We humanize LLMs a lot these days. Since we spend so much time with them, let’s imagine—for this story—that they really are our friends. Just like the people around us, they shape our standards, performance and blind spots.
For years, GPT has felt like the default AI companion. Today, Anthropic’s Claude Opus 5.5 leads this shortlist in factual reliability, scientific reasoning and coding. The brilliant friend you call when a problem gets difficult.
But that friendship comes at a price: Claude cost $5.98 per benchmark task, versus GPT’s $0.72. OpenAI’s GPT-6.1 Sol stays close on capabilities while costing considerably less—the pragmatic friend for more workflows. Everyday costs vary. AA comparison.
Google DeepMind’s Gemini 3.8 Flash is the visual thinker. Its visual understanding approaches the leaders, while reliability, reasoning and coding lag behind Claude and GPT.
Z.ai’s GLM-5.3, from China, is the coding friend. It ranks behind Claude and GPT in our coding score, but factual reliability is weaker. Its tested Max variant accepts text only.
Moonshot AI’s Kimi K3, also from China, is more convincing in scientific reasoning than GLM, with respectable visual understanding. Its weaker terminal performance pulls down the coding average. AA comparison.
France’s Mistral AI brings Large 4 Preview, the affordable wildcard. Its sale pricing is attractive, but it ranks lowest here in scientific reasoning and ties DeepSeek for weakest factual reliability. Its visual performance remains unscored. AA comparison.
xAI / SpaceXAI’s Grok 4.7 is relatively reliable on facts, but trails the leaders in reasoning and coding. Its tested configuration costs considerably more than GPT. AA results.
Meta’s Muse Spark 1.1 illustrates why old friendships deserve a fresh look. Its factual reliability is respectable, but terminal work is a weakness. This older version does not represent Meta’s latest release. AA model profile.
Alibaba’s Qwen3.8 Max offers stronger coding than Grok, but weaker factual reliability and high task costs. DeepSeek’s V4.1 Flash is the cheapest friend in the entire group, although factual reliability is a clear weakness. AA comparison.
| Provider / model | Reliability ↑ | Reasoning /100 ↑ | Coding /100 ↑ | Vision ↑ | USD/task ↓ |
|---|---|---|---|---|---|
| OpenAI / GPT-6.1 Sol | 42 | 42.5 | 55.0 | 86.0% | 0.72 |
| Anthropic / Opus 5.5 | 46 | 46.5 | 63.5 | 87.7% | 5.98 |
| Google / Gemini 3.8 Flash | 30 | 33.0 | 38.5 | 85.6% | 1.24 |
| xAI / Grok 4.7 | 32 | 30.5 | 41.5 | — | 3.74 |
| Meta / Muse Spark 1.1 | 28 | 30.5 | 32.5 | — | 1.38 |
| Alibaba / Qwen3.8 Max | 12 | 30.5 | 45.5 | 82.8% | 5.41 |
| DeepSeek / V4.1 Flash | -5 | 26.5 | 39.5 | 77.0% | 0.27 |
| Mistral / Large 4 Preview | -5 | 23.0 | 40.5 | — | 1.13* |
| Moonshot / Kimi K3 | 20 | 35.0 | 36.0 | 80.5% | 2.00 |
| Z.ai / GLM-5.3 | 14 | 30.5 | 50.5 | — | 2.01 |
Snapshot checked: 12 October 2026. Max settings unless stated: Gemini High; Grok/Muse Xhigh; Claude Max with Default Fallback; Qwen version 0902; Mistral Preview. One model per provider, rather than every provider’s strongest model.
Reliability uses AA Omniscience (-100 to +100): correct answers are rewarded, errors penalized and abstentions neutral. It is not a hallucination percentage. Reasoning is my equal-weight average of HLE and CritPt; coding averages Terminal-Bench and SciCode. These measure scientific reasoning and terminal/scientific coding, with limited coverage of everyday judgment and software quality. CritPt and SciCode are currently marked under review by AA. Methodology.
Vision uses MMMU-Pro, republished by BenchGecko. Dashes: no confirmed score for Mistral, Muse or Grok in the sources checked; GLM is text-only. Image creation and audio/video performance remain unscored because comparable evidence is missing across this selection.
DeepSeek, GLM and Kimi offer open weights; tested Qwen Max is proprietary. Costs are weighted API benchmark averages, excluding self-hosting. *Mistral’s sale rates imply about $0.57 for identical token use; this is an estimate.
Choose your AI circle as carefully as your human one—it’s great to have a few smart friends around, but intelligence isn’t the only quality that matters.