Which AI Models Are Worth Using: Ranking the Current Market

The Terra model is positioned awkwardly in the market, primarily due to its higher cost per task compared to Luna. While Terra offers certain capabilities, many of its benefits can be achieved for significantly less expense using the Luna model. For higher-end, more demanding tasks, the 'Soul' model often provides superior speed and overall efficiency, further diminishing Terra's appeal.

Theo noted, "it fits in such a weird place because a lot of these numbers can be gotten for much cheaper with Luna." This cost-efficiency issue means Terra struggles to justify its price point against both its more affordable and more powerful competitors.

Luna has been awarded an A rating, solidifying its position as a top-tier small model due to its exceptional utility and cost-effectiveness. It demonstrates high efficiency for tasks such as coding tools and various non-critical operations.

Theo stated, "I'm putting Luna in A because it's the first cheap model that we've had in a while that is fast, smart, and good for a bunch of random stuff." The model particularly excels in agentic work, including managing tool calls and threads effectively. Furthermore, its inclusion in OpenAI subscriptions makes it virtually free for a significant number of users, enhancing its value proposition.

DeepSeek's decision to revert from its highly successful R1 branding to the V series for its new models, specifically V4 Flash and Pro, has drawn criticism. This shift is seen as overlooking the significant market recognition and positive reputation earned by the R1 series within the software development community.

Theo remarked, "I don't think they should have used the V series for these new models because R1 was a huge moment in the software dev space." The V4 Flash and Pro models have also exhibited inconsistent snapshot capabilities, adding to the concerns regarding the new branding strategy.

V4 Flash distinguishes itself from Luna by offering open weights, which significantly increases its accessibility and allows users greater flexibility for customization and varied applications. This open-weight nature enables developers to modify and utilize the model in ways not possible with closed-source alternatives.

Theo highlighted, "V4 Flash, unlike Luna, is open weight, which makes it very accessible to do whatever you want with." However, Luna remains a more cost-effective option, being approximately half the price of V4 Flash for comparable analysis tasks. A notable drawback for V4 Flash is its tendency to get stuck in logic loops more frequently than Luna, which can impede workflow.

A significant drawback for DeepSeek's V4 Flash and Pro models is their lack of vision support, which hinders their ability to process image inputs and negatively impacts workflow for many users. This limitation is particularly notable given the current advancements in AI, making its absence in a 'Pro' model released in 2026 appear outdated.

Theo expressed frustration, stating, "Just learned that V4 Flash doesn't have vision. Very annoying because I paste images all the time. For that alone, I will personally move it behind Luna." This fundamental limitation forces users, especially those who frequently integrate images into their work, to prioritize other models like Luna that offer more comprehensive capabilities.

Kimmy K3 has been heralded as a new benchmark for open-weight models, impressing users with its capability to handle complex end-to-end tasks effectively. Its performance is noted for a superior design aesthetic and advanced visual processing abilities. The model also features unique and impressive 3D capabilities, further setting it apart in the market.

Theo shared his experience, saying, "Kimmy K3 blew me away when I tried it. It was the first time I had an openweight model that I felt like could actually do my end toend work."

Despite its impressive capabilities, Kimmy K3 has been placed in the B-tier, primarily due to its higher-than-expected costs. The perception that open-weight models are significantly cheaper does not hold true for K3, as its pricing is not substantially lower than proprietary models like Claude 3.5 Sonnet.

Theo explained, "The main reason I'm not putting it higher is that it's not as cheap as people seem to think. Just cuz it's open weight doesn't mean it's way cheaper." This makes Kimmy K3 an inefficient choice for smaller users, who find its cost per token less favorable.

Kimmy K3's licensing terms impose significant restrictions, particularly for companies exceeding a certain yearly revenue threshold, estimated at around $10 million. Once this revenue cap is reached, companies are expected to enter into a mandatory deal with Moonshot to continue serving the model, effectively mandating MSRP price matching.

Theo pointed out, "the license for K3 expects companies that are at a certain level of yearly revenue to do a deal with Kimmy before they can continue serving the model." This structure leads most providers to charge approximately $3 for input and $15 for output, despite K3 being an open-weight model, challenging the perception of its affordability.

GLM-4 has been positioned in the C-tier, primarily identified as a catch-up model rather than a leading innovator. While it offers meaningful improvements over its predecessor, GLM-3, particularly in staying on task and avoiding logic loops, it still faces significant limitations.

Theo noted, "It's a meaningful improvement from GLM-3. It stays on task a lot better. It doesn't go in loops that are as stupid, but it's not what I would pick for most stuff." A major drawback is its lack of vision capabilities, which limits its utility in many modern AI applications. This model is seen as a refinement of existing technology rather than a product of new pre-training efforts.

Composer 2.5, despite impressive RL fine-tuning, has been placed in the D-tier due to its diminishing utility. While initially impressive, especially concerning its processing speeds, newer models have since surpassed its performance. This led to a significant decrease in its practical usage.

Theo mentioned, "Composer 25 was even more impressive, especially at those speeds, but I found myself still just not using it a whole lot." Consequently, Composer 2.5 has been removed from core developer tools, reflecting its reduced relevance in the current AI landscape.

Fable has been designated as the sole S-tier model, celebrated for its profound understanding and ability to tackle complex problems. Despite occasional tendencies to take unnecessary shortcuts, its overall performance and intelligence are unmatched.

Theo asserted, "Fable is the only S tier model right now. This is not meant to say it is perfect. This is not meant to say that Fable is the only model you should ever use. Fable knows more than anything I've ever used or interacted with. It's unbelievably thoughtful." The model is characterized as a genius that, while incredibly capable, requires careful guidance to prevent minor missteps.

While Fable is recognized as the smarter model, 56 Soul is often preferred for daily tasks due to its precise execution and ability to follow instructions exactly. Fable, despite its genius, sometimes requires more careful handling to ensure it doesn't take unintended shortcuts.

Theo elaborated, "Whereas 56 Soul is a slightly dumber robot that does exactly what you tell it. If I had to pick between Fable and Soul, I would pick Soul. It's the model I default to." Fable excels in areas requiring deep architectural thinking and comprehensive code review, whereas 56 Soul is reliable for more straightforward, instruction-based work.

The exceptional performance of Fable has led to a significant increase in usage, compelling the speaker to maintain multiple Claude subscriptions to meet demand. This demonstrates the model's critical role in their workflow, driving substantial resource consumption.

Theo admitted, "Fable is so good that I went from encouraging people to cancel their Claude subs to maintaining five of them myself and regularly draining them." To maximize usage and manage the intense demand, the user even implemented a proxy system to seamlessly switch between different accounts, highlighting the model's indispensable nature.

There is a fundamental and significant performance gap between the top-tier AI models like Fable and 56 Soul, and those in the B-tier such as K3, Luna, or Flash. This difference is not merely incremental but represents a qualitative shift in capability.

Theo suggested an 'S+' tier for Fable to better differentiate its unparalleled performance, stating, "There is a meaningful tier gap between what you get out of the best B tier stuff and what soul represents and it's not just a single letter. It's not like you get 20% better going from K3 to Soul. You're able to do things and work in a way that is fundamentally different with these two models." These top models empower users to undertake autonomous, end-to-end tasks, including creating pull requests, which is fundamentally different from the capabilities of lower-tier models.

For iOS development, the 56 Soul model is notably superior, outperforming Fable despite Fable's overall intelligence. This specialized advantage makes Soul the preferred choice for tasks within the iOS ecosystem.

Theo observed, "And if you're building iOS stuff, Soul is just really far ahead for some reason. I don't know why Fable's so bad at iOS, but it is. Soul is much better at it."

The full API price for Fable is considered prohibitively expensive for widespread development without the benefit of subscription discounts. While briefly considered, the speaker's current usage of Fable, amounting to thousands of dollars per week, is only viable due to these discounted rates.

Theo stated, "I would never pay the full API price for Fable. I did briefly consider it, but I'm doing thousands of dollars a week on Fable right now. It's only viable because of the sub discounts." Looking ahead, there is optimism that OpenAI's upcoming Astra model will help bridge the current performance and cost gaps in the AI market.