Best AI Models in 2026: TGK's Top 10 and Full Tier List

Updated 29 September 2026 · TGK group ranking

Every new AI model arrives with a chart showing it beat the last one, the company behind it posts its winning score, and every fan base has a reason its favourite is secretly number one. After a while that gets tiring, mostly because none of those numbers answers the question we actually care about: which model do we want beside us while we're writing an article, researching something or building a project?

We've used every model on this list, and in August we ran a direct comparison of Grok 4.6, Opus 5 and GPT-5.6 Sol. We've also published our own AI image generator and AI video generator tests with open data. We're an editorial site, not a benchmark lab, so we haven't tested every release at full capacity and won't pretend otherwise. What we can do is read the independent benchmarks properly, explain what they measure and where they disagree, and set them against our own work. A model that looks brilliant on a chart but wastes time, burns through tokens or hands us an answer we have to rewrite matters just as much as one that scores a little lower and gets the job done cleanly.

Quick Answer

Claude Opus 5.5 is our best model overall and our first choice for coordinating a project, writing and design. GPT-6 Astra is almost level on hard reasoning, although its long sessions get expensive, and GPT-6 Sol at high effort is the model we use most for everyday work. Claude Fable 5.1 ranks above Sonnet 5.5 on raw capability but costs far more in practice, while Sonnet 5.5 is sensible at low or medium effort and one of the most token-hungry models available at max. If you're new to this, the effort setting matters almost as much as the model you pick.

Which model for what?

The coding pick prioritises Bug Hunt's fix count and carries its largest selected bill. Appetite labels refer to max effort unless a setting is named, and "cheaper" compares published task costs, not subscription allowances.

TGK Top 10

Order and tiers follow our group's AI MODEL TIER LIST, 29 September 2026. All costs below use Artificial Analysis's max-effort configurations, which are separate from our lower-effort recommendations.

Token appetite means output volume, not price. Our labels translate AA's rounded output tokens per index task into five bands: Very low below 10k, Low from 10k to under 30k, Medium from 30k to under 70k, High from 70k to under 150k, and Very high at 150k and above. Sources include AA's model pages listed below and its comparisons of Opus 5 and Sol 5.6, Fable 5 and Kimi K3 and Luna and Sol.

Token appetite, explained

Sonnet 5.5 is the clearest example of why the effort setting belongs beside the model name. At max, AA's 28 September assessment measured about 193,000 output tokens per index task, the highest figure it has recorded and roughly 60% more than Opus 5.5 at max and seven times Astra at max. Sonnet also sits off AA's cost Pareto frontier, which means other tested configurations offer at least as much measured capability for less money.

Anthropic's efficiency claims concern lower effort. AA's low-versus-medium comparison reports about 14k and 19k output tokens per task, costing $0.41 and $0.59 respectively. That supports our everyday recommendation at low or medium, while max remains one of the most expensive configurations in this comparison.

A token is a small piece of text the system reads or generates, including billed reasoning that may never appear in the answer. The bill depends on how many tokens the job consumes, the rate for those tokens and the effort setting, plus input context, caching and any additional charges. A short visible answer can still involve considerable reasoning and repeated reads of a long conversation.

Astra illustrates the other side. AA measures roughly 27k output tokens per task at max, but its standard API rates are $10 per million input tokens and $50 per million output tokens. Compare that with Sonnet 5.5's $2/$10 and Opus 5.5's $4/$20, or Sol's $2/$10. Astra can complete a benchmark task efficiently while still becoming expensive over my long sessions, which is exactly what I've been noticing.

The same distinction explains our Fable-versus-Sonnet note. Fable's token use and cost are much worse in our everyday use, while Sonnet at max generates more output in AA's suite.

The Top 10, in detail

Each entry separates measured results, attributed outside views and our own experience.

1. Claude Opus 5.5 · SSS

Benchmarks say: Opus at max leads AA's current index at 58, costing $5.98 per task. Its 41.7 planted fixes put it sixth in our Bug Hunt selection, behind Sonnet, Astra and Fable configurations.

People say: CodeRabbit's 22 September review praises Opus on demanding coding work, while finding that its lower-effort review configuration offers the better starting balance.

We say: Opus is my preferred coordinator and our pick for writing and design. The reduction in excessive dashes has impressed me, although writing quality still depends heavily on the style files and skills you give it. I prefer high effort for the lead role, where I want the model to carry a messy brief through to completion.

2. GPT-6 Astra · SS

Benchmarks say: Astra at max scores 53 on AA. In Bug Hunt, xhigh matches Fable 5.1 max's 43 fixes for $24.22 against $87.18, about 72% less by our calculation.

People say: AA's 9 September assessment highlights efficient token use alongside uneven knowledge-work results. A 5 September Reddit developer account describes clever problem-solving but excessive implementation scope and repeated redirection.

We say: Astra belongs close to Opus for difficult work, and its 3D results were outstanding when it arrived. My sessions have become expensive enough to affect when I choose it, and good task efficiency on a benchmark doesn't cancel that experience.

3. Claude Fable 5.1 · S+

Benchmarks say: Fable at max scores 53 on AA at $7.63 per task. Bug Hunt records 43 fixes in 73.1 minutes for $87.18.

People say: AA's 1 September analysis welcomed its capability gains but found that greater output outweighed a substantial cache-price reduction.

We say: Fable ranks above Sonnet on raw capability in our group list, but its token use and cost are much worse in our work. I would bring it into a difficult plan or ask it for a second opinion rather than keep it running through every routine step.

4. Claude Sonnet 5.5 · S+

Benchmarks say: Sonnet at max scores 56 on AA and leads our Bug Hunt selection with 55.5 fixes, although that configuration costs about $135 and takes almost four hours.

People say: CodeRabbit's 28 September review reports faster reviews and more useful catches than Sonnet 5 in its own pipeline, using low, medium and high settings.

We say: Sonnet makes sense for worker threads and everyday execution at low or medium effort. We keep it fourth, below Fable's raw capability and above Sol, and treat max as a deliberate choice for difficult work rather than a default.

5. GPT-6 Sol · S

Benchmarks say: Sol scores 48 at max for $1.05 per AA task, against 43 at high for $0.38. Bug Hunt's max result is 29.3 fixes for $9.33.

People say: AA's 22 September assessment emphasises lower task costs, with gains and regressions across different evaluations.

We say: Sol remains my everyday model and a useful coordinator. It listens well to writing instructions and follows structure, which makes it our cheaper writing choice. It can also give stupid answers when the effort level doesn't suit the job, and fifth place in S tier reflects that balance.

6. GPT-5.6 Sol · A+

Benchmarks say: The current AA pages put max at 47 and $1.99 per task, and xhigh at 44 and $1.18. Bug Hunt max fixes 43.5 bugs but takes 253.1 minutes.

People say: The same 5 September developer account prefers Sol 5.6's communication and practical judgment to early Astra sessions, despite noticing Astra's technical strengths.

We say: At Extra High, GPT-5.6 Sol has sometimes felt better than the newer GPT-6 models on real work. Plenty of competing models look good on a chart without becoming more useful to us than this older Sol. Extra High means xhigh here, so the max results above describe a different setting.

7. Claude Opus 5 · A+

Benchmarks say: Opus 5 at max scores 51 on the current AA index and costs $5.86 per task, which keeps it competitive with newer models.

People say: AA's 24 July launch assessment presented it as a less expensive way to approach Fable 5's capabilities.

We say: Opus 5 stays seventh because the group still rates its work highly. It was also part of our August comparison, where careful code work mattered alongside writing and factual answers.

8. Claude Fable 5 · A+

Benchmarks say: Fable 5 at max, with Opus 4.8 fallback, scores 50 on AA and costs $8.75 per task, which is more expensive than Fable 5.1 in today's suite.

People say: AA's 9 June assessment placed Fable 5 first at launch and explained the fallback configuration used for its evaluation.

We say: Fable 5 remains A+ and eighth in our group's list, but its cost makes it difficult to recommend for a new everyday workflow.

9. GPT-6 Luna · A

Benchmarks say: Luna at max scores 37 on AA for $0.07 per task, the lowest cost in our Top 10. Its Bug Hunt result is 18.3 fixes for a reported $0.52.

People say: AA's 22 September analysis sees lower cost as Luna's main advance, alongside regressions in its coding-agent evaluation.

We say: Ninth place reflects cost and practicality. Luna, like GPT-5.6 Terra, is a useful alternative when the bigger models hit their usage limits. My own Luna experience is limited because Sol hasn't given me the same allowance problems, so this placement is the group's call rather than a claim that I've run Luna as extensively as Opus or Sol.

10. Kimi K3 · A

Benchmarks say: Kimi at max scores 44 on AA at $2 per task. Its default-setting Bug Hunt run records 21 fixes in 107.8 minutes for $25.27.

People say: AA's 21 July assessment found strong agentic knowledge-work results, which helps explain the attention it got at launch.

We say: I liked Kimi for coding, but it hasn't matched the OpenAI and Anthropic models I keep using. The waitlist frustrates me too, because by the time I get access, two or three newer releases may already be competing for the same job.

Just outside the Top 10

Grok 4.7 is eleventh, Qwen 3.8 Max 0902 twelfth, GLM-5.3 thirteenth and Muse Spark 1.3 fourteenth. Grok is strong now and has always been the relatively cheap option for me, although colleagues say they sometimes hit the limit of a $20 plan with the newer versions, which surprised me.

Muse's reasoning impressed me. A technically experienced friend of TGK used it while designing 3D-printed bridges for his Warhammer game and reported that they held 40 pounds, although he still found Opus the best and continues to use Claude.

Gemini keeps a narrow role as my copilot for smaller pieces of information. Our group ranks the listed Gemini models in a separate Google tier, below E, so that supporting use shouldn't be mistaken for a strong general recommendation.

Full TGK ranking · 29 September 2026

This reproduces all 39 models in the exact order of our group's master document. The full table will move to the AI & Tech Hub and be updated there.

Group notes: Fable 5.1 ranks higher on raw capability, but its token usage and cost are much worse in our use, which makes Sonnet 5.5 the more sensible everyday option. GPT-5.6 Sol in its hardest Extra High reasoning mode has sometimes felt better than the newer GPT-6 models on real work.

Tier key

Deep dive: what the charts actually measure

Bug Hunt: fixes, time and money

Bug Hunt Bench asks coding agents to find defects in two production codebases containing 105 planted bugs, and a model judge checks the repairs against a hidden answer key. Our selected rows, checked against the published data, show how much time and money accompany each score.

TGK calculation: cost per fix is the reported cost divided by planted fixes, rounded to cents. Astra xhigh and Fable max both fixed 43, but Astra cost 27.8% as much. Opus leads AA's broad index yet sits sixth here, which shows how a general capability ranking and a debugging result can disagree.

What extra effort bought

Astra's last step adds two fixes for $8.81 and 31 more minutes, increases of roughly 36% in cost and 53% in time. Sonnet's step from xhigh to max adds 16.5 fixes, $74.49 and 112.3 minutes, while Sol's move from high to max adds 9.3 fixes for another $5.17.

Unplanted fixes and knowing when to stop

The benchmark also records genuine defects fixed outside the planted set. Astra at max averages 55 of these additional fixes, Astra xhigh 53, GPT-5.6 Sol max 48.5 and GPT-6 Sol max 41.3, while Fable 5.1 records 7 at high and 11 at max, Opus 5.5 max 9 and Sonnet 5.5 max 24.5.

Those extras can be useful in an open audit and can create additional review work in a tightly scoped change. The question for us is whether the finished work matches the brief and leaves a manageable amount for a person to check.

Reading Sonnet's launch chart

The Anthropic launch table gives different answers depending on which row you read.

Anthropic links some of Sonnet's FrontierCode failures at max to extra subagent review, timeouts and edits beyond the task, which is a direct example of a model not knowing when to stop.

Artificial Analysis: the broader comparison

AA's Intelligence Index v4.3.2 combines ten evaluations across agents, coding, scientific reasoning and general work. These values were checked on 29 September.

MiMo-V2.6-Pro appears here as a cost comparison and is a different model from MiMo V2.5 Pro in our master list.

Vox's advisor workflow: Opus leading, Fable on call

The workflow shared by Vox (@Voxyz_ai) uses Opus 5.5 on high for the main session, Opus on medium for subagents, and Fable for advice around difficult plans, repeated failures and final review. I like that division because it fits how I think about a coordinator working with separate threads.

Claude Code's advisor documentation confirms that /advisor fable works with Opus. Claude decides when to consult the advisor, and each consultation reads the conversation and adds usage, with the advisor's own read not cached. The Vox post also corrects its original setup instructions: select Opus high through /effort high or /model rather than relying on the original top-level setting. We find it a useful community workflow to try, not something we've measured savings from ourselves.

If you're setting up Claude Code for a real project, our Claude Code security checklist covers what to check before you launch.

How we built this ranking

As a TGK group, we have used every model on this list, and the detailed analysis focuses on the Top 10. Depth of use varies, and my limited experience with Luna is stated above. The dated group master table decides the order and tiers, informed by our work and the external evidence. Some of that work is documented in full, such as the time we built and launched a web app with Codex without a developer on the team.

We judge models by the jobs we actually give them: preserving meaning and voice in writing, researching without invented facts or references, making sensible design decisions, reasoning through messy requests, coordinating a project, knowing when to stop, and the time, tokens and human repair needed to finish. A strong coding score can't settle all of those questions.

Our evidence plays three distinct roles:

We chose AA because it combines varied tasks with cost and token measurements, and Bug Hunt because it puts completed fixes beside time and money. Arena's anonymous comparisons measure user preference, which is useful context, but we haven't turned an Arena position into points in this ranking. Vendor charts help explain claims when we read their conditions and footnotes.

The weighting is editorial: our use determines the recommendation, while external measurements support or challenge it. We haven't averaged incompatible scores or assigned invented percentages to personal impressions, and our own analytical work consists of checking configurations, reconciling disagreements and calculating what extra effort bought.

Easy prompts that every leading model passes tell us little. Silly trick questions sometimes expose absurd mistakes, but we wouldn't build a ranking around them. We aren't defending one company either, since Opus and OpenAI models can make the same mistakes and fall into the same annoying writing rhythm. Grok 4.6 produced the best individual passage in our August test, while GPT-5.6 Sol gave us the better balance across the work.

We'll keep this ranking updated in the AI & Tech Hub, with the date and the reason for every change.