability: reasoning support it. Applying :thinking to an ineligible model returns a 404 model_not_found error.