GPT-5.6 Sol, Terra, and Luna: which OpenAI model should your AI agents use?
Sol, Terra, or Luna? Pick GPT-5.6 by mission difficulty, volume, and cost per successful task — not by putting the strongest model everywhere.
9 min read
Claude Fable 5 and Claude Opus 5 sit at the top of Anthropic’s generally available lineup. The tempting shortcut is: Fable is the most powerful model, Opus is the cheaper one.
That framing is too thin to be useful.
Anthropic presents Fable 5 as its most capable generally available model, especially for demanding reasoning and agents that work for a long time. Opus 5, launched on 24 July 2026, is positioned as coming close to that frontier intelligence at about half the price, aimed at complex agentic coding and daily professional work.
The useful question is not “which model is stronger?”. It is: at what difficulty does Fable 5 actually pay for costing twice as much?
That question matters for agents. A run can issue many calls, hold a huge context, and use several tools. Doubling the model price quickly changes the cost of a mission.
Claude Opus 5
Default · Half the price
Agentic coding, professional work
Claude Fable 5
Escalation · Maximum capability
Long, open, high-stakes missions
The essential differences
Both models share a 1 million token context window and up to 128,000 output tokens. You will rarely pick Fable because it “fits more documents”.
| Criterion | Claude Opus 5 | Claude Fable 5 |
|---|---|---|
| Positioning | Agentic coding and daily professional work | Maximum intelligence, long-running agents |
| Context / output | 1M / 128k tokens | 1M / 128k tokens |
| Adaptive thinking | Available, adjustable | Always on |
| Latency | Moderate | Slower |
On standard API pricing, Fable currently costs exactly twice as much as Opus 5 for both input and output. That ratio is not a decision.
A model that costs twice as much can still be cheaper if it cuts errors, retries, bad decisions, human intervention, and mission length. Paying Fable for work Opus already finishes well is waste.
Claude Fable 5
Maximum capability. Long-running agents, open missions, especially hard reasoning. Adaptive thinking always on. Slower.
Claude Opus 5
High-end default. Agentic coding, professional work, everyday use. Adaptive thinking available and adjustable. Half the price.
Same context size
1 million input tokens and 128,000 output tokens on both sides. The gap is how the model uses that context, not how large it is.
Not Fable everywhere
The best model is not a permanent property of the agent. It can belong to the step, the risk, or the mission.
Fable 5: when maximum capability starts to matter
Fable is not meant to answer a question “a bit better”. It becomes interesting when the system must stay coherent across a long sequence of actions.
Anthropic highlights it for software engineering, knowledge work, vision, scientific research, and agents that work for a long time. The harder and longer the task, the more the gap with other Anthropic models tends to widen.
Take this mission: “Analyse the architecture, find why performance collapses under load, propose a fix, implement it, and verify the result.”
An agent may have to read docs, explore the repo, form hypotheses, run commands, fail, correct, and retest. The first reasoning pass is only a slice of the problem. The model also has to keep the goal, interpret intermediate results, and avoid stacking mistakes.
That is the long-horizon work Fable 5 is positioned for.
Opus 5 is not a watered-down Fable
Calling Opus 5 “Fable, but worse” misses the point. Anthropic describes it as a major step for the Opus line, close to Fable’s frontier intelligence at half the price.
Picture a developer agent used all day: bugs, tests, reviews, refactors, docs, migrations, technical proposals. These are hard tasks. They do not all need the ceiling.
If Opus already completes them reliably, putting Fable on every run doubles the bill for little gain. That is why Anthropic now positions Opus 5 as the default for complex agentic coding and professional work.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 |
| Claude Fable 5 | $10.00 | $50.00 |
Rates change; prompt caching still lowers input cost, but the ratio between the two models remains strongly in Opus’s favour.
Price changes the decision — not always the way you expect
A chatbot conversation may be a handful of calls. An agentic mission can be many.
A deliberately simple example: 500,000 input tokens, 100,000 output tokens, no cache.
- Opus 5: $2.50 + $2.50 = $5
- Fable 5: $5 + $5 = $10
One mission costs twice as much. Multiply by dozens of agents, several missions a day, pipelines, and retries: the model becomes a product decision.
The opposite trap is real. If Opus fails, retries, patches, then needs a review, while Fable lands on the first pass, the theoretical gap disappears.
The useful metric is cost per successful mission, not dollars per million tokens. Ideally you also track total cost, turns, time, retries, tool errors, human interventions, and final quality.
Fable or Opus for coding?
For everyday or even advanced coding, Opus 5 is the better starting point: a feature, a bug, tests, reading a codebase, reviewing a PR, a migration, a refactor.
Fable becomes more interesting when the mission outgrows daily development. For example: reviewing a distributed system under load, or working autonomously until a difficult test suite is green.
The difference is depth and duration — the same idea as choosing between automation, workflow, and agent.
Autonomy is not a reason to pick Fable
“Autonomous agent” is not enough. An agent that sorts information and makes bounded decisions probably does not need Fable. An agent that must explore, form hypotheses, use several tools, and reverse earlier decisions might.
Ask: how hard and how long is the trajectory? Not merely: is this agent autonomous?
| Agent type | Starting point |
|---|---|
| Triage | Lighter than either |
| Docs agent | Opus 5 |
| Reviewer | Opus 5 |
| Developer agent | Opus 5 |
| Complex research | Opus 5, Fable if needed |
| Software architect | Opus 5, then Fable |
| Hard investigation | Fable 5 |
| Long-horizon agent | Fable 5 |
| Critical open mission | Fable 5 |
| High-volume daily agent | Opus 5 |
This table is not a universal rule. It illustrates a principle: Opus should often be the starting point. Fable should be an escalation.
Adaptive thinking and latency
Both models use adaptive thinking. For Fable, Anthropic says it is always on. Opus supports it too, but the intelligence / cost / speed trade-off can be adjusted.
Opus 5
Moderate latency, everyday use
Fable 5
Slower, capability first
Fable is currently classed as slower; Opus as moderate latency. For a fast or repeated interaction, Opus can be the better experience even when the Fable budget exists.
Route instead of choosing once
You do not have to freeze Fable or Opus for the whole system.
One model
Escalation
Fable on every run
Opus first, Fable if needed
Simple to configure
Maximum capability where it pays
Top-tier latency and cost everywhere
Fable review only on the critical path
Examples: Opus executes, Fable reviews a high-stakes mission. Or Fable sets a strategic plan, Opus runs the sub-tasks.
A developer → reviewer → fixer pipeline does not need Fable on every agent. Opus for implementation and fixes, Fable on review when the change is hard to reverse, for instance.
The model can belong to the step, the risk, or the mission. See also fusing models in multi-agent pipelines.
Risk matters as much as difficulty
Some tasks have an especially high cost of error: data migrations, infrastructure, critical architecture, financial operations, security, changes that are hard to undo. Paying for Fable can then be rational.
That still does not replace tests, permissions, validations, approval gates, backups, or deterministic rules. A smarter model does not remove the need for governance — the same point as managing agents without losing control.
The lineup is larger than these two
Anthropic also positions Sonnet 5 (speed / intelligence balance) and Haiku 4.5 (maximum speed). Opus can already be too much for some work. The real choice is not only Fable or Opus: it is how much intelligence this mission needs. For the OpenAI parallel: Sol, Terra, and Luna. For a cross-vendor coding comparison: GPT-5.5 vs Claude vs Gemini vs Cursor.
Stellary
Stellary already structures agents around missions, tools, autonomy levels, and pipeline orchestration. Distinguishing Opus and Fable fits that model: a daily development agent does not have to pay for Fable by default; a particularly open investigation can use it.
That does not mean a fixed model per agent. A reviewer can stay on Opus most of the time and move to Fable on a critical review. The useful future is each mission using the level it needs.
Starting grid
Choose Opus 5
High-end daily use, lots of coding, complex but familiar missions, real cost pressure, a latency / capability compromise, many runs.
Choose Fable 5
Unusually hard mission, long horizon, many steps, very open problem, error cost larger than the model premium, Anthropic’s maximum generally available capability.
Route
Opus by default. Measure outcome, errors, validations. Escalate to Fable on failure, high criticality, or a test suite that will not go green.
Governance
Escalation should not rest only on the model’s self-assessment. Tests, rules, scores, criticality, tool failure, human review.
FAQ
What is the difference between Claude Fable 5 and Claude Opus 5?
Fable 5 is Anthropic’s most capable generally available model, especially for long-running agents. Opus 5 targets complex agentic coding and professional work at a lower cost.
Is Fable 5 better than Opus 5?
Fable 5 is positioned at maximum capability. That does not make it the best economic or operational choice for every task.
How much does Claude Fable 5 cost?
Current standard API pricing is $10 per million input tokens and $50 per million output tokens.
How much does Claude Opus 5 cost?
Current standard API pricing is $5 per million input tokens and $25 per million output tokens.
Is Fable 5 twice as expensive as Opus 5?
Yes, on standard input and output token rates. Prompt caching and the Batch API lower the amounts, but the ratio stays in the same range.
Which Claude model should I use for coding?
Opus 5 is Anthropic’s recommended starting point for complex agentic coding. Fable becomes interesting for the longest, hardest missions.
Do Fable 5 and Opus 5 both have a 1 million token context window?
Yes. Both currently have a 1 million token context window and can produce up to 128,000 output tokens.
Which model should an autonomous agent use?
For a daily agent or a complex but familiar mission, Opus 5 is a strong starting point. For a very open, long, or high-stakes mission, Fable 5 can justify the extra cost.
