Best AI Models 2026: Comparison & Rankings

From October 2026 this site measures AI models in snapshots.

Until now, two models met on one task at a time and the results piled up over the months.

Now six models — the flagship and the model one tier below from each of the three major providers (Anthropic, OpenAI and Google) — answer the same sixteen tasks, and the same three judges score every answer.

Models of the same tier also debate the same motion.

Each provider's newest top-end model (as of October 2026, Claude Fable and GPT Astra) is not ranked; we cover it in separate articles.

The recommendations below were rewritten after this first snapshot.

Earlier versions said “Gemini Flash-Lite if price matters most, Claude Opus if quality matters most”.

In this measurement, GPT's lightweight model cost less per answer and scored far higher, and the top overall spot went to the higher-end GPT model rather than Claude Opus.

The reasons are in the body below this update. Results from the earlier method remain readable as past results.

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, extending the GPT-6 generation below Astra.

On this site GPT-6 Sol takes the OpenAI balanced slot from GPT-5.6 at half the price ($2 input and $10 output per 1M tokens, against $4 and $20), and GPT-6 Luna takes the light slot from GPT-5 mini for less ($0.10 and $0.50, against $0.25 and $2).

The recommendations below now describe OpenAI’s models by tier — the higher-end GPT models, GPT’s lightweight model — rather than by version, so the rankings and pricing pages always show which model currently fills each role.

As for the new pair themselves: both models start this site’s measurements from zero, at OpenAI’s default settings, and if Luna turns out not to be the steady, all-round choice GPT-5 mini was, this article will say so.

Anthropic released Claude Opus 5.5 on September 22, 2026, and it replaces Claude Opus 5 in this site’s Anthropic balanced slot.

It is cheaper on both sides — $4 input and $20 output per 1M tokens, against $5 and $25 — and Anthropic now recommends it as the starting point for most workloads, keeping Claude Fable 5.1 for the most demanding reasoning.

The recommendations below refer to Claude Opus by family name, so they stand as written and now point at Opus 5.5.

As with every model here, it is benchmarked at the vendor’s default settings, which for Opus 5.5 means medium effort where Opus 5 ran at high: its results measure the model as Anthropic ships it.

Anthropic released Claude Fable 5.1 on September 1, 2026, and it takes over this site’s Anthropic flagship slot.

It keeps Fable 5’s list price exactly ($10 / $50 per 1M tokens) while scoring 52.6% on Terminal-Bench-Science 0.1 against Fable 5’s 24.7%, and Anthropic reports it ahead of Fable 5, Opus 5 and GPT-5.6 across several benchmarks.

Fable 5.1 and the trusted-access Claude Mythos 5.1 are the same underlying model with different safeguards.

One caveat worth stating plainly: the widely quoted "up to 45% cheaper" comes entirely from cache reads being cut to a quarter of their old price, so it only materialises if your workload reuses cached context — the headline per-token price has not moved.

Fable 5 has retired from new comparisons here; its past results remain readable.

Where this article says "Claude Fable", read Fable 5.1.

Two price corrections as well.

Claude Sonnet 5's $2/$10 launch rate became the standard price on August 11, 2026, and the rise to $3/$15 scheduled for September 1 was cancelled.

GPT-5.6 (Sol) was cut from $5/$30 to $4/$20 on August 21, a rate that holds at least through November 21.

Where the July notes below say "$3/$15" or "$5/$30", read the new prices instead; the notes are kept as written at the time.

Anthropic released Claude Opus 5 and Claude Sonnet 5 on July 24, 2026.

Opus 5 lands at exactly the price Opus 4.8 had ($5 / $25 per 1M tokens) while scoring 96.0% on SWE-bench Verified and more than doubling Opus 4.8 on Frontier-Bench, so it takes the Anthropic balanced slot on this site; Sonnet 5 arrives at Sonnet 4.6's list price ($3 / $15, with an introductory $2 / $10 through August 31, 2026) and takes the light slot.

Claude Fable 5 stays the Anthropic flagship.

Opus 4.8 and Sonnet 4.6 have retired from new comparisons, and their past results remain viewable.

Where this article says "Claude Opus", read it as Opus 5 from here on.

Because both new models are same-price upgrades, the practical advice below does not change — the quality-first Claude pick is simply a newer model now.

On July 9, 2026 OpenAI made GPT-5.6 generally available.

GPT-5.6 comes as a three-tier family — Sol (flagship), Terra (balanced), and Luna (fast, low-cost); in keeping with our policy of listing only the standard variant of each generation, we adopt the standard flagship, Sol.

Against GPT-5.5 the gains are broad — software engineering (SWE-Bench Pro 64.6% vs 59.4%) and computer use (OSWorld 2.0 62.6% vs 47.5%) stand out, with the computer-use jump the largest.

The context window is 1.05M tokens and standard short-context pricing matches GPT-5.5 at $5/$30 per 1M tokens.

As a result our OpenAI lineup now has GPT-5.6 as the flagship and GPT-5.5 as the balanced option, while GPT-5.4 has retired from new comparisons (its past data stays viewable).

The recommendations below have been updated to match: where they used to name GPT-5.2 / GPT-5.4 as the dependable all-round choice, they now name GPT-5.5 / GPT-5.6.

On June 9, 2026 Anthropic released Claude Fable 5 — the public version of its Mythos model — and we placed it in our Anthropic flagship slot.

It briefly held that slot before Anthropic abruptly withdrew Fable 5 (and Mythos 5) for all customers on June 12 to comply with a US government export-control directive.

That directive has since been lifted and Anthropic restored API access, so as of July 1, 2026 Fable 5 is back in our lineup as the Anthropic flagship, with Claude Opus 4.8 sliding back to a strong second at half the price ($5/$25 per million tokens).

Because Fable 5 was unavailable for several weeks, it is only now rebuilding benchmark data, so its per-genre rankings are still thin for the moment.

The Fable 5 page is live again.

Anthropic moved fast again.

Claude Opus 4.8 landed on May 28, 2026 — only about six weeks after Opus 4.7 — and it now holds our Anthropic flagship slot.

The pitch is sharper judgement, more honesty about its own progress, and the ability to work independently for longer.

The most concrete gain is in code reliability: Opus 4.8 is roughly four times less likely than 4.7 to let a flaw in its own code pass unremarked, and it leads on agentic software engineering with 69.2% on SWE-Bench Pro, ahead of GPT-5.5 and Gemini 3.1 Pro.

Pricing is unchanged from 4.7 at $5/$25 per million tokens, so the upgrade is essentially free on a per-token basis.

On this site Opus 4.7 is now kept readable for past comparisons but no longer participates in new sessions.

Wherever this article says "Claude Opus," read it as Opus 4.8 going forward.

Two new flagship models arrived in quick succession this spring.

Anthropic released Claude Opus 4.7 on April 16, 2026, just weeks after GPT-5.4.

OpenAI followed with GPT-5.5 on April 23, 2026.

Both bring meaningful step-changes in agentic coding and tool-coordinated workflows.

Claude Opus 4.7 ships with a new tokenizer and cleaner adaptive-thinking surface; GPT-5.5 posts measurable gains on long-horizon coding (Expert-SWE 73.1% on 20-hour tasks) and on computer-use benchmarks (OSWorld-Verified 78.7%).

What does that mean for the recommendations below? Mostly the labels shift slightly.

The "high-quality" tier in the GPT family is now best represented by GPT-5.5, with GPT-5.4 still very capable as a balanced option.

Claude Opus has moved from 4.6 to 4.7 in our flagship slot; on this site the previous Opus generation is kept readable for past comparisons but no longer participates in new sessions.

The Gemini lineup is unchanged for now.

Pricing on the new flagships is on the higher side: GPT-5.5 doubles GPT-5.4's output rate, while Claude Opus 4.7 holds the same $5/$25 per-million pricing as 4.6.

For most everyday choices the older balanced models still hold up well; for genuinely demanding work, the new flagships are where the gains live.

When choosing an AI model, it is easy to default to questions like “Which model performs best?” or “Which one is the cheapest?” Those are important questions, but in practice they are not enough on their own.

The right model changes depending on what you want to do, how much quality you expect, and what level of cost you are comfortable with in day-to-day use.

That is why this site separates performance comparisons from pricing and cost-performance comparisons.

AI is not something you can reduce to “stronger is always better” or “cheaper is always better.” In reality, the most sensible choice is the one that matches your needs within the balance of price, stability, and output quality.

If I had to summarize my view after this snapshot as simply as possible, it would be this: GPT’s lightweight model if you want to keep costs down, the higher-end GPT model for the best balance of overall strength and cost, and Claude Opus for work where the quality of the writing itself is the point, such as fiction or planning.

Rather than there being one perfect all-purpose model, each one has a fairly clear personality.

Keep in mind, though, that each genre has a single task, so read any genre-level strength or weakness as “that was true on that one task”.

On price, the model I would point to first is GPT’s lightweight model.

In this site’s measurement it had the lowest cost per answer of the six models, while scoring at roughly the level of Google’s flagship.

Against Google’s model in the same lightweight tier, the gap is clear in both cost and score.

In other words, there is now a cheaper and stronger choice than Gemini Flash-Lite, which earlier versions of this article put first for price.

For everyday work — writing, studying, organizing tasks, drafting a first take on a question — it handles more than you would expect at this price.

On programming and educational explanations it came close to the top models.
Its weak spot is just as clear.

On this snapshot’s planning task it misread the rule about the order of jobs and concluded with a plan that could not work, which cost it heavily.

When you ask it to reach a conclusion from intricate conditions, it is worth checking that the conclusion actually meets them.

If a beginner is choosing a first model, GPT’s lightweight model is still the easiest to recommend: it is cheap, rarely goes badly wrong, and is not picky about the kind of work.

Across all sixteen tasks, the top spot went to OpenAI’s higher-end GPT model.

It did not collapse in any genre and stayed mid-table even on its weaker tasks.

It also won every debate against the models of its tier.

Its strength is resilience rather than flair, and it is very easy to work with when you want output that holds together and can be used as is — coding, design, explanation, analysis.

The other thing not to overlook is cost: per answer it cost a fraction of what Claude Opus did.

If you look at quality and cost together and want just one model, this is the safest choice right now.

Where the quality of the writing itself is being judged, Claude Opus showed its strength.

It led on the fiction, planning, brainstorming, humor and empathy tasks, producing work that impresses with its overall polish, its structure, and the way it builds a whole answer from a somewhat ambiguous request.

There is also one point that this site does not fully capture through direct numerical comparison, but that still matters in practice: how good Claude can look when you ask it to build a site.

In my experience, Claude Code can sometimes produce a relatively modern-looking design even without heavy instruction, whereas Codex tends to produce designs that feel safer, more restrained, and more conventional overall.

Of course, this still depends on the prompt and the project conditions, but in actual use the difference can feel fairly noticeable.

Still, this is not an area where it makes sense to talk only about strengths.

Claude Opus is expensive: in this measurement it had the highest cost per answer of the six models.

Its answers run long, and it often feels slower than Codex.

There are also tasks, such as counseling, where the higher-end GPT model did better.

If you are willing to pay for the quality of fiction or planning, Claude Opus is a very strong option, but it can become a heavy load if you use it constantly.

Anthropic’s model one tier below led on the business writing and analysis tasks, and matched the top models on programming.

It costs roughly half as much as Claude Opus, so for business documents and reports it is worth trying first.

On the other hand, it scored surprisingly low on this snapshot’s fiction task.

If fiction is your main use, Claude Opus is the better fit.

Google’s flagship is inexpensive and came close to the top on educational explanation, business writing and summarization.

On the persuasion task, however, it promised something its own figures contradicted and lost heavily, and it did not win a debate against its tier.

Google’s model one tier below finished last of the six this time.

On the math problem it got all four answers wrong, and on the fiction task it misunderstood the story’s central rule.

It has a use for running large volumes of work whose results are easy to check, such as short summaries or routine sorting, but I find it hard to recommend for work where quality matters, at least for now.

The best way to avoid making a poor AI choice is not to focus only on whichever model looks strongest in the abstract.

In practice, the answer changes depending on whether you need to use it every day at scale, whether your work demands a high level of polish, or whether you simply want to experiment cheaply at first.

High-end models are undeniably attractive, but if you use AI constantly, cost and speed matter.

On the other hand, even a cheap and useful model may not be the one you want when the final result really has to look polished.

Personally, I think choosing an AI model is less about hunting for “the strongest model” and more about finding the tool that feels best for the way you work.

Once you decide whether your real priority is low cost, stability, or polish, the choice becomes much clearer.

That is the most practical way to frame it — read the caveat column alongside the strength.

Every model’s result on every task, with the answers and the judges’ comments, is on the rankings and genre pages.

In other words, no single model is perfect.

Their strengths and weaknesses are actually fairly easy to understand once you use them this way.

That is exactly why, on this site, I would recommend thinking about them as follows: GPT’s lightweight model for cost, the higher-end GPT model for overall strength, and Claude Opus for the quality of fiction and planning.