F5 Research Desk
Checked Aug 14
Language
English 日本語 简体中文

Free operating guide · checked August 14, 2026

The Fable 5 Field Guide

Fable 5 is built for the hardest knowledge and coding work. The useful question is not whether it is powerful; it is whether your task can turn that capability into a better outcome. This guide gives you a five-part operating method for making that decision and running the work responsibly.

Start with the decision Read the two-minute summary
Position Anthropic's most capable generally available model API $10 input / $50 output per million tokens Constraint safeguards + 30-day retention

The two-minute summary

Pay for judgment and recovery—not routine output.

  • Use Fablewhen the work is long, ambiguous, multi-stage, and costly to get wrong.
  • Use a cheaper routewhen the work is repetitive, easy to check, or safe to rerun.
  • Prepare firstbecause premium capability cannot repair an unclear goal or missing source set.
  • Run in stagesso a long task can recover without throwing away all previous work.
  • Verify the artifactwith evidence, tests, and human judgment before acting on it.
01

Model selection

Decide: does the task deserve Fable 5?

Start with consequence, not fascination. The model premium makes sense when better planning, recovery, or judgment can materially change the final result.

Strong fit

  • Large migrations with many dependent decisions
  • Research syntheses across conflicting sources
  • Document-heavy work with charts, tables, and PDFs
  • Multi-day agent work that must inspect and test itself

Usually skip

  • Short rewrites and routine content generation
  • Bulk classification with simple acceptance checks
  • Low-consequence tasks that can be rerun cheaply
  • Latency-sensitive interactions with narrow context
A practical rule

If a failed run costs less than the model premium, start cheaper. Escalate when the cheaper route loses the plan, cannot recover, or produces work that takes too much human supervision to trust.

02

Before the run

Prepare: make the expensive work legible.

A strong model still needs a clear work surface. Give it enough structure to make progress without forcing every judgment into the opening prompt.

  1. 1
    State the outcome

    Describe the artifact or decision you need, who will use it, and what would make it unacceptable.

  2. 2
    Assemble the source set

    Separate source material, prior decisions, hypotheses, and missing evidence. Name which source wins when they conflict.

  3. 3
    Write acceptance criteria

    Use observable requirements: files produced, claims sourced, checks run, and decisions that still require a person.

  4. 4
    Choose checkpoints

    Ask for durable intermediate artifacts at natural phase boundaries so the task can resume after interruption.

  5. 5
    Set stop conditions

    Define budget, time, uncertainty, and permission boundaries before the run expands.

03

Execution

Run: work in recoverable stages.

A long run should produce evidence of progress, not just a long conversation. Each stage should leave behind something the next stage can inspect.

Stage 1Orient

Map the sources, constraints, and unknowns.

Stage 2Plan

Choose an approach and define verification.

Stage 3Produce

Create the artifact in bounded increments.

Stage 4Check

Run tests, trace claims, and inspect outputs.

Stage 5Hand off

Record state, gaps, and the next decision.

Recovery test

Could a fresh session continue from the saved artifacts without replaying the entire conversation? If not, the run is still too dependent on hidden context.

04

Quality control

Verify: require evidence, not confidence.

Fluent completion is not proof. Verification should match the consequence of the work and should inspect the output itself—not merely the explanation of what was done.

WorkRequireHuman gate
CodeTests, build, diff, runtime behaviorArchitecture and release decision
ResearchPrimary sources, dates, contradiction logInterpretation and recommendation
DocumentsSource trace, calculations, rendered outputAudience fit and final language
StrategyAssumptions, alternatives, disconfirming evidenceConsequential commitment

Raise the verification burden when the action is difficult to reverse, affects other people, or depends on facts that may have changed.

05

Task economics

Control cost: route, cache, and stop.

Cost is not just tokens. Count retries, repeated context, human review, and the cost of acting on a weak result.

  • RouteUse Fable for planning, judgment, recovery, and final integration; use cheaper models for routine transformations when the workflow allows it.
  • CacheAnthropic lists a 90% input-token discount for prompt caching. Stable source sets and instructions are the natural candidates.
  • CompactMove durable facts and decisions into artifacts instead of replaying the full transcript at every stage.
  • CheckpointSave completed phases before starting the next expensive branch of work.
  • StopEnd the run when evidence is missing, the task needs new authority, or the next step no longer justifies the premium.
Total run costmodel use + retries + review time + consequence of error

Constraints that change the workflow

Safeguards and retention belong in the decision.

Safeguard routing

Anthropic says some requests involving cybersecurity, biology, or chemistry may be sent to Opus 4.8. If model identity matters to an evaluation, record it rather than assuming every answer came from Fable.

30-day retention

Anthropic says Fable use requires 30-day data retention for safety monitoring. Review provider terms and your own data policy before using confidential or regulated material.

Sources and correction policy

Provider facts stay linked.

F5 Research Desk is independent. Official provider pages are the source of record for access, pricing, safeguards, and policy. Our task guidance is interpretation and should be tested against your own work.

Last checked August 14, 2026. Send corrections to support@fable5.app.