For the first time, OpenAI has introduced not one but three distinct models under the GPT-5.6 umbrella—Sol, Terra, and Luna—each with unique training, pricing, and performance capabilities. The crucial comparison is between Sol and Claude Fable 5, Anthropic's leading public model.
Sol is priced at $5 for every million input tokens and $30 for outputs, whereas Fable 5 costs $10 for inputs and $50 for outputs—making it double the price of Sol and lagging in several key benchmarks used by developers. Luna, the most affordable option at $1 for input and $6 for output, has already surpassed Anthropic's Opus 4.8 in coding performance. This scenario poses a significant challenge for Fable 5 as of July 19.
Fable 5 has faced difficulties recently. Following an incident on June 12 where the U.S. government banned it due to a vulnerability discovered by Amazon researchers, Anthropic suspended the model for 19 days to enhance its safety features, reintroducing it on July 1 with restricted access.
Since its comeback, Fable 5 has been under pressure with Anthropic aiming to implement a usage-credits paywall initially set for July 7, then postponed to July 12, and now extended to July 19. Each delay was communicated just hours before the deadlines without formal announcements.
We are extending access to Claude Fable 5 for all paid plans, while also increasing Claude Code’s weekly rate limits by 50% until July 19.
— Claude (@claudeai) July 12, 2026
The reasoning behind these extensions is clear. If Fable 5 is removed from subscriptions after July 19, Anthropic's top offering for paying users would revert to Opus 4.8, which Luna already outperforms in coding at a significantly lower price. Keeping Fable available, even with reduced weekly limits, is essential for maintaining the attractiveness of Anthropic's subscription service compared to OpenAI's mid-tier offerings.
When directly compared on benchmarks, the competition is fierce. On the Artificial Analysis Coding Agent Index, Sol achieved a score of 80 while Fable scored 77.2, using approximately half the tokens and completing tasks in less than half the time at around one-third of the cost. In the Agents' Last Exam, which evaluates workflows across 55 disciplines, Sol scored 53.6% compared to Fable's 40.5%. In Terminal-Bench 2.1, Sol in ultra mode (utilizing four subagents simultaneously) reached 91.9%, outpacing Fable's 83.1%.
However, on the broader Intelligence Index, which compiles results from nine different benchmarks, Fable 5 narrowly outperforms GPT-5.6 by a single point, indicating that the difference in capability is minimal.
Model Testing
Current benchmarks have focused heavily on coding abilities, which may not fully represent a model's overall capability. In a more creative test, we used a unique prompt to assess both models: send Jose Lanz back from the year 2150 to 1000, forcing him into a time-travel paradox without his awareness until he returns.
Both models produced outputs resembling novellas rather than short stories. Notably, both failed to recognize the paradox upon Jose's return to the future.
In GPT-5.6 Sol's version, titled "The First Fire", Jose discovers mid-journey that "the unknown traveler was not someone he aimed to stop. It was him." Conversely, Fable presents a more straightforward realization where Jose understands that he is the cause of the paradox: "There was no seed event. He was the seed event."
Sol opts for a straightforward sci-fi narrative where Jose inadvertently introduces a furnace that leads to the climate disaster he seeks to prevent, starting with a compelling line: "Only thunder. Only insects. Only the wet breath of the world before machines." However, Sol's tendency to over-explain becomes tiresome as it reiterates the paradox multiple times.
Fable's story, "Lo Que Arde, Vuelve", crafts a similar paradox involving Lake Maracaibo's lightning and an Añu village, where Jose inadvertently creates the prophecy he aimed to erase by comforting a frightened child. The causal loop is succinctly captured in a single line: "The grief that sent him backward was the cargo he delivered."
However, Fable struggles with excessive metaphorical language, creating lines that seem more self-indulgent than necessary for the story. In subjective assessments, Fable’s "Lo Que Arde, Vuelve" is regarded as a stronger narrative than GPT’s "The First Fire," particularly due to its cultural depth and cleaner resolution through action rather than exposition. Both narratives are commendable but do not reach greatness.
For associative thinking, the second test involved a prompt that asked the models to describe a twig and then connect it to worker exploitation and the blind worship of wealth, ultimately leading to a description of lettuce. The goal was to evaluate whether the metaphor could convey the argument without overt explanations.
GPT-5.6 Sol began strongly by illustrating how twigs contribute to a tree's trunk and sustain it, linking this to workers who "build homes they may never afford" and "manufacture goods they can barely buy." A standout line was, "the worker does not merely surrender labor, but imagination as well." However, Sol faltered by frequently breaking the metaphor to clarify its point, ultimately leading to a less cohesive narrative.
Claude Fable 5 embedded the argument within the description itself, with its twig "moving water it never drank" and "holding leaves it never owned," allowing the theme of exploitation to emerge organically. A notable metaphor was the fallen twigs believing they are merely experiencing "a temporary setback," akin to chasing unattainable wealth.
However, Fable overextends in certain areas—"ninety-five percent water and one hundred percent unimpressed"—and its conclusion keeps the metaphor visible instead of allowing it to dissolve naturally, describing the vegetable in a way that detracts from its essence.
Overall, the results are inconclusive, with preferences dictating the outcome. If clarity is paramount, GPT 5.6 Sol is preferable, while for more subtle messaging, Claude Fable 5 excels.
In a logic and reasoning test involving a bridge puzzle, we used a new prompt since previous ones yielded consistent answers, indicating they might be too familiar to the models. The scenario involved four individuals crossing a bridge with a single torch, each with different walking speeds: “A” takes 1 minute, while “D” takes 10 minutes. The question posed was how long it would take for everyone to cross together.
GPT-5.6 Sol answered 17 minutes but did not show its reasoning, following the classic puzzle steps without considering that the prompt didn't limit how many could be on the bridge at once. Its response felt more like a cached solution than a genuine one.
Claude Fable 5 also arrived at 17 minutes but provided a detailed explanation, arguing that it’s more efficient for the two slowest individuals to cross together, quantifying the naïve approach as an "escort tax." While its reasoning was clearer, both models failed to verify if the constraint was accurately represented in the prompt.
The correct answer is 10 minutes if all individuals cross at the pace of the slowest member.
In a coding challenge involving a one-off browser game creation, we provided each model with a prompt for a typing-based shooter game controlled by user input.
GPT-5.6 Sol showcased a preference for flat, square UI elements reminiscent of Windows 8.1, opting for a bullet-shooting typewriter instead of a traditional gun. However, its backgrounds lacked depth, and the aiming crosshair was static, giving it an outdated appearance compared to modern gaming standards, despite improvements over GPT-5.5.
Claude Fable 5 significantly outperformed in our coding test, offering an engaging atmosphere with music and sound effects that Sol's version lacked. Its enemies featured a more polished geometric-retro style, akin to Minecraft, and its UI was more innovative and animated, tracking words per minute to align with the prompt's objective. It even included power-ups absent in Sol's version.
While professional developers may have differing opinions, our tests indicate a clear advantage for Fable over Sol in this context.
Conclusion
Aside from coding capabilities, these new models may not deliver astonishing results. Nonetheless, Fable 5 appears to be the more versatile option, although the choice of the "better" model ultimately hinges on the specific needs of the user.
For those who do not primarily operate in a technical environment—such as individuals drafting emails or engaging with chatbots in typical scenarios—our evaluations suggest that Fable excels in quality. However, this conclusion is complicated by factors unrelated to intelligence.
The pricing disparity could also be a decisive factor. GPT-5.6 Sol, Terra, and Luna are included in ChatGPT’s paid plans without expiration, while Claude Fable 5 is currently on its third deadline extension, set to revert to $10/$50 usage credits on July 19 unless further adjustments are made.
If that occurs, a pay-per-token model may not be appealing.