GPT-5.6 Sol, Terra, and Luna

Model selection

Three routes, one 372K context limit

GPT-5.6 Sol

gpt-5.6-sol

Highest-capability route for complex coding, research, and long-horizon agentic work.

Default reasoning: low
Multi-agent metadata: v2
Ultra: Advertised · experimental

GPT-5.6 Terra

gpt-5.6-terra

Balanced route for general development, analysis, and agentic workloads.

Default reasoning: medium
Multi-agent metadata: v2
Ultra: Advertised · experimental

GPT-5.6 Luna

gpt-5.6-luna

Faster, economical route for routine coding and higher-throughput workloads.

Default reasoning: medium
Multi-agent metadata: v1
Ultra: Not advertised

Availability is controlled per endpoint and API key. Query GET /v1/models with the same key you will use for inference.

Context window

372,000 tokens, combined

The limit includes instructions, conversation history, tool definitions and results, input representation, reasoning, and generated output. Occasional metered totals above 372K can include accounting or protocol overhead; they do not increase the advertised model limit.

Hard product limit

372K combined input and output.

Normal large input

Keep input at or below about 350K.

Tools or long output

Target 330K–340K to preserve headroom.

Reasoning controls

Low through max is verified

low

Latency-sensitive transformations and straightforward edits.

medium

Balanced everyday analysis and development work.

high

Difficult implementation, debugging, and multi-step analysis.

xhigh

Complex architecture and long-horizon agent work.

max

Hardest tasks where quality matters more than latency.

Ultra is experimental. Sol and Terra advertise ultra, but bounded probes have not confirmed reliable production behavior. Luna does not advertise it.

Fast / Priority

Requestable, not guaranteed

Send service_tier: "priority" separately from reasoning_effort. Requests are accepted, but audit logs have often reported the actual upstream tier as default. Jatevo therefore does not currently guarantee accelerated delivery or a fixed speed multiplier.

Priority delivery depends on endpoint policy and upstream availability.

Agent capabilities

Streaming works; tool reliability remains caveated

Verified

Basic inference and streaming have successful live traffic.

Advertised metadata

Text and image input, search, WebSocket preference, and parallel tools.

Fresh direct forced-tool probes were not consistently conclusive. Validate tool calling on your target endpoint before depending on it for non-idempotent production actions.

API examples

OpenAI-compatible requests

Chat Completions with reasoning

curl https://api.jatevo.ai/v1/chat/completions \
  -H "Authorization: Bearer $JATEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "messages": [
      {"role": "user", "content": "Review this architecture and identify its three highest risks."}
    ],
    "reasoning_effort": "high",
    "max_tokens": 2000
  }'

Streaming

curl -N https://api.jatevo.ai/v1/chat/completions \
  -H "Authorization: Bearer $JATEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-terra",
    "messages": [
      {"role": "user", "content": "Create an implementation plan."}
    ],
    "reasoning_effort": "medium",
    "stream": true
  }'

Request Priority delivery

curl https://api.jatevo.ai/v1/chat/completions \
  -H "Authorization: Bearer $JATEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-luna",
    "messages": [
      {"role": "user", "content": "Summarize this incident report."}
    ],
    "reasoning_effort": "low",
    "service_tier": "priority",
    "max_tokens": 800
  }'

Operations

Context and timeout handling

context_length_exceeded

Compact old turns, remove unnecessary tool results, reduce attachments, and reserve output headroom.

Timeout or incomplete stream

Use streaming, realistic read timeouts, bounded exponential backoff, and preserve the request ID.

Do not blindly replay non-idempotent tool actions after a timeout. Confirm whether the action completed first.

Need client-specific reasoning configuration?

Open reasoning guide