Insights · AI Model Rankings
Claude Fable 5 is holding the top spot, but not by much. GPT-5.6 Sol is a few points behind it, and a 2.8-trillion-parameter open-weight model out of Beijing just made the whole conversation more interesting. Here's an honest read on where things actually stand.
Photo by Immo Wegmann on Unsplash
Every few months someone asks us, more or less directly, "so which AI model should I actually be using?" It used to be a fairly easy question. Right now it isn't, and if you've tried to answer it yourself by skimming a leaderboard, you've probably noticed the numbers don't quite agree with each other from one site to the next. That's not a bug in any single tracker — it's just what a genuinely crowded field looks like. We pulled together the models actually worth knowing about this month, what they're good at, and where the leaderboard noise is hiding a simpler story underneath.
The current lineup
| Model | Developer | Type | Known for |
|---|---|---|---|
| Claude Fable 5 | Anthropic | Closed | Long-horizon, autonomous reasoning |
| GPT-5.6 Sol | OpenAI | Closed (expanding) | Raw speed — up to 750 tok/sec |
| Kimi K3 | Moonshot AI | Open-weight | Coding & agent tasks, huge context |
| Claude Opus 4.8 | Anthropic | Closed | Coding, computer-use agents |
| GLM-5.2 | Z.ai | Open-weight | Best open SWE-bench score before K3 |
| Gemini 3.5 Pro | Closed | 2M-token context, ecosystem reach | |
| DeepSeek V4.5 | DeepSeek | Open-weight (MIT) | Setting the price floor for everyone |
| Grok 4.5 | xAI | Closed | Cheapest pricing in the top 10 |
How close the top three actually are
Before getting into each model individually, it's worth seeing the gap for what it is. Pulled from the same snapshot of Artificial Analysis's public Intelligence Index, taken July 21, the top three models are separated by less than three points on a hundred-point scale. That's not "one model dominating," that's three labs essentially tied, at least on this particular measure.
Three labs, three different starting points and budgets, separated by under three points on a hundred-point scale.
A quick note before going further, because it matters more than most rankings admit: no two trackers score these models identically. Different weighting, different benchmark mixes, different update cadences. We're using Artificial Analysis's public index here because it's transparent about its methodology and gets referenced by the other trackers we checked, but treat any single number as directional, not gospel.
Claude Fable 5 — the current leader, with an asterisk worth knowing
Fable 5 is the first release in what Anthropic calls its "Mythos" tier — a step up from the existing Opus line, aimed specifically at long, mostly unsupervised work: planning a task, executing it, checking its own output, and course-correcting without a person hovering over every step. It's been sitting at or near the top of every major tracker since launch, and the margin over second place, while real, is narrower than the headlines suggest.
The asterisk: Fable 5 went dark for nineteen days in June. Anthropic pulled public access on June 12 to comply with U.S. Department of Commerce export controls, and restored it on July 1 once the relevant controls were lifted. It's been stable since. Worth knowing if you're building something on top of it and want to understand why some June screenshots floating around show it as unavailable.
Photo by Zulfugar Karimov on Unsplash
GPT-5.6 Sol — the fast one
OpenAI's answer to Fable 5 arrived as a three-model family — Sol, Terra, and Luna — with Sol as the flagship. The interesting part isn't really the benchmark score, which is close behind Fable 5's without quite catching it. It's that Sol runs on Cerebras's wafer-scale chips, hitting speeds up to 750 tokens per second, noticeably faster than anything OpenAI had previously shipped. Access started restricted to select partners at the June preview and has been widening gradually since. If you tried to sign up the day it was announced and couldn't, that's why.
Kimi K3 — the one everyone's actually talking about
K3 is the story of the summer, and it's not really about the ranking. Moonshot AI released a 2.8-trillion-parameter model — by parameter count, the largest open-weight system anyone has shipped — and it landed close enough to Fable 5 and Sol on independent benchmarks that the gap between "the best model" and "the best model you don't have to pay a premium lab for" basically closed overnight. Moonshot itself said K3 trails the top two on overall performance but beats everything else, including Anthropic's own Opus 4.8, on coding and general-agent tasks specifically. Third-party testing has broadly backed that up.
The model comes with a 1-million-token context window, native image understanding, and an always-on reasoning mode. It's also compatible with the OpenAI SDK, which lowers the bar for anyone already building on GPT or Claude to try swapping it in. Demand was strong enough that Moonshot paused new sign-ups within three days of launch; full model weights are scheduled for July 27, so the open-source community hasn't even gotten its hands on the whole thing yet.
Claude Opus 4.8 — still doing the heavy lifting
Easy to overlook Opus 4.8 now that its younger sibling Fable 5 is getting all the attention, but it briefly topped the index itself back in May and remains one of the strongest models available for coding and computer-use agent work specifically — 69.2% on SWE-bench Pro, 84% on Online-Mind2Web. If your workload leans heavily toward agents that need to actually operate software rather than just reason about it, Opus 4.8 is still very much in the conversation, not a retired predecessor.
GLM-5.2 — the open-weight model K3 overshadowed
Before Kimi K3 showed up, GLM-5.2 was the open-weight model people pointed to, and it's still genuinely good — the strongest open SWE-bench Pro score (62.1%) before K3 took the title, and meaningfully cheaper to run than most closed alternatives. If K3's size makes it impractical for your infrastructure, GLM-5.2 is the more modest, easier-to-self-host alternative worth a look.
Photo by jonakoh_ on Unsplash
Gemini 3.5 Pro — the ecosystem play
Gemini doesn't always top the raw capability charts, and Google doesn't seem especially bothered by that. What it offers instead is reach: a 2-million-token context window, deep integration across Search, Workspace, and Android, and pricing that got dramatically more aggressive the same week it launched, when Google cut its top-tier AI subscription by 60%. If you're already living inside Google's ecosystem, Gemini's advantage isn't benchmark supremacy, it's that it's already everywhere you're working.
DeepSeek V4.5 — the quiet price-setter
DeepSeek doesn't generate the same headline cycle Kimi K3 did, but its models have had an outsized effect on this entire market for over a year now: every time DeepSeek ships something competitive under a permissive license, every closed-model lab feels pressure to justify its own pricing. V4.5 continues that pattern. It's not the single best model on this list, but it's arguably the most influential one on what everything else costs.
Grok 4.5 — the value pick
Grok 4.5 doesn't win on raw capability against the top three, but it's the cheapest model in the top 10 by a real margin, and its tight integration with X gives it a genuine edge for anything that benefits from real-time social data. Not the pick if you need frontier reasoning. A reasonable pick if you need something competent, fast to access, and inexpensive to run at volume.
Photo by Jakub Żerdzicki on Unsplash
So which one should you actually use?
Depends what you're doing with it, and we mean that more literally than it might sound.
Claude Fable 5 is built for exactly this, and it currently leads the pack for a reason. If you're handing off a task and checking in at the end rather than every step, this is the one to reach for.
GPT-5.6 Sol's Cerebras-powered throughput is a genuine differentiator for anything latency-sensitive — real-time assistants, high-volume pipelines, interactive tools where a few hundred milliseconds actually matters to the person waiting.
Kimi K3 or GLM-5.2, depending on your infrastructure budget. K3 is bigger and currently more capable; GLM-5.2 is easier to actually run somewhere you control.
Claude Opus 4.8 is still purpose-built for this and performs like it. Don't assume "older" means "worse" here — it's a different tool tuned for a different job than Fable 5.
DeepSeek V4.5 or Grok 4.5, depending on whether you want open weights or a hosted API. Both exist specifically because someone decided the frontier shouldn't cost frontier prices.
Photo by sidney zou on Unsplash
Frequently asked questions
Is Claude Fable 5 really the "best" AI model?
On the index we're citing here, yes, narrowly. On a different tracker with different weighting, GPT-5.6 Sol edges ahead on some measures. Treat "best" as "leading by a small margin on most current trackers," not as a settled fact.
Can an open-weight model like Kimi K3 actually replace a closed frontier model?
For a growing number of use cases, yes, especially coding and agent tasks where K3 specifically claims to outperform even some closed models. For workloads needing the absolute frontier of general reasoning, the closed top two still hold a real, if narrowing, edge.
Why do different leaderboards rank these models differently?
Different benchmark mixes, different weighting formulas, and different update schedules. A model can lead one index and sit third on another without either tracker being wrong — they're measuring slightly different things.
Is Claude Fable 5 available to use right now?
Yes. It was suspended for nineteen days in June for export-control compliance and has been publicly available again since July 1, 2026.
The bottom line
If you'd asked us in January which model was "most powerful," there was a cleaner answer. Right now there isn't, and honestly, that's a healthier place for the industry to be. Three genuinely different organizations, with three very different resource levels, are within a few points of each other at the top of the field — and one of them is giving its weights away for free in a few days. Pick based on what you're actually building, not on which bar happens to be tallest in whichever chart you saw first.
Further reading on this topic
We'll keep updating this ranking as new models land and the trackers refresh their numbers. Check the Insights section for the next update.
Benchmark scores, pricing, and access details referenced in this article reflect publicly available tracker data and company statements as of the publish date and change frequently — always check current figures directly with the model provider or an independent tracker before making a decision based on them. This article does not contain affiliate links; where future articles do, they will be disclosed per our Affiliate Disclosure.