Claude Fable 5 vs Opus 5: which model should your AI agents use?
Fable 5 or Opus 5? Choose by mission difficulty and cost per successful task — not by putting Anthropic’s strongest model everywhere.
10 min read
With GPT-5.6, OpenAI is no longer shipping a single “more powerful” model. The family is three clearly positioned models: Sol (hard work), Terra (capability versus cost), Luna (volume and price).
An agent reviewing architecture does not need the same compute as an agent labelling tickets. A workflow run thousands of times a day can lose its economics if every step uses the flagship.
The real question: when should you use Sol, Terra, or Luna?
GPT-5.6 Luna
Speed · High volume
Triage, classification, extraction
GPT-5.6 Terra
Balance · Everyday work
Review, synthesis, AI workflows
GPT-5.6 Sol
Capability · Hard missions
Debug, research, open-ended agents
The essential differences
OpenAI positions Sol as the flagship for complex reasoning, coding, and demanding professional work. Terra is the balanced choice for everyday workloads. Luna is the fastest and cheapest, aimed at high volume.
That does not make Luna “bad”, Terra “average”, and Sol “good”. They optimize different constraints.
GPT-5.6 Sol
Maximum capability. Hard reasoning, difficult coding, research, open-ended agents. Announced 1.05 million token context window, up to 128,000 output tokens.
GPT-5.6 Terra
Balance. Analysis, review, daily tasks, AI workflows. OpenAI places it roughly at the “mini” level of previous GPT generations.
GPT-5.6 Luna
Speed and cost. Classification, triage, extraction, large pipelines. First-pass work, then escalate the hard cases.
Not one model everywhere
The best model is no longer the one to use everywhere. A reviewer does not need the same compute as a debugging agent.
Sol: when quality comes first
Spend more compute when a mistake or shallow reasoning costs more than the model call.
Natural fits: complex codebases, hard bugs, architecture, multi-step research, large document sets, several tools, an autonomous agent on an open mission, a demanding review.
Example: “Find the regression, propose a fix, change the project, and confirm tests pass.” The number of steps is unpredictable. Saving a few cents can be a false economy if a lighter model multiplies retries.
Terra: often the most interesting model for agents
Terra is for workloads that must balance intelligence and cost. On 30 July 2026, OpenAI cut its API price to $2 per million input tokens, $12 output, $0.20 cached input.
It is often enough for tickets, content review, structured generation, synthesis, docs, human-assisted agents, transforms, ordinary code, and multi-step AI workflows.
A bounded reviewer (receive, check, find issues, suggest fixes) needs intelligence inside a defined perimeter. Terra is often the better economic choice than Sol.
Luna: do not underestimate it
On 30 July 2026 OpenAI cut Luna’s API price by 80%: $0.20 per million input, $1.20 output, $0.02 cached. Sol stays at $5 / $30. The gap is huge.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 |
| GPT-5.6 Terra | $2.00 | $12.00 |
| GPT-5.6 Luna | $0.20 | $1.20 |
One hundred thousand pieces of feedback to classify is not where you want Sol. Luna does the first pass; Sol only steps in when the case requires it.
Rates above are standard API pricing (short context). They change: check the official page before you commit spend.
Stop picking a single model
A system has no obligation to use the same model for every task.
One model
A routed family
Sol on every step
Luna / Terra / Sol by mission
Simple to configure
Cheaper at volume
Uniform quality
Quality where it matters
Flagship latency everywhere
Fast replies on triage
Example team: triage → Luna, reviewer → Terra, developer → Sol. Cheaper, faster, and still high quality if routing follows real difficulty.
| Agent type | Model | Why |
|---|---|---|
| Triage agent | Luna | Speed and cost |
| Classification | Luna | Structured task |
| Extraction | Luna | Little reasoning needed |
| Docs assistant | Terra | Solid compromise |
| Writing agent | Terra | Enough quality |
| Reviewer | Terra ou Sol | Depends on difficulty |
| Everyday coding | Terra | Balance |
| Hard debug | Sol | Stronger reasoning |
| Multi-step research | Sol | High ambiguity |
| Open-ended agent | Sol | Planning and adaptation |
The table is a starting point, not a law. Prompt, tools, context, input size, and the quality bar matter as much as the agent’s role.
Must an autonomous agent always use Sol?
No. Separate the agent type from the difficulty of its missions.
“Check each new ticket, pick the team, apply the label”: the agent decides, the task is bounded. Luna or Terra can be enough.
“Analyse this codebase, find why performance dropped, and propose a plan”: even a single mission can justify Sol.
The metric is not “is this agent autonomous?” — it is how much intelligence does this mission actually need? Same principle as choosing automation, workflow, or agent.
Several models in one workflow
Feedback → Luna (classify) → Terra (analyse) → simple cases done / hard cases → Sol.
Development: Terra (first implementation) → tests → Luna (summarise) → Sol (hard review).
Or: Sol (plan) → Terra (subtasks) → Sol (final review).
There is no perfect combination. It depends on the workflow — see also model fusion in multi-agent pipelines.
Do not read cost in isolation
A cheaper model can cost more if it produces more errors, retries, tokens, tool calls, or human fixes. Sol everywhere wastes budget on trivial work.
Useful metric: cost per correctly finished task. One Sol run at $0.20 with no correction can beat five Luna runs at $0.02 plus a human intervention.
The question is not “which model is cheapest?” It is “which is the cheapest model that reaches the quality this task needs?”
Speed
Luna
Fastest in the family
Terra
In the middle
Sol
Capability first, Fast mode in the API
OpenAI announced Fast mode for Sol in the API on 30 July, when latency matters. For a user click, latency is critical. For an agent mission that already lasts minutes, a few extra seconds matter less than the final quality.
ChatGPT, Codex, and the API
The three models are not exposed the same way everywhere. The OpenAI API offers Sol, Terra, and Luna. Codex and ChatGPT Work expose them on eligible plans. In standard ChatGPT conversations, Terra and Luna are not always selectable as distinct models.
A recommendation system must consider what is actually available (product, plan, channel, provider) — not only the theoretically best model. For a cross-vendor coding comparison: GPT-5.5 vs Claude vs Gemini vs Cursor. For Anthropic’s top tier: Fable 5 vs Opus 5.
Why this matters for agents
An agentic run chains analysis, tools, and retries. A small per-call gap becomes huge across thousands of missions. Hence routing: simple → Luna, medium → Terra, hard → Sol, with escalation if a level fails.
Routing can use complexity, budget, latency, role, context size, tools, and success history.
Stellary
In Stellary, agents are workspace members: role, tools, autonomy, missions. The platform tracks activity and cost. Distinguishing Sol, Terra, and Luna is a natural fit — a triage agent should not pay Sol by default.
That does not mean a fixed model per agent. A reviewer can stay on Terra 90% of the time and switch to Sol for an unusually hard review. The interesting future: each mission uses the level it needs. For control: manage AI agents without losing control.
Starting grid
Choose Luna
Short task, fairly clear rules, high volume, cost-critical, latency-sensitive, errors easy to detect or escalate.
Choose Terra
Real analysis, Sol would be oversized, regular use, quality/cost compromise, several AI calls in the workflow.
Choose Sol
Hard problem, costly mistakes, reasoning at the centre, exploration, several tools, complex coding or research, open mission.
Route
Volume on Luna, daily work on Terra, hard problems on Sol, escalate if needed. A single model is still fine on a simple system.
FAQ
What is the difference between GPT-5.6 Sol, Terra, and Luna?
Sol is the flagship for the hardest work. Terra balances capability and cost. Luna prioritizes speed, volume, and low cost.
Which GPT-5.6 model should you use for coding?
Sol is OpenAI’s recommended starting point for complex coding. Terra is often better for everyday or more repetitive development.
Which OpenAI model should an autonomous agent use?
Sol is usually the best starting point for genuinely open-ended missions. A simpler agent can still run well on Terra or Luna.
Is GPT-5.6 Luna suitable for AI agents?
Yes, especially for triage, classification, extraction, and other simple, high-frequency workloads.
Is GPT-5.6 Terra weaker than Sol?
Terra is a different compromise. Sol pushes capability further; Terra aims for strong results at a lower cost.
How much does GPT-5.6 Sol cost?
At publication, standard API pricing is $5 per million input tokens and $30 per million output tokens, excluding Fast mode and very long context.
How much do GPT-5.6 Terra and Luna cost?
Since 30 July 2026, Terra is $2 per million input and $12 output. Luna is $0.20 input and $1.20 output.
Can you use Sol, Terra, and Luna in the OpenAI API?
Yes. All three are available in the OpenAI API. Availability in ChatGPT and Codex depends on the plan.
