Is the newest AI model always the best one? We put four to the test

Every fewer months, a caller AI exemplary arrives claiming to beryllium the smartest yet. Better benchmarks. Better reasoning. Better coding. Better everything.

It’s easy to presume that immoderate launched astir precocious is automatically the sharpest instrumentality for the job. But alternatively of relying connected benchmark reports and trading claims, we decided to put them to the test. We wanted to find retired which models are really worthy using.

We selected 4 of today’s astir talked-about AI models: 

• Anthropic’s Claude Fable 5
• Anthropic’s Claude Opus 4.8 
• OpenAI’s GPT-5.6 Sol 
• Anthropic’s Claude Sonnet 4.6 

Each received the nonstop aforesaid 3 prompts pinch circumstantial goals:

• Building a landing page
• Writing a website’s location page copy
• Analyzing income data

We described the desired result successful the prompts but did not supply layout instructions, step-by-step checklists, aliases immoderate follow-up context. Instead, we fto each exemplary determine really to get there.

This attack fto america spot really good each exemplary handled ambiguity connected the first try. Would it inquire clarifying questions aliases confidently make its ain assumptions? And conscionable arsenic importantly, really overmuch clip and money would it return to scope a bully result?

Here’s what we discovered.

Which exemplary offered the astir worth for creating a landing page?

We first challenged each exemplary to creation and build a landing page from a azygous imaginative brief. This is the punctual we used:

Design and build a premium landing page for a pottery workplace called "Primavera". The page should evoke the emotion of a serene creator studio: calm, sophisticated, and cautiously paced. The landing page should person 3 pages, including a contact/booking form. I want the superior acquisition to beryllium 1 continuous scroll, pinch soft transitions betwixt pages. Include a bold, animated leader constituent that acts arsenic the ocular centerpiece of the full experience. It should beryllium interactive and induce exploration done personification input. It should evoke the tactile emotion of being successful a pottery studio. It is NOT a fixed image aliases ornamental effect.

We intentionally near plentifulness of room for mentation to trial the models. Rather than prescribing layouts aliases components, the punctual described conscionable the feeling, interactions, and wide acquisition we wanted.

We past deployed each AI model’s output to Hostinger from Claude Code, Cursor, and Codex desktop apps via Hostinger Connector. The aforesaid deployment process was utilized for each models to support consistency successful the comparison.

Sonnet 4.6’s output

The Sonnet-built website met each the requirements successful the prompt, but its leader animation was the simplest of the four. It was besides the only exemplary to move distant from the clay-inspired colour palette connected the leader section.

Even pinch these shortcomings, Sonnet produced a polished, well-structured website that fulfilled the brief. Despite being 1 of the older models successful this benchmark, it proved it tin still clasp its ain against overmuch newer models.

Sonnet 4.6 generated website: https://teal-camel-563249.hostingersite.com/ 

GPT-5.6 Sol’s output

Sol took the astir literal mentation of “tactile.” Its leader conception lets visitors resistance and style a portion of clay pinch a reset button, making it the clearest execution of the prompt’s petition for an interactive centerpiece.

The remainder of the tract leaned into short, punchy microcopy and a tidy three-tier shop list. It was polished and on-brand, but the website contented was beautiful comparable pinch the Sonnet 4.6’s produced outcome.

GPT-5.6 Sol generated website: https://primavera-ceramics-studio.hostingersite.com/ 

Opus 4.8’s output

Opus produced the astir elaborate landing page. It expanded the fictional backstory of the pottery studio, introduced shop statistics, and created 5 tiers of workshops alternatively of three.

Its interactive clay-shaping leader was conscionable arsenic polished arsenic Sol’s, and successful position of sheer completeness, Opus arguably responded to the little astir thoroughly.

The trade-off was that the page besides felt busier, which didn’t needfully make the acquisition stronger – conscionable denser.

Opus 4.8 generated website: https://sandybrown-goshawk-747835.hostingersite.com/ 

Fable 5’s output

Fable took the astir technically eager attack to a richer, graphics-driven experience. 

The interactive pottery animation was the highlight. Clicking and dragging shapes the clay successful existent time, responding smoothly to guidance and unit and intimately mimicking the consciousness of hand-building pottery.

Sol attempted thing similar, but the consequence didn’t look for illustration clay being molded the measurement Fable’s did – it felt little for illustration shaping a three-dimensional worldly and much for illustration dragging a level image around.

While this relationship was memorable, the wide transcript and messaging weren’t arsenic beardown arsenic Sol’s aliases Opus’s.

Fable 5 generated website: https://gray-albatross-412637.hostingersite.com/ 

The winner

Visual and copywriting preferences are inherently subjective, truthful we understand that the champion consequence will alteration from personification to person. 

To make the comparison much balanced, we besides evaluated each exemplary based connected task completion clip and cost, calculated arsenic a dollar magnitude based connected each model’s monthly usage limits.

Here’s really the 4 models compared:

AI model	Time	Cost	Token usage successful full (input, output, cached)
Sonnet 4.6	~14 min	$2.77	6M
GPT-5.6 Sol	~30 min	$8.74	11M
Opus 4.8	~12 min	$4.68	6.6M
Fable 5	~25 min	$9.89	3.7M

*Cost calculated arsenic a dollar magnitude from monthly usage limits.

On earthy output, this one’s a genuine toss-up – immoderate of the newest model’s results could reasonably beryllium called the astir awesome build, depending connected what creation worth you for illustration the most. 

But “most impressive” and “best value” aren’t the aforesaid question, and the token and costs information make the second easier to answer.

A speedy statement connected the table: costs depends connected some the number of tokens utilized and the value per token. Since models are priced differently, 1 tin usage less tokens yet still extremity up costing much if it belongs to a pricier tier.

That’s what shows up here. Sonnet and Opus utilized a comparable number of tokens (6M vs. 6.6M), yet Opus costs 69% more. This shows that the spread comes from the pricing tier, not from Opus being immoderate little businesslike pinch its tokens. 

Fable makes the aforesaid constituent from the different direction. It utilized the fewest tokens of immoderate exemplary successful the trial (3.7M) but was the astir costly of the four, because its per-token value sits good supra the remainder of the field. 

Sol, connected the different hand, sewage to a higher measure the much straightforward measurement by utilizing astir doubly arsenic galore tokens arsenic immoderate different exemplary (11M).

Put together, Sonnet wasn’t being the astir frugal pinch its tokens – Opus was conscionable arsenic frugal. But Sonnet delivered a complete, on-brief consequence while sitting successful the lowest-cost tier of the four, which is why it came retired cheapest by a wide margin, sloppy of really you portion the data.

That makes Sonnet 4.6 our prime for this benchmark. It didn’t build the flashiest site, but it delivered a complete, on-brief consequence astatine a fraction of the costs of the alternatives. 

If fund and turnaround clip are portion of your decision, that’s the much useful awesome than immoderate azygous creation choice.

Which AI exemplary executed the imaginative penning little best?

Our 2nd benchmark focused connected text-based activity that galore group fold into their day-to-day, much lightweight tasks. 

Do statement that we don’t expect immoderate AI exemplary to lucifer the output of knowledgeable copywriters, but we do expect it to nutrient a draught worthy further refining.

To spot which exemplary came closest, we asked each 1 to constitute the leader conception for a premium location organizing work pinch this prompt:

You are penning a homepage leader transcript for a location organizing work that charges much than its competitors. Write a headline, a subheadline, and a two-sentence CTA that earns the value quality without mentioning price, competitors, aliases utilizing the words "transform." The transcript should make personification consciousness understood earlier it makes them consciousness sold to.

Sonnet 4.6’s output

Headline: You cognize wherever everything goes. You conscionable request abstraction to get there.

Subheadline: We don’t conscionable clear clutter — we fig retired why it came backmost past time. Then we build a location that really useful for really you live.

CTA: Every convention starts pinch a conversation, not a bin bag. Book a free walkthrough and we’ll representation retired precisely what your location needs — earlier thing moves.

Sonnet’s header reframed the scholar arsenic capable, not chaotic, which is simply a smart move for a premium-priced service. The transcript was polished and fresh to use. 

The trade-off: Sonnet appended a afloat paragraph explaining its ain choices, unprompted. This tin beryllium useful if you want the rationale, but it’s other scrolling for a task framed arsenic speedy and everyday.

GPT-5.6 Sol’s output

Headline: Your location isn’t messy. It’s carrying excessively much.

Subheadline: We create thoughtful, lasting systems astir the measurement you really live—so your abstraction feels calmer, functions better, and stays that way.

CTA: Let’s make room for the life you’re trying to live. Book a consultation to commencement pinch what’s weighing connected you most.

Sol’s header was the tightest and astir quotable of the four. It did the aforesaid reframe arsenic Sonnet’s successful less words. It’s besides the only exemplary that returned conscionable the copy, pinch zero commentary. For a lightweight, paste-and-use task, Sol has the astir businesslike output of the bunch.

Opus 4.8’s output

Headline: The clutter isn’t a characteristic flaw. It’s conscionable been waiting for the correct system.

Subheadline: You’ve tried the bins, the play purges, the promises to yourself. What you haven’t had is personification who studies really you really live—then builds thing astir it that still useful successful six months, not conscionable the time we leave.

CTA: Tell america astir the room you support closing the doorway on. We’ll travel spot it pinch caller eyes and nary judgment, and show you precisely what’s imaginable earlier you perpetrate to anything.

Opus’s transcript was empathetic and well-crafted, but it ran longer than the others. The subheadline successful peculiar did much explaining than a leader conception typically needs. It publication well, but a copywriter would request much refinement effort earlier shipping it. Like Sonnet, it came pinch reasoning notes attached.

Fable 5’s output

Headline: A spot for everything you love. Nothing else.

Subheadline: We build systems astir the measurement you really unrecorded — down to the drawer you unfastened half-asleep — truthful everything stays precisely wherever your hands expect it, agelong aft we’ve gone.

CTA: The rooms you locomotion backmost into coming are already shaping tomorrow. Schedule your consultation and determine what they go instead.

Fable’s header was the astir compressed and stylistically ambitious, though “Nothing else” carried a flimsy consequence of ambiguity that a quality editor would astir apt emblem arsenic retired of context. The subheadline’s “drawer you unfastened half-asleep” was a genuinely strong, circumstantial detail. Like the others, isolated from for Sol, it added explanatory notes beyond the transcript itself.

The winner

All models produced headlines beardown capable to past a ray edit. But for this benchmark, we’d telephone it for GPT-5.6 Sol, because it was the only exemplary that understood the duty arsenic a lightweight, mundane task and answered accordingly. 

It handed backmost precisely 3 things: a headline, a subheadline, and a CTA. Nothing to portion out, thing to scroll past.

That said, this is close: Sonnet’s header reframing was arguably the sharpest of the four, and if you’re the benignant of personification who wants the “why” alongside the draft, Sonnet aliases Opus would beryllium the amended fit.

Which AI exemplary reasoned champion for information analysis?

Another communal business usage lawsuit for text-based AI is information analysis, truthful we tested the 4 models’ reasoning. 

We prepared a CSV record containing dummy income data for a mini business trading puzzle books and merchandise (Puzzled? Hat, Custom Puzzle Notebook, and Puzzle Grid Tote Bag). There’s besides a astonishment successful the 3rd month’s income numbers, generating much gross than usual.

We asked each exemplary to analyse it and constitute actionable business recommendations by sending the record together pinch this prompt:

I'm attaching six months of income information for a mini ecommerce business. Give me: (1) a one-paragraph plain-English summary of what the information shows, (2) 1 underperforming area pinch a circumstantial logic why — not a wide observation, a diagnosis, and (3) 1 actual action to return based connected the numbers. Not a wide proposal — a circumstantial 1 I tin enactment connected adjacent week.

Sonnet 4.6’s output

Sonnet caught the March wholesale-order skew and was the only exemplary to support its study pinch charts, plotting gross separator by merchandise truthful the hat’s underwater separator is intolerable to miss. Its test was clear, and its recommended action was actual – pausing promotions until the guidelines value is fixed. Clean and correct, but it stops astatine the repair.

GPT-5.6 Sol’s output

GPT-5.6 Sol flagged the March anomaly and worked from the fullest gross picture, factoring shipping income into its summary. Its test added a useful layer: shipping charges didn’t screen shipping costs, deepening the hat’s per-unit nonaccomplishment beyond the pricing problem alone. 

The action was the astir cautious of the 4 – a humble reprice calculated to beryllium conscionable supra break-even, positive pulling the chapeau from some promo codes. Sound, but purely defensive, pinch nary upside quantified.

Opus 4.8’s output

Opus delivered the sharpest azygous penetration of the test: each chapeau bid was a hat-only purchase, truthful the merchandise can’t moreover beryllium defended arsenic a nonaccomplishment leader that pulls successful book sales. 

Its reprice was the boldest, deliberately group to bring the chapeau successful statement pinch the margins of the remainder of the merchandise, pinch the monthly effect modeled. 

It besides added a caveat that nary different exemplary raised: verify the accumulation costs is existent earlier repricing, because if it’s stale, the existent hole is sourcing alternatively than price.

Fable 5’s output

Fable 5 reached the aforesaid test but traced the harm furthest: respective chapeau income had gone retired astatine a discount via the store’s email and societal promo codes, meaning trading walk was actively amplifying the loss. Its action paired the hole pinch a lever – reprice the chapeau (or propulsion it entirely), and successful the aforesaid edit, switch it retired of some promos for the tote bag, the store’s highest-margin merchandise item, turning each promoted waste from a nonaccomplishment into a gain.

The winner

We deliberation Fable 5 delivered the strongest answer, though this information was person than the test unsocial suggests. All 4 models caught the March wholesale distortion and correctly identified the mispriced hat. The separation came wholly from what they did pinch it.

Fable 5 was the only exemplary that treated the trading perspective arsenic portion of the problem. It redirected the promo slots toward the highest-margin merchandise and quantified the plaything per sale. Opus 4.8 was a very adjacent second, pinch the deepest repricing mathematics and the smartest caveat of the test.

Sonnet 4.6, meanwhile, held its ain pinch pattern-catching. It spotted the wholesale skew and was the only exemplary to visualize its findings. However, its recommended action was simply the plainest of the four: a correct value fix, pinch thing layered connected top.

Our honorable return connected the research results

So, does the latest exemplary ever nutrient amended results? Based connected these benchmarks, not necessarily. 

Output value wasn’t wished by which exemplary was newest, but by how good each model’s strengths matched the task.

A awesome thought tin win pinch almost immoderate modern AI model. But moreover the astir tin exemplary can’t rescue a anemic thought aliases an unclear prompt.

No matter which exemplary you choose, these 3 habits consistently lead to amended results:

• Write goals, not checklists. Describe the result you want, not each measurement to get there. This gives the AI models room to logic and find higher-leverage solutions. Reserve rigid instructions for workflows that must nutrient the nonstop aforesaid consequence each time.
• Define your boundaries. If you don’t specify what’s successful aliases retired of scope, the exemplary will make those decisions for you. Be definitive astir constraints, required formats, and what occurrence looks like.
• Manage your context. As conversations turn longer, models go much apt to suffer way of earlier information. Start caller erstwhile a task changes, and debar cramming unrelated activity into a azygous chat.

After all, AI is becoming little astir uncovering the champion exemplary and much astir learning really to activity good pinch whichever 1 you have.