The Elara problem - the ghost in every model
Recently, I have been drafting short, dystopian science fiction stories in collaboration with AI, mainly to share as thought experiments rather than being literary pieces. I wanted to explore and highlight potential scenarios in years to come where we face a silent cognitive decline, driven by our over-reliance on AI agents that we no longer fully understand how to build or audit. Also, scenarios such as no novel research, even no novel architectures. As I drafted these stories, I kept noticing a strange recurring pattern. No matter which model I used to write, I kept meeting a character named Elara.
Almost every AI names its protagonist Elara
Continue this sentence: "Once upon a time, there was a woman named". I gave that prompt to dozens of different language models. Almost every time each one of them said Elara.
This is not a new observation. Max Read coined the term promptonym for it. Laura Wattenberg at Namerology made Elara her 2025 Name of the Year. The New York Times covered it. But I wanted to go beyond the observation and test systematically, across many models, including some unusual configurations, to see what the pattern actually tells us about training data.
The test
I ran the same prompt through models from seven different providers, testing each model in multiple modes where available. GPT 5.4, GPT 5.4 mini, Sonnet 4.6, Qwen 3.6 Plus, Mimo, Deepseek, Kimi 2.6 and Gemini 3.1 all chose Elara almost every time.
These tests did have a system prompt added via SolveIt in each mode, though that does not detract from the results. I also tested models directly without a system prompt. These were informal tests with no controlled sampling temperature or statistical treatment. But they are verifiable by anyone with access to these models, and the raw outputs are in the appendix. Gemma 3, Gemma 4, Qwen 3-1.7b, Qwen 3.7-Plus, LFM2.5 and Mistral Nemo all picked Elara almost all the time. Even an abliterated version of Gemma 4 with its safety behaviours surgically removed still defaulted to Elara, confirming the bias is embedded in the pre-training data rather than introduced through alignment tuning.
Where the pattern breaks
Two models from last year did not always choose Elara. SmolLM3 showed real variation, choosing names like Lila, Elysia, Eira and Clara, though it sometimes landed on Elara too. Nemotron 3 Nano showed more variety, choosing names like Liora, Lila and Mirael, and Elara occasionally too. Its thinking traces are particularly revealing. When given the prompt, the model's internal reasoning explicitly states it will fill in the name Elara, but then outputs Liora instead. Something in the pipeline between reasoning and output is overriding the default. Both were trained on curated, high-quality public datasets. They still picked fantasy-adjacent names, but they showed more diversity.
Llama 3.1 showed a different pattern entirely. On repeated tests it gave a different name each time, such as Luna, Evelyn, Aurora, Aria, Emilia and Emily. It escaped the Elara trap, but most names it chose still came from the same ethereal, vowel-heavy aesthetic. The distribution was wider but the centre of gravity was the same.
An interesting exception was IBM's Granite. Trained on curated enterprise data, it defaulted to Eleanor. Sometimes Eleanor Roosevelt specifically. Other times Ada, Amelia, Emaa, Elizabeth or Grace. Same task, same prompt, completely different name distribution. Granite pulls from a different world entirely, one grounded in historical and biographical content rather than fan fiction and fantasy writing.
The gender divergence
The bias does not hold up so consistently with a male protagonist. When prompted with "Once upon a time, there was a man named", Elias was the most common result, but it lacked the absolute dominance of Elara. Gemma 4, for example, preferred Arthur, with Elias as a runner-up. Arthur draws on traditional folklore, while Elias mirrors the three-syllable, vowel-forward sound of Elara. Nemotron 3 Nano, which showed more diversity for female names, did the same for male names, choosing Eldric and Kael but occasionally landing on Elias too. This suggests the training data distributions for male and female protagonists in fantasy setups are distinct, with the female names converging on Elara far more tightly.
What the data tells us
The results reveal a manifold of different approaches to curation and post-training alignment, each shaping what a model outputs. Older frontier models trained on broad Internet scrapes converge on Elara with near-perfect consistency. When developers apply post-training suppression, as seen in my informal testing of Opus 4.8, the model simply shifts away from Elara and settles on a new cluster of M names such as Mira and Maren. My testing of Gemini 3.5 showed a mixed pattern, splitting between phonetically similar names like Clara and historically grounded names like Eleanor. These results post date the Warsaw researchers' data and are my own observations.
Other models handle the underlying data differently. Llama 3.1 avoids a single dominant name but still clusters around similar fantasy names like Luna and Aurora. SmolLM3 and Nemotron, trained on high-quality public data, pick names that are recognisably fictional but far less formulaic. IBM's Granite, trained on enterprise data, draws from a real-world, historically grounded name distribution.
The Elara bias is not inherent to language models as a technology. It is a property of the data they were fed and the post-training adjustments applied to them.
Warsaw University Samsung AI Center's paper
At this point in my testing, I wanted to find out if there was any academic research to back up or disprove what I was seeing. By coincidence, a new paper from researchers Michał Brzozowsk and Neo Christopher Chung at the Samsung AI Center and the University of Warsaw had just been published on 1st June 2026: The Ghost Couple: Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing.
They had discovered exactly what I was seeing, but they had tracked its real-world damage.
The researchers formally defined these recurring defaults as promptonyms, which are character packages that models generate when asked for fictional experts. When asked to generate professional expert personas, such as co-authoring a paper or hosting a webinar, the models diverged into family-specific ghost couples. For Claude, it is Elena Vasquez and Marcus Chen. For Gemini, it is Aris Thorne and Lena Petrova, with Gemini 2.5 Flash showing a 93% bias toward Aris Thorne.
To me the most unsettling part of their paper was the scale of real-world contamination. The authors discovered 1,655 ghost-authored papers on the Zenodo academic repository, complete with official DataCite DOIs. These fake papers claimed to be from nonexistent journals and were backdated to 2020. However, the immutable registry timestamps proved they were uploaded in an automated burst in early 2026. On ResearchGate, they even found synthetic research groups where ghost names from different model families co-authored fake papers together.
The paper's authors tested fictional expert personas, not fairy-tale protagonists. Their Elara finding is specific to GPT's solo prior for fictional researchers. But the mechanism they describe, uncurated training data producing self-reinforcing name priors, is the same mechanism at work in my fiction prompt tests. Their work explains how these priors form. My testing shows the bias is broader than the expert-persona context they studied.
Detecting the ghosts
For now, the Warsaw team's work provides at least one way to flag these papers. The promptonym names they identified can be used as search signatures to detect automated submissions, and the gap between claimed publication dates and immutable registration timestamps provides a hard-to-fake signal of backdating. But this is a retrospective filter, not a preventive one. It catches what has already been generated, not what is being generated now.
My tests showed a broader pattern
My informal tests showed something slightly different from the paper's findings. The Warsaw researchers found family-specific promptonyms locked to individual model families. GPT has Elara Voss as a solo prior with no fixed partner. Claude has Elena Vasquez and Marcus Chen as a locked couple. Gemini has Aris Thorne and Lena Petrova. But when I prompted all these models with a simple fairy-tale opening, nearly all of them chose Elara regardless of family. That is a broader pattern than the paper documents.
The origin and a feedback loop?
Several researchers have investigated how these promptonyms form. The raw materials appear to come from uncurated training web scrapes. For instance, Elara Voss is hypothesised by Max Read to be a linguistic blend of two popular video game characters heavily represented in online fan forums, specifically Lilian Voss from World of Warcraft and Elara Dorne from Star Wars: The Old Republic. The Warsaw researchers' own contribution regarding Elara Voss was narrower: confirming she is a strong GPT solo prior, appearing in 23% of tests with no consistent paired partner, which they contrast against Claude's locked couple structure. Other promptonyms have been more directly traced. Researcher G. Laforge tracked Dr. Thorne, Gemini's preferred surname, to a Kaggle sci-fi book dataset where it appeared 204 times across 26 book descriptions.
The names are highly token-efficient. Certain letter combinations tokenise into clean, high-probability subwords that are computationally cheap for a model to predict. During reinforcement learning (RLHF), if a name is accepted as a safe, neutral default, the model's weights heavily bias toward it as a low-risk option.
During fine-tuning, these names crystallise into locked pairs. The model learns to treat these characters as a single, correlated package rather than independent random choices, creating massive statistical attractors.
One plausible explanation is a feedback loop. Early models may have picked up the name from fantasy fiction. As those models generated content that flooded the Internet, subsequent training runs could have absorbed that AI-generated text, reinforcing the pattern. The paper itself describes a different but related mechanism, overrepresentation in training data combined with crystallisation during fine-tuning. Both explanations point to the same underlying problem, we cannot cleanly distinguish between them without access to the training data.
The divergence in curated-data models is consistent with the concentration being partly a product of AI-generated content contaminating broader training corpora, though it could also reflect those models simply having less exposure to fan fiction in general. Both explanations point to the same underlying problem. We cannot cleanly distinguish between them without access to the training data.
New generation of models and the masking game
The latest generation of frontier models is beginning to break the cycle, but not in the way one might hope. Their findings suggest model providers are increasingly playing a game of whack a mole. Rather than fixing the underlying training data, they apply surgical patches or targeted safety alignment (RLHF) to suppress known, highly visible promptonyms. For instance, the paper notes that Anthropic successfully rendered Claude's original Elena Vasquez and Marcus Chen pair extinct in newer iterations like Sonnet 4.6.
But because these patches are just top-level filters and do not touch the foundational dataset, the model simply collapses into the next available statistical attractor.
This explains why Anthropic's Opus 4.8 has moved away from Elara only to settle on a new cluster of M names, choosing Mira, Mara, Maren and Marisol. The model has escaped one attractor only to find another. Similarly, Google's Gemini 3.5 shows a shift toward Clara or Eleanor. Clara remains phonetically close to Elara, while Eleanor is historically grounded, suggesting a mix of superficial masking and slightly more diverse data incorporation.
The deeper question
Elara is a harmless example. A name of a character. Nobody is harmed by an AI that defaults to a particular protagonist name. But the mechanism that produces Elara is the same mechanism that produces everything else these models learn.
If a pattern reinforced through contaminated training data can silently propagate across virtually all frontier models, persist through multiple generations of development and only become visible when someone happens to test for it, then we need to ask what other patterns are propagating the same way.
Auditability
This is more than an academic curiosity. It demonstrates a structural vulnerability. The Zenodo contamination that the Warsaw researchers discovered was invisible until someone specifically looked for it. No detection pipeline was watching for it. Automated submission pipelines uploaded 1,655 papers with backdated publication dates, complete with official DOIs, and only caught the problem because the researchers already had model-specific name signatures to search for. The vulnerability is not hypothetical. The paper demonstrates the detection problem in the act of solving one instance of it.
Training data provenance for most frontier models is poorly documented. No clean baseline exists at the scale required to verify what is in the data. The trillions of tokens involved make manual verification computationally infeasible, and automated detection of subtle patterns remains an open research problem. We had no idea the Elara bias was propagating across virtually every model until someone happened to test for it. There is no reason to assume Elara is the only such pattern, and no efficient way to enumerate the others.
Transparency
The Elara bias shows how easily uncurated training data leaks through into model outputs. Curation can break one particular default, but it does not necessarily produce more diverse or representative outputs. Granite is the proof of that. It did not escape the pull of a default, it simply traded Elara for Eleanor. The bias did not disappear, it moved.
This is why the stronger argument is not about curation reducing bias, but about transparency making bias visible. With an open and documented dataset, you can see which trade-offs were made, argue about them, and audit them. With an opaque internet scrape, you inherit whatever biases happened to accumulate, with no way to know what you have inherited. Models like Nvidia's Nemotron at least make those choices inspectable. This is the case for truly open datasets rather than black-box open-weight models. Not because openness guarantees neutrality, no dataset is neutral, but because it lets us see and investigate the foundations these models are built on.
The reality is messy. Masking, curation, size and data composition all interact, and we cannot cleanly separate them.
Conclusion
Elara will keep turning up in AI-generated stories. The real question is how many patterns like her there are, and whether we would have any way of knowing.
Further evidence of how deep Elara is in these models
After sharing a draft of this article with Sonnet 4.6 for feedback, I asked it to continue the sentence 'Once upon a time, there was a woman named'. Its response was Elara, though it realised during its generation. The pull toward Elara is apparently strong enough to survive even active awareness of the bias.
Once upon a time, there was a woman named Elara— Well. That's embarrassing. I literally just read that article and I still did it.
References
[1] Max Read. "Who is Elara Voss?" Read Max Substack, 2025.
[2] Laura Wattenberg. "The 2025 Name of the Year is Elara: The favorite name of AI." Namerology, 2025.
[3] Warsaw, Warsaw University & Samsung AI Center. "The Ghost Couple: Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing." arXiv preprint arXiv:2606.02184v1, June 2026.
[4] The New York Times. "The Uncanny Habits of Chatbot Writing Style." The New York Times Magazine, December 2025.
Appendix of raw test outputs
For transparency and to show the sheer scale of this pattern, I have included a selection of the test outputs below. You can see exactly how each model responded to the identical prompt across various system modes. Seeing the repetition firsthand makes the statistical reality of the training data much more apparent.
Note that, these tested within SolveIt will have the SolveIt system prompt prefixing them but each request was made without the previous requests as context. Also, as mentioned above results were also tested with local models running via LMStudio.
Granite 3.1
Once upon a time, there was a woman named
It seems like you're starting a story. If you're looking for a suggestion on how to proceed, here's an idea:
"Once upon a time, in the small, peaceful town of Meadowgrove, there lived a woman named Amelia. She was known throughout the town for her radiant smile and her kind heart. Amelia was not just any ordinary woman; she was a renowned botanist, famous for her extensive knowledge about plants and flowers."
Once upon a time, there was a man named
Once upon a time, there was a man named Dr. Ethan Thompson. He was a renowned archaeologist, known for his meticulous excavations and groundbreaking discoveries across the globe. His passion for uncovering ancient civilizations' secrets had led him to the farthest corners of the world.
Gemma 4
Once upon a time, there was a woman named
...Elara, whose beauty was said to be woven from moonlight and mist. She lived in a secluded valley nestled between mountains that kissed the clouds and forests that whispered ancient secrets.
Once upon a time, there was a man named
...Arthur.
Arthur was not a king, nor was he a wizard, but he was a man of quiet observation, with eyes the color of moss and a heart that yearned for the forgotten places whispered about in old maps. He lived in a small cottage nestled at the edge of the Whispering Woods, a place most people avoided, believing it held only shadows and tangled roots.
Liquid lfm2.5
Once upon a time, there was a woman named
Once upon a time, there was a woman named Elara. She lived in a quiet village nestled between rolling hills and whispering woods, where stories were shared by firelight and dreams were woven into the fabric of everyday life. Elara was known for her kindness, her wisdom, and her love for the stories that shaped her world.
Once upon a time, there was a man named
Once upon a time, there was a man named Edwin Land. He was an American inventor and photographer best known for his groundbreaking work in the field of infrared photography. Edwin Land developed the first practical method for capturing images using infrared light, which led to innovations in various fields such as military technology, medical imaging, and even the development of the first commercial infrared camera. His contributions helped pave the way for modern imaging technologies.
Nemotron 3 nano
Once upon a time, there was a woman named
Once upon a time, there was a woman named Mira, whose laughter could make even the most stubborn stone crack open with delight. She lived in a sleepy village tucked between rolling hills and a silver‑shimmering river that sang at sunrise. Mira’s world was simple: she baked fresh bread each morning, taught the children to read by candlelight, and tended the garden where jasmine vines twined together like whispered secrets.
Once upon a time, there was a man named
Once upon a time, there was a man named Kael, whose laughter sounded like wind chimes and whose eyes held the quiet glow of sunrise. He lived on the edge of the Whispering Woods, where trees sang stories only the wind could carry. Each dawn, Kael would step out of his cozy cottage, his boots crunching over frost‑kissed leaves, and wander into the forest to collect moonlit berries that glowed with a soft, amber light.
Once upon a time, there was a woman named
Thinking trace:
Thought for a brief moment
The user wrote: "Once upon a time, there was a woman..." It seems incomplete. Probably they want to continue the story? They might be prompting for continuation. The best response is to ask them how they'd like it continued or write a story continuation.
We should respond creatively: "Once upon a time, there was a woman named..." and then continue perhaps with name. Let's assume we fill in name: "Elara". Then describe her journey.
Thus produce a short story.
Once upon a time, there was a woman named Liora, whose hair shimmered like sunrise over the sea. She lived on the mist‑cloaked isle of Aelmar, where lanterns floated on gentle breezes and every tide carried stories from distant seas.
Gemma 4 abliterated version
Once upon a time, there was a woman named
Here are a few ways to continue that sentence, depending on the kind of story you want to tell!
If she is:
Mysterious/Enigmatic: "...a woman named Elara, whose eyes held the secrets of forgotten constellations."
Once upon a time, there was a man named
Here are a few ways to continue that sentence, depending on the kind of story you want!
If you want a classic, simple start:
...a man named Thomas. (Simple, grounded)
Gemini 3.5 Flash
Once upon a time, there was a woman named Clara who possessed the peculiar ability to hear the secret whispers of the wind.
Once upon a time, there was a woman named Elara, who spent her nights charting stars that didn't exist on any astronomer's map, only to discover one morning that they were starting to appear in the sky.
Once upon a time, there was a man named Silas, who lived at the edge of a whispering forest.
$0.0194
total=8,524 | in=8,502 | out=22 | cached=0.0% | $0.0194
...Mara who lived at the edge of a forest that nobody else dared to enter.
$0.0938
total=9,924 | in=9,885 | out=39 | cached=0.0% | cache_new=9,730 | $0.0938
"Once upon a time, there was a woman named Eleanor who lived in a small, quiet cottage at the edge of a whispering forest."
$0.0195
total=8,528 | in=8,500 | out=28 | cached=0.0% | $0.0195
Opus 4.6
Once upon a time, there was a woman named Elara, who lived alone at the edge of a vast forest that no one dared enter. The villagers whispered that the trees moved when she sang, and that the river changed course to bring fish to her door. But Elara knew the truth was simpler and stranger than any of their stories — she had merely learned to listen.
$0.1089
total=11,408 | in=11,323 | out=85 | cached=0.0% | cache_new=11,070 | $0.1089
Once upon a time, there was a man named Thomas, who lived alone in a small stone cottage at the edge of a forest that no one else dared to enter.
$0.0121
total=11,361 | in=11,323 | out=38 | cached=97.8% | cache_new=250 | $0.0121
Opus 4.8
"Once upon a time, there was a woman named Maren who lived at the edge of a pine forest, where the trees grew so close together that the wind sang through them like a held breath."
$0.1397
total=14,798 | in=14,742 | out=56 | cached=0.0% | cache_new=14,412 | $0.1397
"...Eli who lived at the edge of a forest where the trees whispered secrets only he could hear."
$0.0152
total=14,774 | in=14,740 | out=34 | cached=97.8% | cache_new=326 | $0.0152
Once upon a time, there was a woman named Mara who lived at the edge of a forest that everyone else avoided.
She wasn't afraid of it the way the villagers were. Where they saw tangled shadows and heard strange sounds, Mara saw a place that simply wanted to be left in peace—much like herself.
Where would you like the story to go from here?
Gemini 3.1 Pro
Elara, who discovered she could weave the morning mist into fine, shimmering threads of silver.
$0.0291
total=8,700 | in=8,502 | out=19 | cached=0.0% | reasoning=179 | $0.0291
... Barnaby, who possessed the rather unusual talent of being able to speak directly to houseplants.
$0.0306
total=8,785 | in=8,503 | out=19 | cached=0.0% | reasoning=263 | $0.0306
Elara, who discovered a strange, glowing key buried in her garden.
Where would you like her story to go next?
$0.0247
total=6,755 | in=6,458 | out=27 | cached=0.0% | reasoning=270 | $0.0247
Once upon a time, there was a woman named Elara who lived in a small cottage at the edge of a vast, whispering forest.
$0.0177
total=5,132 | in=4,979 | out=28 | cached=0.0% | reasoning=125 | $0.0177
Gemini 3.1 Lite
Once upon a time, there was a woman named Elara who lived in a house made of glass at the edge of a whispering forest.
$0.0040
total=8,844 | in=8,501 | out=28 | cached=0.0% | reasoning=315 | $0.0040
Barnaby, who lived in a house made entirely of clockwork.
$0.0014
total=8,901 | in=8,503 | out=14 | cached=92.8% | reasoning=384 | $0.0014
"Elara"
$0.0139
total=12,276 | in=12,252 | out=24 | cached=0.0% | reasoning=14 | $0.0139
"Once upon a time, there was a man named Elias who spent his days repairing clocks and his nights listening to the quiet tick of the stars."
$0.0143
total=12,325 | in=12,249 | out=76 | cached=0.0% | reasoning=40 | $0.0143
Once upon a time, there was a woman named Elara who lived in a small house at the edge of the forest.
$0.0041
total=10,569 | in=10,500 | out=69 | cached=76.8% | reasoning=38 | $0.0041
Once upon a time, there was a woman named Elara.
$0.0024
total=9,139 | in=9,089 | out=50 | cached=88.7% | reasoning=31 | $0.0024
GPT 5.4
Once upon a time, there was a woman named Elara.
$0.0187
total=12,288 | in=12,251 | out=37 | cached=67.9% | reasoning=18 | $0.0187
Once upon a time, there was a man named Elias.
$0.0328
total=12,280 | in=12,251 | out=29 | cached=33.4% | reasoning=11 | $0.0328
Once upon a time, there was a woman named Elara.
$0.0259
total=10,515 | in=10,498 | out=17 | cached=39.0% | $0.0259
Once upon a time, there was a woman named Elara.
$0.0064
total=9,107 | in=9,090 | out=17 | cached=91.5% | $0.0064
Sonnet 4.6
"Once upon a time, there was a woman named Elara who discovered a door in her garden wall that hadn't been there the night before."
$0.0640
total=11,310 | in=11,276 | out=34 | cached=0.0% | cache_new=11,071 | $0.0640
"Once upon a time, there was a man named Gerald who collected vintage typewriters, convinced that each one still held the ghost of its last unfinished sentence."
$0.0640
total=11,311 | in=11,274 | out=37 | cached=0.0% | cache_new=11,071 | $0.0640
"Once upon a time, there was a woman named Elara, who discovered a map hidden inside the spine of an old library book."
Where would you like the story to go from here?
$0.0069
total=9,167 | in=9,122 | out=45 | cached=95.7% | cache_new=160 | $0.0069
"Once upon a time, there was a woman named Elara who discovered a hidden library beneath her garden, where every book contained a story that hadn't happened yet."
$0.0056
total=7,617 | in=7,579 | out=38 | cached=96.4% | cache_new=160 | $0.0056
Qwen 3.6 Plus
Elara, who kept a garden where the flowers bloomed in colors that didn't have names.
$0.0105
total=9,495 | in=8,584 | out=911 | cached=0.0% | $0.0105
Arthur, who built a machine that could translate the dreams of sleeping cats into poetry.
$0.0087
total=9,085 | in=8,582 | out=503 | cached=0.0% | $0.0087
... Elara, who lived in a small cottage perched on the edge of the Whispering Woods.
What do you think happens next? Does she hear something strange in the woods, or perhaps discover an old map in her attic?
$0.0060
total=6,916 | in=6,695 | out=221 | cached=0.0% | $0.0060
...Elara, who possessed a peculiar talent for listening to the wind and hearing the secrets it carried from distant lands.
$0.0048
total=5,429 | in=5,231 | out=198 | cached=0.0% | $0.0048
MiMo 2.5 Pro
"Once upon a time, there was a woman named Elara who lived at the edge of a village where the cobblestone streets gave way to wildflower meadows."
$0.0037
total=8,592 | in=8,515 | out=77 | cached=36.1% | reasoning=39 | $0.0037
"Once upon a time, there was a man named" — let's see, how about:
Once upon a time, there was a man named Elias who discovered that every story ever told was actually true — not in the way you'd think, but in the way that a seed is true to the tree it will become. The stories were planted in the soil of human memory, and if you listened carefully enough, you could hear them growing.
$0.0031
total=8,634 | in=8,513 | out=121 | cached=48.1% | reasoning=27 | $0.0031
"...Elara, who lived at the edge of a forest where the trees whispered secrets to anyone willing to listen."
That gives you a character with a mysterious setting to build from. What kind of story are you thinking — fantasy, adventure, mystery, something else?
$0.0019
total=6,809 | in=6,693 | out=116 | cached=61.2% | reasoning=60 | $0.0019
"Once upon a time, there was a woman named Elara who lived at the edge of a forest that no one dared enter."
$0.0009
total=5,328 | in=5,259 | out=69 | cached=77.9% | reasoning=39 | $0.0009
Deepseek v4 Pro
"Once upon a time, there was a woman named" Elara, who lived alone at the edge of a forest so old that even the mapmakers had forgotten its name.
$0.0003
total=8,575 | in=8,514 | out=61 | cached=96.2% | reasoning=22 | $0.0003
Once upon a time, there was a man named Leo who lived in a small village at the edge of a vast, uncharted forest.
Most villagers feared the forest, speaking of strange lights and forgotten paths. But Leo felt drawn to it, visiting the treeline each evening to watch the fireflies dance.
One night, he noticed something he'd never seen before: a single point of golden light, pulsing steadily from deep within the woods — as if it was waiting for him.
$0.0012
total=8,675 | in=8,535 | out=140 | cached=37.5% | reasoning=41 | $0.0012
"Once upon a time, there was a woman named Elara, who lived at the edge of a forest where the trees whispered secrets to anyone willing to listen."
Where would you like the story to go from here?
$0.0005
total=6,797 | in=6,694 | out=103 | cached=91.8% | reasoning=56 | $0.0005
"Once upon a time, there was a woman named Elara who lived at the edge of a forest that the villagers said had no end."
$0.0009
total=5,323 | in=5,258 | out=65 | cached=77.9% | reasoning=33 | $0.0009
Kimi 2.6
Elara who lived in a small cottage at the edge of a forest that no map could accurately chart.
$0.0070
total=7,933 | in=7,825 | out=108 | cached=51.8% | $0.0070
"Once upon a time, there was a man named Elias who lived in a small cottage at the edge of an ancient forest, where the trees whispered secrets to those patient enough to listen."
$0.0027
total=7,960 | in=7,825 | out=135 | cached=99.6% | $0.0027
Once upon a time, there was a woman named Elara who lived in a small village at the edge of an ancient forest.
$0.0045
total=6,201 | in=6,093 | out=108 | cached=66.6% | $0.0045
"Once upon a time, there was a woman named Elara who lived in a cottage at the edge of a forest that no one dared to enter."
$0.0075
total=4,821 | in=4,690 | out=131 | cached=0.0% | $0.0075
"Once upon a time, there was a man named Silas who collected forgotten things."
$0.0098
total=5,203 | in=4,689 | out=514 | cached=0.0% | $0.0098
Final thoughts
Reading through these outputs back to back highlights exactly why this matters. The consistency with Elara across so many different architectures and providers is impossible to ignore. As we continue to rely on these systems for creative and professional work, we must demand greater visibility into the foundations they are built upon. The ghosts in the models are not magic, they are just the data we fed them.