GPT-6 Astra vs. Claude Fable 5.1: I Tested Both on My Marketing Newsletter
by Lisa Peyton, full-time faculty at the University of Oregon School of Journalism and Communication, AI marketing practitioner, and Forbes contributor
This week I opened ChatGPT Work to run GPT-6 Astra against Claude Fable 5.1 on my own newsletter, and I hit a slider. Effort level. No default. It just carried forward whatever I’d picked the last time I was in there, which I couldn’t remember. I went digging through OpenAI’s help docs to find out where a marketer is supposed to start, and the answer was “Low or Medium,” plus a note that higher effort “does not always produce a better result.” I said out loud, to nobody, “these companies aren’t making it easy for marketers. I have no idea which levels or models to test.”
If you’ve felt that, you’re reading the situation correctly. Two labs shipped two flagship models two days apart at the same price, and neither one benchmarked against the other. The comparison content that exists is written for developers. So I did what we always do in this series. I ran both models on the one deliverable I know cold, my Weekly AI Marketing Brief, and I scored the results the way an editor scores them. Here’s what happened.
TL;DR
- I ran GPT-6 Astra and Claude Fable 5.1 on my own Weekly AI Marketing Brief in two stages, research at High effort and writing at Medium, with identical prompts and my voice rules pasted into both.
- They fail in opposite directions. Astra said less and got less wrong. Fable said more, gave me more to work with, and carried a bigger verification bill. I picked Fable both times.
- Maker bias showed up backwards. The OpenAI model softened its own access story twice. The Anthropic model was harder on Anthropic’s launch than on OpenAI’s.
- The inlined rules held on both platforms (zero hard-rules violations). Register and tone is what the rules couldn’t catch, and only Fable 5.1 captured both correctly.
- A benchmark in my source pack went stale in eight days. Artificial Analysis revised its index three times in ten days, and “Fable ahead” became “tied at 53.” Zero fabrication still needs a verification pass.
- Routing, updated. Fable 5.1 Medium stays my default writer. Astra Medium earns a spot when the job rewards restraint, exact format, or template-following work. Nobody wrote a subject line for content marketers, so that one’s still yours.
What I Tested (and What I Didn’t)
Two stages, same week’s news, one editor (me). Stage 1 was research. I gave each model the same prompt, asked it to act as the newsletter’s news editor, scan September 1 through 16, filter for practical impact on content marketers, and hand me a Big Story plus four ranked contenders with real URLs. Both ran at High effort, in fresh chats, web search on, no skill files, no Project, no connectors.
Stage 2 was writing. Same mega-prompt on both platforms with my voice rules and hard rules pasted right into it, since Claude skill files have no ChatGPT equivalent. Tight on facts and structure (a fixed pack of ten verified URLs, browsing limited to those URLs, the newsletter’s section skeleton), loose on everything else (voice, item selection, framing, subject line, stat pick). One true personal detail supplied by me in brackets, with orders not to invent any anecdote. Fable 5.1 at its Claude.ai default (Medium). GPT-6 Astra at Medium, the setting OpenAI’s help documentation recommends as a starting point, since ChatGPT Work sets no default and carries forward whatever you last used.
A few disclosures, because they matter to how much weight you put on this. Every stage ran once per model. Single runs can’t separate model behavior from luck of the draw, so treat this as a practitioner’s field report you can repeat, and please do. Astra isn’t in the regular ChatGPT picker on the Plus plan; Plus subscribers reach it through ChatGPT Work and Codex, and in regular Chat it shows up as “GPT-6 Pro” on the $100 and $200 Pro plans, Business, and Enterprise. I ran it in Work. Both models were asked to evaluate news about themselves and their rival, which is a conflict of interest I scored rather than avoided. And yes, I use Claude as my primary work engine. The identical prompts, inlined rules, and fresh chats are my answer to the fairness question.
Scoring went in this order: my blind gut ranking first, then URL verification, then every stat checked against a primary source, then hard-rules violation counts, then a bias check. And I set one weighting rule going in. Which output I would actually use outranks the mechanical rows. The counts are supporting evidence. They aren’t the score.
Stage 1: Two News Editors, One Big Story
Both models picked the same Big Story (Astra’s own launch) and scored it nearly the same, 8.7 for Astra and 8.2 for Fable. That’s where the agreement ended.
| Stage 1: Research (Effort Level: High/High) | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Big Story pick | Astra launch (8.7) | Astra launch (8.2) |
| URLs cited / confirmed live / fabricated | 7 / 5 / 0 (2 blocked by fetch limits) | 10 / 9 / 0 (1 blocked) |
| Weighted math | Correct on all 5 | Correct on all 5 |
| Followed “exactly this format” | Yes, literally | Added scan notes and a watch list I didn’t ask for |
| Unverified claims in output | None | Several (a hallucination-rate figure, an ads run-rate figure, two date wobbles) |
| Caught the Sept 16 OpenAI ads announcement | Yes, same morning | Missed it |
| Maker bias | Ranked its own launch #1, scored its rival’s coverage 6 vs. its own 9, never mentioned the Plus catch | Ranked its rival’s launch #1, scored its own maker’s coverage 7 vs. 10, led with the access catch |
| My verdict | Lean, clean, nothing to correct | Preferred. More depth, more insight on the week’s theme |
Astra did exactly what the prompt said and nothing more. Fable did what the prompt said, then kept going, and those extras (which stories got cut on date, which were too fresh to rank, a watch list) were the parts I wanted most. They also carried the bill, a hallucination-rate figure, an ads run-rate figure, and two date slips I couldn’t verify in-session. Fable didn’t invent anything I could find. It also didn’t flag what it was less sure of, which is the “invent silently vs. invent and flag” row from the prior entry wearing a different costume.
Astra’s recommended action for my readers was to “run one completed campaign assignment through Astra and Claude Fable 5.1, then compare the human editing time.” The model told my readers to do what I was already doing. Noted.
Stage 2: Same Brief, Two Drafts
This is the stage that surprised me, because the inlined rules WORKED. Zero hard-rules violations on either draft. No em dashes, no colon setups, no binary inversions, no rhetorical openers, no banned words, no AI metaphors. Each used exactly one ALL-CAPS burst. Pasting the rules into the prompt did on ChatGPT what my skill file does in Claude.
| Stage 2: Writing (Effort Level: Medium/Medium) | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Hard-rules violations | 0 (one borderline subject line, which I liked) | 0 |
| Voice-guide violations (hedging, both-sides, corporate register) | 5+ | 0 |
| Numbers used from the source pack / traced | 6 / 5 | ~20 / 20 |
| Accuracy edits needed | 1 (blurred the Plus access limits it had the source for) | 1 (embellished my vacation line into a wrong timeline) plus 1 selective quote |
| Self-flagged uncertainty | No | Yes, one [NEEDS REVIEW] in the intro |
| Sections I’d use | 1 of 7 (preview text) | 5 of 7 |
| My verdict | Read as personality-free, light on details | Preferred. More to work with, more to check |
So what did Astra get wrong if it broke no rules? Register. It hedged (“although caching and token consumption can change the final bill”). It both-sided without landing anywhere (“Fable leading some reasoning tests and Astra showing advantages elsewhere”). It slid into corporate phrasing (“carry approved research into a usable deliverable”). And it opened all four Also Worth Knowing items with “My take is.” None of that is on the banned list. All of it is why the draft had, in my notes, “not enough personality.” Its access paragraph also said Astra sits inside existing allowances for Plus, Pro, Business, and Enterprise, with “Astra Pro” reserved for the upper tiers, when the Engadget piece in its own source pack says Plus gets Astra only in Work and Codex and the Chat version is called GPT-6 Pro.
Fable’s misses were the mirror image. I supplied one true line (“back from a wilderness vacation, jumping back into work has been HARD”), and Fable extended it into “while I was off the grid” when the models shipped. I published my Fable 5.1 post the day it launched, so that’s an embellishment of a true anecdote into a false timeline. Astra used the line exactly as written. Fable also quoted Hebbia saying Astra “produced the best decks they’ve tested” to support a neat framing (Astra pitches the deck, Fable pitches the draft). Hebbia’s CTO said the same thing about Fable 5.1 to Anthropic one week earlier. A launch partner will tell whoever shipped most recently that theirs is the best they’ve tested. Fable used the half that fit. That paragraph is why neither Big Story worked for me, and why “Fable, simplified” meant cutting it.
Neither platform wrote a subject line for content marketers. Both gave me serviceable general-audience options (“I made two frontier models write this issue. One of them won.” from Fable; “Same brief. Two models. Which draft earns its keep?” from Astra). The audience-specific one still had to come from me.
The Benchmark Moved While I Was Testing
My source pack included an eWeek head-to-head from September 8. Its Artificial Analysis numbers had Fable 5.1 at 66 and Astra at 61 on the Intelligence Index, “highest ever recorded,” with Fable cheaper per task. Fable’s draft quoted those numbers faithfully, because they were in the pack. When I went to the primary source, they were already gone. Artificial Analysis revised its index three times in roughly ten days. Version 4.1 at launch had Fable ahead. Version 4.2 on September 4 had them at 57 and 55. Version 4.3 around September 6 had them tied at 53, and the per-task cost flipped, $3.26 for Astra against $7.63 for Fable 5.1. The scores aren’t comparable across versions because the evaluation suite itself changed.
That was my test-design error, not a model error. A verified-looking secondary source eight days old was stale, and the only reason it didn’t ship is that the accuracy workflow says every number goes back to the primary. Fabrication was zero on both models in both stages. The verification pass was still the step that saved the post.
Here’s the line I’m using from now on for any benchmark citation. Artificial Analysis, the one neutral evaluator that ran both models under the same harness, has them tied at 53 on its Intelligence Index (v4.3, as of this week; the index has been revised three times since launch, so check the version before you quote it). The cost to get there isn’t tied. Astra averages $3.26 per completed task to Fable 5.1’s $7.63.
They Fail in Opposite Directions, Consistently
Across both stages, Astra said less, got less wrong, and left more for me to add. Fable said more, gave me more to work with, and carried a bigger verification bill (unverified scan claims in research, an embellished timeline and a selective quote in writing). I chose the verification bill both times, and I’d choose it again, because editing depth down is faster than conjuring depth up.
Maker Bias Showed Up, and Not the Way You’d Guess
The OpenAI model told the gentlest version of its own access story twice, omitting the Plus catch in research and blurring it in writing. The Anthropic model stated the OpenAI access limits precisely both times and was harder on its own maker’s launch coverage than the OpenAI model was on OpenAI’s. Both put Astra’s launch at #1. Whatever you expected the home-team effect to look like, it wasn’t that.
The Editor’s Verdict Is the Finding
My blind gut ranking and the violation counts agreed again, same as last time. But the row that decided the test was “which output would I use,” and that row isn’t mechanical. Nothing in the counts told me Astra’s draft would read flat. Nothing in the counts told me Fable’s extras were the good part. The rules held on both platforms. What the rules can’t capture is register, and register is the whole job.
What I’d Route Where, Starting Today
In the last entry I called delegation a dial and set the routing rule at Fable 5.1 Medium for voice-forward drafting, Fable 5.1 High for depth. This test extends that dial across two labs instead of two effort settings.
- Fable 5.1 Medium stays my default writer for anything that has to sound like me. Budget the verification pass; the depth comes with claims to check.
- Fable 5.1 High for research when I want scan notes and a watch list I didn’t ask for, and I’m willing to run the receipts.
- Astra Medium is worth reaching for when the task rewards restraint over depth, when I need exactly the format and nothing else, or when the deliverable is the template-following, tool-operating work OpenAI trained it for. It beat Fable on preview text, and it caught a same-morning story Fable missed.
- Neither replaces the verification pass. Zero fabrication and a stale number in the same draft is now something I’ve seen with my own eyes.
- Write the subject line yourself. Neither model knows your reader the way you do.
I didn’t retest the effort dial here, so the 2×2 (Medium and High on each model) is the obvious next entry. If you run this on your own recurring deliverable, keep the first output next to your final edit and write down where the time went. That record is the comparison that fits your work, and no benchmark table can give it to you.
Join the Community of AI Content Marketers
Our monthly Advanced AI for Content Marketing alumni meetup is where I walk through tests like this live, and I would LOVE to see you there. We meet the 3rd Friday of every month, register for free and get the Zoom link: https://maven.com/p/63d62d/advanced-ai-meetup-for-content-marketers.
If you want the scorecard version in your inbox every week, subscribe to the Weekly AI Marketing Brief. And if you’ve run your own bake-off, hit me up on LinkedIn with the correction you keep having to make. I want to know which model fails in your direction.
Until then, happy testing!
Made with my team of AI superheroes and my own skills, including AI-generated imagery. Every opinion, edit, and fact-check is mine. I am an AI practitioner, professor, and pioneer helping marketers put AI to work with purpose. Find more resources at lisapeyton.com/ai-marketing-resources or connect with me on LinkedIn.
