opendunes
Evaluate and optimise
Copy page ▾

GPT-6

Move an app from GPT-5.6 to GPT-6: pick one of three tiers, check the effort, and drop the sampling fields.

  1. Change the model ID.
  2. Update the request.
  3. Read the answer.
  4. Set the effort.
  5. Compare before you switch.

1. Change the model ID

GPT-6 comes as three tiers, each with a -pro ID: the same model, served with pro reasoning, which works longer before one final answer. There is no Terra in GPT-6.

Model IDModel IDThe name a request calls a model by, lab/model, copied from Models.More on Models →ForEffort levels
openai/gpt-6-astrathe most capable: long, many-step work across code and documentslow to max; always reasons
openai/gpt-6-astra-proAstra, with pro reasoninglow to max; always reasons
openai/gpt-6-solstrong reasoning on demanding worknone to max
openai/gpt-6-sol-proSol, with pro reasoningnone to max
openai/gpt-6-lunaefficient, repeated work at scalenone to max
openai/gpt-6-luna-proLuna, with pro reasoningnone to max
FromTo
Model IDopenai/gpt-5.6-sol, -terra or -lunaone of the six above
Context1 050 000 tokens1 050 000 tokens
Longest answer128 000 tokens128 000 tokens

Each ID also exists with :batch at its end, for batch work. GET /v1/model/{lab}/{model} answers each model's entry, reasoning included. The request keeps its shape; model changes.

JSON
{
  "model": "openai/gpt-6-sol",
  "reasoning": {
    "effort": "medium"
  },
  "messages": [
    {
      "role": "user",
      "content": "Summarise the incident report in five lines."
    }
  ]
}

2. Update the request

You send todayOn GPT-6Send instead
effort none on Astrarefusedlow
effort minimalnot among the levelslow, then compare
temperature, top_p, logprobs or top_logprobsOpenAI takes them at effort none onlynothing: steer with the prompt
prompt_cache_retentionreplacedprompt_cache_options: {"ttl": "30m"}
Watch out

On OpenAI's own API, GPT-6 Astra calls tools on the Responses shape only, and Sol and Luna call functions on the chat shape at effort none only. Run your tool calls on /v1/chat/completions before you switch: see Tools and function calling.

One addition changes the effort from one turn onward, and keeps the start of the conversation as it was, so a cached prompt stays cached. Put a system message with empty content and a configuration_update just before the user turn it applies to.

JSON
{
  "model": "openai/gpt-6-sol",
  "reasoning": {
    "effort": "medium"
  },
  "messages": [
    {
      "role": "user",
      "content": "Summarise the incident report."
    },
    {
      "role": "assistant",
      "content": "The outage began at 09:14 when …"
    },
    {
      "role": "system",
      "content": "",
      "configuration_update": {
        "reasoning": {
          "effort": "high"
        }
      }
    },
    {
      "role": "user",
      "content": "Now find the root cause and propose a fix."
    }
  ]
}
RuleWhy
keep each update where it was when you send the history backa moved update changes the start, and the cache is lost
never two updates side by siderefused
never on a -pro IDrefused

3. Read the answer

OpenAI can stop a GPT-6 answer on biology or cybersecurity grounds. It comes back as a 403 whose body says the model refused, sometimes after the answer began. Show the message; sending the same prompt again returns the same refusal. See Errors and retries.

OpenAI lists three ways Astra behaves differently: it asks more questions before it acts, follows instructions in long files more closely, and formats answers with more lists and tables. Say in the system prompt how much it may do without asking, and the style you want.

On /v1/responses, a tool can be marked "async": true: the model calls it and goes on working, and you send its result in a later request, with the same call_id.

4. Set the effort

RouteField
/v1/chat/completionsreasoning.effort, or reasoning_effort

Keep the level you ran on GPT-5.6 as the starting point, and set it in every request. Move Astra off none. Keep the request's level fixed for the whole conversation, and change it for one stretch with a configuration_update.

5. Compare before you switch

Send the same set of real prompts to both IDs, and compare three things.

CompareWhere
the answersyour own check: a test suite, a reviewer, the task's pass rate
what each costusage.cost, in DADAThe Algerian dinar, the one currency the docs show.More on Balance and charges →, on every answer
how much each reasonedusage.completion_tokens_details.reasoning_tokens

GPT-6 comes as tiers, and each tier has its own lowest effort, so the comparison is a sweep with an effort per tier. The script sends one task to the old ID and to each tier, and one more run puts Astra at none to show the refusal. It prints one line per run: how the answer ended, what the model reasoned, and the charge in DA, then the first words of the answer. A refused request prints its status; show a refusal as it is, because sending the same prompt again returns it again.

RunModelEffort
fromMODELEFFORT
astraASTRAlow
astra, noneASTRAnone
solSOLnone
lunaLUNAnone
import os
from openai import APIStatusError, OpenAI

API_KEY = os.environ["OPENDUNES_API_KEY"]
BASE_URL = "https://opendunes.com/v1"
MODEL = "anthropic/claude-sonnet-4.5"
ASTRA = "anthropic/claude-sonnet-4.5"
SOL = "anthropic/claude-sonnet-4.5"
LUNA = "anthropic/claude-sonnet-4.5"
EFFORT = "medium"
MAX_TOKENS = 8000
TASK = (
    "Summarise the incident report in five lines.\n\n"
    "The outage began at 09:14 when the payment queue filled. "
    "The team restarted the worker at 09:40. Orders placed in "
    "between were retried. No card was charged twice."
)
RUNS = [
    ("from", MODEL, EFFORT),
    ("astra", ASTRA, "low"),
    ("astra, none", ASTRA, "none"),
    ("sol", SOL, "none"),
    ("luna", LUNA, "none"),
]

client = OpenAI(base_url=BASE_URL, api_key=API_KEY)

for label, model, effort in RUNS:
    try:
        reply = client.chat.completions.create(
            model=model,
            max_tokens=MAX_TOKENS,
            messages=[{"role": "user", "content": TASK}],
            extra_body={"reasoning": {"effort": effort}},
        )
    except APIStatusError as error:
        print(f"{label:12} refused, status {error.status_code}")
        continue
    usage = reply.usage
    details = usage.completion_tokens_details
    thought = details.reasoning_tokens if details else 0
    print(
        f"{label:12} {reply.choices[0].finish_reason:7}",
        f"thought {thought}",
        f"cost {usage.cost}",
    )
    print("   ", (reply.choices[0].message.content or "")[:60])

The samples put anthropic/claude-sonnet-4.5 in every seat, so they run as they stand. Set MODEL to openai/gpt-5.6-sol, -terra or -luna, and the others to the IDs below.

SeatModel ID
ASTRAopenai/gpt-6-astra
SOLopenai/gpt-6-sol
LUNAopenai/gpt-6-luna

With the GPT-6 IDs set, the line astra, none comes back refused, and the other tiers answer. Read the three tiers against your own check, and keep the effort you ran on GPT-5.6 as the first level to try on each.

Switch once the new ID holds; going back is the same one-line change. Run your red-team suite on the new ID too: Red teaming.

Last modified on