opendunes
Evaluate and optimise
Copy page ▾

GPT-5.6

Move an app from GPT-5.5 or GPT-5.4 to GPT-5.6: pick one of three tiers, and the request keeps working as it is.

  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-5.6 comes as three tiers, each with a -pro ID: the same model, served with pro reasoning, which works longer before one final answer.

Model IDModel IDThe name a request calls a model by, lab/model, copied from Models.More on Models →For
openai/gpt-5.6-solthe most capable: complex reasoning, coding, agents
openai/gpt-5.6-sol-proSol, with pro reasoning
openai/gpt-5.6-terraa balanceBalanceWhat you can spend, in DA.More on Balance and charges →: everyday coding, reasoning and agents
openai/gpt-5.6-terra-proTerra, with pro reasoning
openai/gpt-5.6-lunafast, high-volume work: chat, classification
openai/gpt-5.6-luna-proLuna, with pro reasoning
FromTo
Model IDopenai/gpt-5.5 or openai/gpt-5.4one of the six above
Context1 050 000 tokens1 050 000 tokens
Longest answer128 000 tokens128 000 tokens
Effort levelsnone · low · medium · high · xhighthe same, and max

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

JSON
{
  "model": "openai/gpt-5.6-terra",
  "reasoning": {
    "effort": "medium"
  },
  "messages": [
    {
      "role": "user",
      "content": "Find the slow query in this log and explain it."
    }
  ]
}

2. Update the request

Nothing you send today is refused. Two additions are worth taking.

AdditionWhat it does
a -pro IDpro reasoning: for hard work where quality matters more than time and tokens
prompt_cache_options{"mode": "explicit", "ttl": "30m"} caches only the parts you mark with cache_control; it replaces prompt_cache_retention
JSON
{
  "model": "openai/gpt-5.6-sol",
  "prompt_cache_key": "souk-bot-faq",
  "prompt_cache_options": {
    "mode": "explicit",
    "ttl": "30m"
  },
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "The shop's full FAQ: …",
          "cache_control": {
            "type": "ephemeral"
          }
        },
        {
          "type": "text",
          "text": "Do you deliver to Constantine?"
        }
      ]
    }
  ]
}

How the prompt cache works, and what it saves: Caching.

3. Read the answer

ChangeWhere
writing to the cache is now charged, at the model's cache-write rateusage.prompt_tokens_details.cache_write_tokens, beside cached_tokens
answers run shorter by default than on GPT-5.5check whether a "keep it short" line in your prompt still helps
a pro answer counts all the reasoning behind itusage.completion_tokens_details.reasoning_tokens

Where cache writes cost more than the reads save, cache fewer parts: mark only the start that repeats. See Chat.

4. Set the effort

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

OpenAI gives medium as the default. From GPT-5.5 or GPT-5.4, keep the level you ran, then try one level lower: OpenAI reports GPT-5.6 reaching the same results with fewer tokens. max is new, above xhigh: compare the two before you pay for it.

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-5.6 comes as tiers, so the comparison is a sweep: one task to the old ID and to each tier, with two more runs for the two things this move adds, one effort level lower and a -pro ID. The script 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.

RunModelEffort
fromMODELEFFORT
solSOLEFFORT
sol, lowerSOLLOWER
sol-proSOL_PROEFFORT
terraTERRAEFFORT
lunaLUNAEFFORT
import os
from openai import OpenAI

API_KEY = os.environ["OPENDUNES_API_KEY"]
BASE_URL = "https://opendunes.com/v1"
MODEL = "anthropic/claude-sonnet-4.5"
SOL = "anthropic/claude-sonnet-4.5"
SOL_PRO = "anthropic/claude-sonnet-4.5"
TERRA = "anthropic/claude-sonnet-4.5"
LUNA = "anthropic/claude-sonnet-4.5"
EFFORT = "medium"
LOWER = "low"
MAX_TOKENS = 8000
TASK = (
    "Find the slow query in this log and explain it.\n\n"
    "12:01 SELECT * FROM orders WHERE status = 'late' took 4210 ms\n"
    "12:01 SELECT id FROM users WHERE id = 7 took 3 ms\n"
    "12:02 UPDATE stock SET count = count - 1 WHERE sku = 'B-112' took 12 ms"
)
RUNS = [
    ("from", MODEL, EFFORT),
    ("sol", SOL, EFFORT),
    ("sol, lower", SOL, LOWER),
    ("sol-pro", SOL_PRO, EFFORT),
    ("terra", TERRA, EFFORT),
    ("luna", LUNA, EFFORT),
]

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

for label, model, effort in RUNS:
    reply = client.chat.completions.create(
        model=model,
        max_tokens=MAX_TOKENS,
        messages=[{"role": "user", "content": TASK}],
        extra_body={"reasoning": {"effort": effort}},
    )
    usage = reply.usage
    details = usage.completion_tokens_details
    thought = details.reasoning_tokens if details else 0
    print(
        f"{label:11} {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-5.5 or openai/gpt-5.4, and the others to the IDs below.

SeatModel ID
SOLopenai/gpt-5.6-sol
SOL_PROopenai/gpt-5.6-sol-pro
TERRAopenai/gpt-5.6-terra
LUNAopenai/gpt-5.6-luna

Read sol against sol, lower first: OpenAI reports GPT-5.6 reaching the same results with fewer tokens, so the lower level is the cheaper one if your check agrees. Then read sol-pro against sol, in thought and in the charge: pro reasoning is worth its cost only where the answer is better.

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