July 2026. Three frontier models sit at the top of every leaderboard: Anthropic's Claude Fable 5, OpenAI's GPT-5.6, and Google's Gemini 3.5. Everyone wants to know which one is king. It depends on what you're doing, and the article says so up front. Here's the side-by-side after running all three against real prompts, with the Artificial Analysis scores and each vendor's pricing page open in another tab.
The Three-Horse Race, At a Glance
Before we go deep, here's the cheat sheet. All three are verified shipping models as of July 2026 (cross-checked against Artificial Analysis and each vendor's own pricing pages; no rumored names here):
| Model | Vendor | AA Intelligence Index | Pricing (in / out per MTok) | Strongest lane | Output speed | |-------|--------|----------------------|------------------------------|----------------|--------------| | Claude Fable 5 | Anthropic | 59.9 (#1) | $10 / $50 | Long reasoning + creative writing | ~63 t/s | | GPT-5.6 Sol | OpenAI | 58.9 (#2) | $5 / $30 | Math, logic, multimodal, tools | ~78 t/s | | Gemini 3.5 Flash | Google | Flash tier (Pro pending) | $1.50 / $9 | Speed, multimodal, long context | 289 t/s |
If you just want a one-line answer: Fable 5 is the smartest in long-form reasoning and the most genuinely creative; GPT-5.6 is the sharpest at hard logic and the most flexible knob; Gemini 3.5 is the fastest and the best at juggling images, audio, and text in one pass. Now the details, because none of that is the whole story.
Claude Fable 5: Anthropic's Mythos-Class Flagship
Don't let the name confuse you. Claude Fable 5 is real: Anthropic's Mythos-class flagship, the first of the Claude 5 family, released June 9, 2026, with capability above Opus. As of July 2026 it sits at #1 on the Artificial Analysis Intelligence Index at 59.9, a single point ahead of GPT-5.6 Sol. It carries a 1M token context window, 128K max output, and on Anthropic's own SWE-Bench Pro it scores 80.3%, eleven points clear of the previous Opus. That ordering matters, because a lot of people still assume Opus is the top. It isn't anymore.
Where Fable 5 genuinely pulls ahead:
- Long-horizon reasoning. I've watched it hold a thread across 20+ steps without losing the plot. Most models start paraphrasing themselves or drifting around step 8. Fable 5 doesn't.
- Creative writing that doesn't read like a template. This is the big one for content work. GPT-5.6 and Gemini 3.5 both default to safe, structured prose. Fable 5 takes risks: voice, rhythm, actual opinion. It's the only frontier model where the output sometimes surprises me.
- Complex code refactors. Multi-file, cross-dependency stuff where you have to keep the whole module in your head. Fable 5 handles this cleanly; Opus 4.8 was already strong here and Fable 5 is a step up.
- Safety alignment. Anthropic's red-teaming shows there's a built-in classifier that redirects roughly 5% of sensitive sessions to Opus 4.8 with a clean HTTP 200. It's the model I'd hand a sensitive policy question without flinching.
The tradeoffs are real, though. Fable 5 is expensive ($10 in / $50 out per million tokens), double Opus 4.8 and roughly 10x what Gemini 3.5 Flash charges on input. At ~63 tokens/second it's also the slowest of the three. For real-time chat, that's a non-starter. And Chinese, while competent, is not its native register; it reads like a brilliant non-native speaker. For pure Chinese work, Qwen3.7 Max (1M context, $2.50/$7.50, the strongest Chinese of any model we've tested) still wins on naturalness.
GPT-5.6: The Reasoning Powerhouse with a Knob
GPT-5.6 is OpenAI's current flagship, released July 9, 2026, and the standout feature isn't raw intelligence; it's reasoning_effort. You can dial reasoning from high to xhigh to max to ultra, trading latency and cost for depth. On Artificial Analysis it lands at 58.9, one point behind Fable 5, but at roughly half the cost per task ($1.04 vs $2.75). That single knob changes everything about how you'd use it in production.
OpenAI actually shipped three tiers, not one: Sol ($5/$30), Terra ($2.50/$15), and Luna ($1/$6). A 5x price spread inside a single generation; pick wrong and you pay 5x too much.
Where GPT-5.6 wins:
- Math, logic, formal problem-solving. On the hardest reasoning evals it's at or near the top of Artificial Analysis's leaderboard. Fable 5 trades blows here; Gemini 3.5 trails.
- Multimodal understanding. Feed it an image plus text plus audio context and it holds it together better than anyone except Gemini.
- Tool use and structured workflows. The function-calling ecosystem around GPT-5.6 is the largest and most battle-tested.
- The effort dial. This is the killer feature for routing. Easy query on
high, hard query onmax. The cost curve is yours to shape.
Where it falls flat: creative writing trends conservative and a little corporate. The Chinese output has a "translated" feel: technically correct, slightly off in cadence. And max effort is genuinely expensive; on real workloads xhigh is the value sweet spot and most users will never need max.
Gemini 3.5: Google's Agentic Multimodal Bet
Google DeepMind announced Gemini 3.5 at I/O 2026 on May 20, with Flash shipping immediately and Pro still in private preview. Google positions it around four verbs: perceive, reason, use tools, interact. That's agentic in the true sense: not just "can call a function" but "can hold a multimodal context and act on it across steps." On June 2, 2026, Google cut Flash pricing 40%, landing it at $1.50 in / $9 out per million tokens, the cheapest tier in this comparison.
Where Gemini 3.5 wins:
- Speed. Flash outputs at 289 tokens/second, 4x faster than the previous Flash tier. It feels instant. Nothing else in this comparison comes close.
- Native multimodal. It's the best at understanding images alongside text in a single pass (84.2% on CharXiv Reasoning). Fable 5 and GPT-5.6 can do multimodal, but Gemini 3.5 was built for it from the ground up.
- Long context. 1,048,576-token window. Drop an entire book or a whole codebase in and ask questions. The other two cap out lower.
- Agentic coding. On Terminal-Bench 2.1 it hits 76.2%, a Flash-tier model beating the previous Pro tier on agent benchmarks. That's the headline of the release.
The weakness is personality. Gemini 3.5's creative writing is factual, dry, and a little encyclopedic. If you want a tarot reader with actual voice, or a character that feels alive, you'll be disappointed. It's the model I'd trust to summarize a 500-page legal document, and the last one I'd ask to write a short story.
Head-to-Head: Eight Dimensions
Here's how I'd score them across the dimensions people actually care about (subjective impressions from real prompts, layered on top of the Artificial Analysis numbers; treat them as a starting point, not gospel):
| Dimension | Winner | Notes | |-----------|--------|-------| | Hardest reasoning | Fable 5 (59.9) / GPT-5.6 max (58.9) | Tied at the top; Fable 5 edges on prose-heavy reasoning, GPT-5.6 on formal math | | Creative writing | Fable 5 | Not even close; the only one that doesn't sound like a template | | Coding | Fable 5 (complex) / GPT-5.6 (general) | Fable 5 wins on refactors; GPT-5.6 on broad productivity | | Multimodal | Gemini 3.5 | Native from the ground up; Fable 5 and GPT-5.6 are good but bolted on | | Speed | Gemini 3.5 Flash | 289 t/s vs 63-78 t/s, not a contest | | Cost | Gemini 3.5 Flash / Qwen3.7 Max / DeepSeek V4 | Both frontier models are 3-10x more expensive | | Chinese | Qwen3.7 Max (not in this trio) | Of the three, Fable 5 is most natural; GPT-5.6 most "translated"; Gemini 3.5 in between | | Agentic tools | Sonnet 5 / Gemini 3.5 | Fable 5 can do it but Sonnet 5 is the cheaper agentic workhorse |
Who Wins What (TL;DR)
- Hardest reasoning: Fable 5 (59.9 on AA) or GPT-5.6 on
max. Flip a coin on formal math; Fable 5 on anything prose-shaped. - Agentic tool use: Sonnet 5 (cheaper, near-Opus quality) or Gemini 3.5. Fable 5 can, but it's overkill cost-wise for most agentic loops.
- Speed-sensitive anything: Gemini 3.5 Flash. Period.
- Creative writing in English: Fable 5. By a mile.
- Creative writing in Chinese: None of these three. Qwen3.7 Max is the answer. (Yes, really.)
- Long-context book/codebase work: Gemini 3.5, because of that 1M+ window.
- Cheapest acceptable quality: DeepSeek V4 ($0.14/$0.28, the price butcher), Qwen3.7 Max, or Gemini 3.5 Flash.
Routing, as a design stance
Everything above is editorial, and none of it is a shelf list. Inside x-rush the practical version of this trio problem disappears by design: AI Chat and the other text features take the brief, route each task internally to whatever fits its language, difficulty, and tooling needs, and bill a flat 0.5 credits per text generation no matter which engine served it. Vendor names never reach the interface, so a leaderboard reshuffle like the September AA v4.3 refresh is a backend note, not a broken product promise.
That stance is the working answer to "which model should I standardize on?" in a year when the top of the board changed twice. Standardize on the brief, not the name.
The Takeaway
There is no king of 2026. Claude Fable 5 is the smartest in long-form reasoning (59.9 on AA) and the only one that writes like a person. GPT-5.6 is the sharpest at hard logic (58.9, half the cost per task) and the most flexible knob in the business. Gemini 3.5 is the fastest (289 t/s) and the best at juggling modalities. They're each the best at something important, and each genuinely bad at something the others do well.
If you're picking one model for everything, you're doing it wrong. The right answer in 2026 is a decision per task: Fable 5 where reasoning or voice matters, GPT-5.6 where you want a controllable effort dial, Gemini 3.5 where speed or multimodal matters, Qwen3.7 Max for Chinese content at scale, DeepSeek V4 for cost-bound volume. Inside x-rush that decision is made for you, per task, and the price is the same either way.
Pick a lane. Or let a router pick it for you.
Open AI Chat with a real brief and judge the answer, not the engine. The engine, whichever it is this quarter, is not your problem.