This page compares two models on one job: turning a customer interview transcript into a published case study. It covers finding the real before and after, keeping selected quotes verbatim, holding a fixed template across a batch, and cost per story.
Jul 30, 2026 · 11 min read
GPT-5.6 Terra is the production choice for a steady flow of customer stories. Claude Fable 5 is the choice for the flagship story, and its citation mode is the better tool for building an auditable quote ledger.
The cost gap is the certain part. With the same token assumptions, one story costs about $0.30 on Fable 5 and about $0.06 on Terra, and both vendors offer a 50 percent asynchronous batch discount that brings the example to about $0.15 and about $0.031, 5. Keep that in proportion: human editing costs far more than inference, so a draft that saves one substantial rewrite can outweigh the model bill for a large batch of stories.
The quality evidence points the other way and it is indirect. Fable leads an independent creative-writing tracker, and a blind test of the GPT-5.6 tiers put Terra's prose behind its own siblings and called it more utilitarian8, 9. Fiction is not a case study, which has to stay inside the transcript, so that supports a hypothesis about narrative instinct rather than proving anything about customer stories. The professional benchmarks disagree outright, and the page says so rather than picking the flattering one.
A polished sentence the customer never said is the one error that can cost you a reference. Check every quote against the transcript by substring.
At roughly a fifth of the cost per story, and half that again through batch, the cheaper model makes a proper length check affordable on every draft.
For the story that goes on the homepage, narrative instinct is worth paying for. Judge it on editing minutes saved rather than on the token bill.
Provenance and strict schemas do not always combine in one request. Extract with citations first, then draft to the schema and carry the ledger across.
This page compares the two models through their API in one neutral setup, on the parts of a case study that decide whether it can be published.
Those parts are reading the transcript for the real before and after, choosing quotes and keeping them exactly as spoken, separating what the customer said from what the writer inferred, holding a fixed template and length across a batch, and the cost per story once the programme is running.
Content management systems, story templates as design artefacts and approval workflows are left out on purpose. They belong to the app around the model, so the same model behaves differently in a chat product, through the API, or inside a workspace. Judging them here would compare wrappers rather than the writing.
The published facts that affect story production. One interview is nowhere near either context limit, so the price rows and the citation behaviour are what decide this.
Figures from Anthropic and OpenAI documentation, checked July 30, 2026. Worked example: 20,000 input and 2,000 output tokens costs about $0.30 on Fable 5 and about $0.06 on Terra, or about $0.15 and about $0.03 through batch. A two-pass workflow totalling 45,000 input and 6,000 output tokens costs about $0.75 and about $0.16. Citation mode and strict structured output cannot be enabled in the same request, which is why the recommended workflow is two passes.
Read the evidence column closely. The narrative rows rest on fiction tests, the professional rows contradict each other, and the one clear capability difference is about quote provenance.
Better-choice calls map to what each source measured. The narrative evidence is creative writing rather than case-study work, the two professional figures come from the same vendor table and disagree, and no public evaluation tests either model on turning an interview into a published story.
Score both models on interviews whose finished case study you already approved, because the published version is your answer key. Grade quote accuracy mechanically and narrative by eye, in that order.
Take three to five completed interviews and make sure one has missing metrics and one is rambling or self-contradictory. Those two decide the comparison, because a clean transcript makes both models look good.
Identical prompt, transcript, template, length range and source package on both sides, at comparable effort settings. Do not edit, regenerate or add hints before scoring.
Run evidence extraction and drafting as separate passes on each model, since citation mode and strict structured output cannot share a request on one side and the comparison has to be like for like. Test the API surface the team will deploy.
Verify every quoted string appears verbatim in the transcript, then score whether the real before and after was found, whether facts came only from the source, whether quotation and paraphrase stayed apart, whether the template held, and how many minutes of editing each draft needed. Remove the model names and use two reviewers.
The most useful thing in this table is the contradiction. Two exact-version professional results published side by side put the two models in opposite orders.
No public benchmark tests either model on turning an interview into a published case study with attributed quotes. That is the gap this page cannot close, and it is why the method section grades against stories your team already approved.
Both models get the same rule: build the evidence ledger before any prose, and never let the drafting step touch the words inside quotation marks. What differs is how much of the contract you write out.
For Claude Fable 5, use a concise handoff with the intent, the boundaries and a definition of done, since its guide recommends brief steering rather than restating the same behaviour several times4. Ask for the evidence ledger with exact quotes and source locations first, then the story. Tell it to label an unsupported but plausible connection as an editorial gap rather than filling it, and keep the effort setting moderate for a routine story so it does not explore narrative options you did not ask for.
For GPT-5.6 Terra, make the production contract explicit: the schema fields, the length range, the success criteria and a validation block. OpenAI's guidance is a lean prompt with the effort set deliberately, and it warns that a broad instruction to be concise can leave the draft too short6. So specify what belongs in every section rather than asking for brevity, and render the JSON into your publishing template outside the model.
A Claude Fable 5 prompt: ledger first then the story
Turn this interview into a customer case study of
900 to 1,050 words using the supplied template.
First build an evidence ledger: facts and exact quotes,
each with its source location in the transcript.
Then write the story. Never alter the words inside
quotation marks.
Where a connection is plausible but unsupported, label it
an editorial gap instead of filling it.
Lead with the customer's problem, show the decision and the
implementation, and end with evidenced results.A GPT-5.6 Terra prompt: the production contract written out
Produce JSON matching the supplied case-study schema:
headline
summary
challenge
decision
implementation
results
exact_quotes
unsupported_claims
The rendered story must be 900 to 1,050 words.
Quotes must be exact substrings of the transcript.
Do not infer a metric that is not stated.
Return a validation block with the word count, any missing
field, and any quote that failed exact matching.The shared failure is the serious one, because it puts words in a named customer's mouth and reads perfectly. The model-specific failures are about cost and flatness.
One question first. Is this a production story in a steady programme or a flagship story someone will read closely? Then follow the branch that matches most of your work.
A starting point, not a rule. Score both against case studies you have already published.
If this is a steady programme of production stories, make GPT-5.6 Terra the default and use batch pricing, since the cost per story is roughly a fifth and the retries are cheap enough to enforce a length range properly5. Supply one approved story as a style example, because compliance without causal movement is its likely failure.
If the story is a flagship that someone will read closely, or the interview is rambling and self-contradictory, use Claude Fable 5 and keep the effort setting moderate so a routine story does not turn into an elaborate one4. Where you need an auditable quote ledger, its citation mode is the real capability difference, and it needs its own pass because it cannot run alongside strict structured output2.
Whatever you choose, the quote check belongs outside the model. Verify every quoted string against the transcript with a substring check, and treat a paraphrase inside quotation marks as a defect rather than a style choice. The benchmarks disagree about which model reasons better on professional work, so let your own editing time decide instead7.
One limit applies to Playgram rather than the models. A customer-marketing team that needs stories drafted straight into its content system, versioned there and pushed through an approval queue needs that system. Playgram is a chat workspace, so the ledger and the draft come back in the conversation and the publishing happens where it always did.
GPT-5.6 Terra is the safer default for volume and Claude Fable 5 for the stories that matter most. The most defensible workflow extracts the quotes with provenance, drafts to a schema, and checks every quotation outside the model.
The evidence is genuinely mixed here rather than merely thin. Two exact-version professional results published in the same vendor table put the models in opposite orders, the narrative signals come from fiction tests, and nothing public measures turning an interview into a published story with attributed quotes7, 8, 9. The one clear capability difference is quote provenance, and the one clear economic difference is cost per story.
The safest final step is to test the shape of your own interviews, not a generic prompt from the internet. A fair test needs the same setup for both models: the same transcript and template, the same prompt and the same place to run them, so the result reflects the models and not the tool around them. In practice that is harder than it sounds, since most teams end up running one model in one app and the other in a different one, on two separate subscriptions, which tilts the comparison before the first draft comes back. The cleaner the setup, the more the difference you see is really Claude Fable 5 vs GPT-5.6 Terra, and not just which one happened to be easier to reach that day.
Upgrade as needed, and only pay for what you actually use
Save ~17% with the annual plan
Pro
Perfect for small and medium teams
Unlimited users & infinite memory
Multi-LLM chats
Unlimited use of DeepSeek V4 Flash
US & EU data residency
Enterprise
Get in touch
For organizations with advanced needs
Unlimited users & SSO
Priority Support
Unlimited use of DeepSeek V4 Flash
US & EU data residency
30-days money back guarantee